mCardiacDx: Radar-Driven Contactless Monitoring and Diagnosis of Atrial Fibrillation.

Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice IEEE J Transl Eng Health Med . 2026 Mar 4;14:144–163. doi: 10.1109/JTEHM.2026.3670383 Search in PMC Search in PubMed View in NLM Catalog Add to search mCardiacDx: Radar-Driven Contactless Monitoring and Diagnosis of Atrial Fibrillation Arjun Kumar Arjun Kumar 1 Department of Computer Science, KAIST, Daejeon, 34141, South Korea 2 Department of Software, Ajou University, Suwon, 16499, South Korea Find articles by Arjun Kumar 1, 2 , Noppanat Wadlom Noppanat Wadlom 1 Department of Computer Science, KAIST, Daejeon, 34141, South Korea 2 Department of Software, Ajou University, Suwon, 16499, South Korea Find articles by Noppanat Wadlom 1, 2 , Jaeheon Kwak Jaeheon Kwak 2 Department of Software, Ajou University, Suwon, 16499, South Korea Find articles by Jaeheon Kwak 2 , Si-Hyuck Kang Si-Hyuck Kang 3 Department of Internal Medicine, SNUBH, Suwon, 13620, South Korea Find articles by Si-Hyuck Kang 3 , Insik Shin Insik Shin 1 Department of Computer Science, KAIST, Daejeon, 34141, South Korea Find articles by Insik Shin 1, ✉ Author information Article notes Copyright and License information 1 Department of Computer Science, KAIST, Daejeon, 34141, South Korea 2 Department of Software, Ajou University, Suwon, 16499, South Korea 3 Department of Internal Medicine, SNUBH, Suwon, 13620, South Korea Corresponding Author: I. Shin ✉ Corresponding author. Received 2025 Oct 2; Revised 2026 Feb 9; Accepted 2026 Feb 25; Collection date 2026. © 2026 The Authors This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ PMC Copyright notice PMCID: PMC13068127  PMID: 41970944 Abstract Arrhythmia is a common cardiac condition that can precipitate severe complications without timely intervention. Among them, atrial fibrillation (AF) is the most common form. While continuous monitoring is essential for timely diagnosis, conventional approaches such as electrocardiogram (ECG) and wearable devices are constrained by their reliance on specialized medical expertise and patient discomfort from their contact nature. Existing contactless monitoring, primarily designed for healthy subjects, face significant challenges when analyzing reflections from AF patients due to disrupted spatial stability and temporal consistency caused by underlying irregular heart contractions. In this paper, we introduce mCardiacDx, a radar-driven contactless system that accurately analyzes these complex reflections and reconstructs heart pulse waveforms (HPWs) for AF monitoring and diagnosis. The key technical contributions of our work include a novel precise target localization (PTL) technique that accurately locates heart reflections despite spatial disruptions, an encoder-decoder model (HPR-Net) that effectively transforms these reflections into HPWs, addressing temporal inconsistencies, and a final analysis module for AF monitoring and diagnosis. Our evaluation on a dataset of 48 subjects (24 healthy, 24 with AF) in a seated, normal breathing, real-world setting shows that both mCardiacDx and the PTL technique significantly outperform the state-of-the-art approach in monitoring and diagnosing AF. Objective: To develop a contactless radar-driven system, mCardiacDx, that overcomes reflection disruption challenges in AF patients to accurately reconstruct interpretable heart pulse waveforms (HPWs) for monitoring and diagnosis.Methods and procedures: We introduce a PTL technique to locate heart reflections despite spatial disruptions, and an encoder-decoder model (HPR-Net) to robustly process reflections and reconstruct interpretable HPWs, addressing temporal inconsistencies. The HPWs are then processed by a final analysis module for AF monitoring and diagnosis. mCardiacDx is validated against a state-of-the-art approach (baseline) on a dataset of 48 subjects (24 healthy, 24 with AF) in a seated, normal breathing, real-world setting. This validation confirms the system’s robustness and generalizability to real-world seated scenarios variations in posture and environment.Results: mCardiacDx significantly outperforms the baseline in both monitoring and diagnosis. HPW fidelity (Dynamic time warping (DTW) score) for AF patients improves from 5.92 to 2.92. HR/RR interval median absolute percentage error (MedAPE) reduced (e.g., HR from 9.10 % to 2.94 %; RR interval from 8.42 % to 2.95 %). Our system achieves superior diagnostic performance with 0.93 accuracy, and 0.91 recall (sensitivity), significantly surpassing the baseline’s accuracy of 0.85 and recall of 0.75, while both maintain a specificity of 0.96.Conclusion: mCardiacDx is a robust, non-contact system for continuous cardiac care, addressing a critical gap in real-world AF monitoring and diagnosis. Keywords: Contactless arrhythmia monitoring and diagnosis, contactless atrial fibrillation monitoring and diagnosis, heart pulse waveform Clinical Impact: mCardiacDx offers a passive and continuous solution for AF monitoring and diagnosis that reduces patient discomfort from wearable devices and enabling consistent detection of cardiac events in home or clinical settings. Clinical and Translational Impact: mCardiacDx provides a non-contact solution for AF monitoring and diagnosis, addressing the critical limitation of existing wearable and contactless devices, and demonstrating high potential for passive, continuous integration into regular home cardiac care. I. Introduction Cardiovascular diseases (CVDs) are the leading cause of death worldwide, with an estimated 20.5 million deaths annually [1] , [2] , [3] . As a major category of CVDs, arrhythmias encompass various conditions where the heart beats with an abnormal heart rate or rhythm [4] . Among these, atrial fibrillation (AF) is the most common arrhythmia, where abnormal electrical signals in the atria of heart lead to asynchronous and irregular heart contractions, increasing the risk of severe complications such as stroke or heart failure [5] , [6] , [7] , [8] . The prevalence of arrhythmia is currently 2-9 % of the general population [9] , while AF alone affects approximately 52.5 million individuals worldwide [10] , and this is expected to rise in the coming years. Therefore, regular monitoring and early diagnosis of AF are critical to mitigate the risk of adverse events and provide a reference for effective clinical care [11] , [12] Given the significance, AF diagnostic technologies are rapidly advancing, particularly with the integration of artificial intelligence into these tools. Specifically, the standard method for diagnosing AF in medical practice involves conducting an electrocardiogram (ECG) examination at healthcare facilities like hospitals, which is inadequate for regular monitoring considering the challenges posed by regular commuting and the substantial involvement of medical professionals. To address this demand, ambulatory ECG devices such as Holter monitors [13] , [14] , [15] and ECG patches [16] , [17] have been developed. However, these require electrodes to be attached to the skin for long durations, which can disrupt daily activities like sleeping or showering. Additionally, the use of electrodes is not suitable for certain patients, such as newborns and those with skin injuries [18] , [19] . Smartwatches with single-lead ECG sensors [20] and photoplethysmography (PPG)-based devices [21] , [22] , [23] have also emerged. However, the former requires active user contact to the electrode button, limiting its use for truly continuous monitoring, while the latter is known to cause user discomfort and has a high rate of non-compliance, with 88 % of users removing them before bedtime [24] . To achieve regular cardiac monitoring with passive and continuous benefit, contactless sensing has emerged as a promising solution that leverages Wi-Fi [25] , [26] , RFID [27] , Ultra-Wideband (UWB) [28] , and Millimeter-wave (mmWave) radar [29] , [30] to detect thoracic wall vibrations. These works monitor the mechanical activity of the chest wall to measure basic heart rate and its variability, often quantified by the RR interval (the time between two successive heartbeats). Furthermore, recent efforts attempt to reconstruct cardiac waveforms from radar reflections like ECG [31] , [32] , [33] , seismocardiogram (SCG) [34] , and photoplethysmogram (PPG) [35] , [36] . Existing contactless methods are based on spatially single, stable, and temporally aligned reflections for cardiac waveform reconstruction in healthy subjects with normal heartbeats (synchronous and regular contractions). However, these underlying assumptions fail in patients with AF, where abnormal heartbeats (asynchronous and irregular contractions) create spatially dispersed and temporally misaligned reflections, fundamentally compromising waveform fidelity and diagnostic reliability. For instance, a study on atrial fibrillation (AF) showed that asynchronous and irregular heart contractions produce invalid ECGs without clear cardiac features, preventing AF diagnosis [37] . The same author in [38] demonstrates successful ECGs reconstruction. However, its reliance on calibration with a traditional ECG patch undermines the claim of being contactless, and its lack of output fidelity makes it unreliable for a true diagnosis. Its performance also declines significantly on unseen patients, indicating poor generalization beyond the training data. Similarly, a recent study [39] introduces a contactless AF monitoring system that achieve clinical-level performance on a large dataset. However, its reliance on a knowledge transfer model ambiguously maps radar features to true cardiac activity, limits interpretability and confidence in the diagnosis. Furthermore, these methods are often limited to controlled, idle postures, restricting their applicability to unconstrained common postures, such as sitting, where body motion is unavoidable. As noted in [40] , [41] , body posture significantly affects ECG signals, with seated positions causing more pronounced distortions, showing that methods for controlled postures cannot generalize to seated common scenarios. To sum up, a truly contactless system for monitoring and diagnosing AF in seated common scenarios does not yet exist. To bridge this gap, we propose mCardiacDx, a radar-driven contactless heart pulse waveforms (HPWs) measurement system that enables passive and continuous monitoring and diagnosis for both healthy and cardiac patients in seated common scenarios. Fig. 1 illustrates the usage scenarios of mCardiacDx, where we set up a low-cost mmWave radar [8] to sense the subject’s thorax wall vibration. We then introduce a custom-designed a signal processing technique and encoder-decoder model to precisely detect heart variations and transfer them into HPWs. The system then employs its heart health analysis module to perform AF monitoring and diagnosis based on these HPWs. In this clinical context, mCardiacDx serves as a contactless AF monitoring and diagnosis tool that identifies AF by analyzing HRV features derived from HPWs. Here, diagnosis refers to AF detection based on HRV-derived features and is not intended to replace standard ECG-based evaluation. Accordingly, mCardiacDx is positioned as a monitoring and screening support tool that complements routine clinical assessment and may facilitate timely referral for confirmatory ECG testing when AF-related irregularity is detected. In this way, mCardiacDx provides accurate HR and RR interval estimation from reconstructed HPWs, enabling continuous rhythm monitoring in seated common scenarios. To realize mCardiacDx, we identify and address two main challenges as follows. FIGURE 1. Open in a new tab The usage scenarios of mCardiacDx. mCardiacDx provides contactless heart pulse waveforms for cardiac monitoring and diagnosis of AF. C1: Locating spatially dispersed and variable-magnitude reflections. Different from previous works that focused on healthy subjects with synchronous and regular heart contraction, mCardiacDx focuses on reconstructing HPWs during asynchronous and irregular contraction in AF patients. For healthy subjects, the synchronous and regular heart contraction generates a unified mechanical force on the chest wall, which produces a coordinated and uniform chest-wall vibration, leading to reflections concentrated at a single stable location with consistent magnitude in signal space [42] . However, for AF patients, asynchronous and irregular contractions generate a scattered mechanical force. In particular, asynchronous and irregular contractions cause the heart chambers (atria and ventricles) to contract at varying times and force, which prevents the force from unifying, thus transferring as multiple, scattered impulses/forces to the chest wall. We infer that this force produces non-uniform chest-wall vibrations, where each vibration acts as a distinct source, resulting in reflections dispersed across multiple unstable locations each exhibiting varying magnitude in signal space [43] . Our statistical analysis across subjects (N = 210) confirms this observation: AF subjects exhibit a higher mean number of reflection locations (3.8 vs. 1.54) and greater variability (standard deviation (SD)) in both location count (0.93 vs. 0.64) and reflection magnitude (0.24 vs. 0.05). These findings in AF patients challenge existing works, which rely on the assumption of a single stable location. This assumption prevents them from locating reflections of varying magnitude dispersed across multiple unstable locations needed for accurate HPW reconstruction. To address this challenge, we introduce a novel precise target localization (PTL) technique, which employs a dynamic processing strategy to locate reflection samples of varying magnitude dispersed across multiple unstable locations for accurate HPW reconstruction. C2: Interpreting temporally misaligned reflections. The same underlying physiological factors that cause spatial dispersion and magnitude variation also cause reflections to be temporally misaligned with the underlying ventricular depolarization, which disrupts the predictable pattern of reflections. Healthy subjects exhibit temporally aligned reflections, whereas those from AF are temporally misaligned. Our statistical analysis between the radar reflections (Phase) and ground truth signal using Zero-Normalized Cross-Correlation (ZNCC) [44] provides quantitative evidence of this degradation, with healthy subjects showing a strong temporal alignment (mean ZNCC of 1.00) while AF patients exhibit significant misalignment (mean ZNCC of 0.47). The temporal misalignment prevents existing works, which are designed for temporally aligned reflections, from correctly interpreting such misaligned reflections or distortions for accurate HPW reconstruction. To address this challenge, we propose the heart pulse reconstruction network (HPR-Net), an encoder-decoder model that utilizes graph attention networks (GAT) to interpret the localized multi-bin features derived from PTL and effectively handle their inherent temporal misalignment for accurate HPW reconstruction. We prototype the end-to-end mCardiacDx using a Texas Instruments (TI) AWR1642BOOST mmWave radar [45] . We leverage a large-scale dataset of 210 subjects for system development, from which we reserve an independent test set of 48 subjects (24 healthy and 24 with AF) for evaluation in a seated, normal breathing, real-world setting in collaboration with a medical professional. Section IV provides comprehensive details regarding hardware specifications, participant demographics, the synchronization protocol, and network training parameters. We then compare mCardiacDx’s performance against the state-of-the-art method implemented as a baseline, with all measurements relative to the ground truth ECG. Since PTL is a signal processing technique, we examine its performance by integrating PTL into the baseline model, replacing the baseline’s conventional technique for locating reflections sources.This serves as component analysis to compare the performance of PTL against a conventional technique. Our results show that both mCardiacDx and the PTL-integrated baseline (baseline+PTL) outperform the baseline in monitoring and diagnosing AF. Our results show mCardiacDx and baseline+PTL achieve superior performance in HPW reconstruction, cardiac rhythm monitoring (HR and RR interval), and AF diagnosis compared to the baseline. For HPW reconstruction, mCardiacDx achieves high fidelity, with dynamic time warping (DTW) [46] scores for AF patients improving from 5.92 (baseline) to 3.78 (baseline+PTL) and 2.92 (mCardiacDx), indicating accurate waveform alignment during AF. In HR and RR interval estimation, both methods outperform the baseline, with mCardiacDx achieving the lowest median absolute percentage error (MedAPE) [47] (MedAPE reduced from 9.10 % to 2.94 % for HR and from 8.42 % to 2.95 % for RR interval). This improvement is further supported by stronger linear correlation with ECG (HR: increased from 0.361 to 0.704 and 0.915; RR: from 0.264 to 0.648 and 0.902) and improved Bland–Altman agreement with ECG, as evident by reduced bias toward zero (HR: from −4.89 bpm to 2.43 bpm and 0.44 bpm; RR: from 0.065 s to −0.029 s and −0.008 s). For diagnosis, both methods outperform the baseline, with mCardiacDx achieving the best performance, improving recall (sensitivity) from 0.75 to 0.91, f1-score from 0.83 to 0.93, and accuracy from 0.85 to 0.93, while maintaining a high specificity of 0.96. Results show mCardiacDx is a robust contactless system for pervasive and continuous cardiac care in seated common scenarios. Our contributions can be summarized as follows: • We propose a novel Precise Target Localization (PTL) technique that can locate the heart reflections dispersed across multiple unstable locations. It addresses the spatial instability caused by asynchronous and irregular contractions, making it possible to reconstruct HPWs even during complex rhythm irregularity such as AF. • We design a novel Heart Pulse Reconstruction Network (HPR-Net) leveraging a Graph Attention Network (GAT) to interpret the temporally misaligned reflections from PTL and transforms them into HPWs that enable monitoring and diagnosis of AF. • We prototype and validate the mCardiacDx system on a dataset of 24 patients with AF in real-world settings. To our knowledge, this is the first contactless system to demonstrate superior monitoring and diagnostic performance through interpretable cardiac waveforms without calibration with any wearables. II. Understanding Contactless Heart Pulse Waveform A. Radar Driven Cardiac Sensing The human heart consists of four chambers—the left atrium (LA), right atrium (RA), left ventricle (LV), and right ventricle (RV)—that work together to manage blood circulation through synchronous and regular contraction. The atria receive blood and pump it into the ventricles, which then store and propel blood throughout the body. The mechanical activity of the heart is generated under the stimulation of electrical activity, where mechanical contraction is initiated by electrical depolarization [48] . This activity is regulated by the heart’s electrical conduction system, which initiates at the sinoatrial (SA) node and propagates through the atrioventricular (AV) node, right bundle branch (RBB), left bundle branch (LBB) and the His–Purkinje fibers (PF) network. In healthy subjects, the atria and ventricles contract and relax in synchronous and regular manner due to normal electrical pathways as shown in left side of Fig. 2(a) [49] , [50] , [51] , [52] , [53] . However, this contraction is significantly disrupted in the case of AF patients due to a chaotic pathways, as shown in right side of Fig. 2(b) [54] , [55] , [56] , [57] , [58] , [59] . FIGURE 2. Open in a new tab Contactless cardiac sensing and heart’s morphology with electrical pathway. (a) An illustration of radar-based cardiac sensing capturing subtle chest vibrations. (b) A comparison of the electrical pathways in a healthy heart (left) and a heart with AF (right). As illustrated in Fig. 2(a) , radar-based cardiac sensing captures subtle vibrations on the chest wall during each heartbeat, which are a direct result of the heart’s contraction. The radar transmits mmWaves to the chest and receives the reflections. These reflections are represented by the 2-D channel impulse response (CIR) matrix, visualized as a radargram, with dimensions of fast time and slow time due to their different sampling rates [28] , [60] . Fast-time corresponds to the range of the reflecting surface, defining range bins where each bin represents a specific spatial location on the chest. Slow-time consists of fast-time snapshots of the reflections (or reflection samples). These samples are taken at a much lower rate, capturing their temporal evolution over repeated transmissions. Existing works analyzes these reflection samples across both dimensions to extract the chest wall vibration, which is then transformed into cardiac waveforms. We define the chest wall vibration at time as , which can be extracted from the phase of the reflection samples using Equation 1 , where is the mmWave wavelength. Existing contactless works rely on the fundamental assumption that a singular, stable source on the chest wall generates the cardiac reflection samples, which subsequently maps to a singular, stable location (a single range bin) in the radar data. This stability is the key principle for accurate signal processing in healthy subjects and forms the foundation of most current contactless cardiac sensing approaches. B. Limitations of Contactless HPW Reconstruction During Atrial Fibrillation To achieve contactless monitoring and diagnosis of AF, reconstructing HPWs from mmWave cardiac reflection samples is a straightforward approach. However, we found that such a solution is not feasible due to the limitations of existing methods. In particular, we implement a state-of-the-art contactless method an extended version of [34] as a baseline for reconstructing HPWs from mmWave radar reflection samples. In this context, the QRS waveform, particularly R peaks of the ECG correspond to ventricular depolarization, whereas the pulse peaks ( ) of the HPWs represent the mechanical pulse generated by the subsequent mechanical response (ventricular contraction) that follows depolarization. We denote the ECG R peaks as , the ECG-derived HPW pulse peaks as , and the reconstructed HPW pulse peaks as . Physiologically, the ECG R-peak precedes the mechanical pulse peaks ( , ) due to the inherent electromechanical delay; therefore, radar-derived pulse peaks are expected to occur after the corresponding R-peak. Our results show that the baseline performs well for healthy subjects but fails for AF patients ( Fig. 3 ). For healthy subjects, the reconstructed HPWs are transient, accurate in number, and closely match the ground-truth HPWs derived from the ECG (ECG-derived HPWs). In this case, the pulse peaks align well with both and ( Fig. 3(a) ). However, for AF patients, the reconstructed HPWs are distorted, reduced in number, and poorly match with the ECG-derived HPWs. Here, the fails to align with and ( Fig. 3(b) ). This misalignment leads to significantly higher median absolute percentage errors (MedAPE) of 9.10 % and 8.42 % for heart rate (HR) and RR intervals (time between two successive heartbeats, measured between consecutive and analogously between and ) in AF patients compared to much lower errors of 2.66 % and 2.73 % in healthy subjects, all computed relative to the ECG. FIGURE 3. Open in a new tab Performance comparison of baseline HPW reconstruction against the ground truth (ECG and ECG-derived HPWs): (a) Successful reconstruction for a healthy subject (b) Reconstruction failure for an AF patient. The primary cause of this failure stems from two interconnected challenges that the baseline is not designed to address. First, the uncoordinated and non-uniform movements in AF patients result in spatially dispersed reflection samples of varying magnitudes. This spatial instability conflicts with the baseline heart reflection selection technique, which identifies the heart reflection samples by assuming a singular, stable source. As a representative case shown in Figure 4 , a healthy subject exhibits high-magnitude reflection samples within a single, stable range bin ( Fig. 4(a) ), whereas an AF patient shows varying magnitude reflection samples dispersed across multiple unstable range bins ( Fig. 4(b) ). Our statistical analysis in (I.C1) supported this by quantifying the significantly higher standard deviation (SD) in both reflection samples range bin count and magnitude for AF patients. Without a single, stable location, the baseline cannot accurately locate cardiac reflection samples in AF patients. Second, this spatial instability causes temporal misalignment of reflection samples. The baseline translator uses a single-step convolutional encoder–decoder to learn a static transformation, which assumes a predictable temporal pattern. However, as shown in Figure 5 , AF causes reflection samples to become temporally misaligned with the underlying ventricular depolarization (R-peaks). Cross-correlation analysis performed on entire 60-second segments as a single window between radar and ECG signals highlights this. Despite the different physical domains (mechanical vs. electrical), cross-correlation of normalized signals is appropriate here because both modalities are physiologically coupled via the heart’s electromechanical periodicity. Healthy subject shows a strong temporal alignment (mean ZNCC of 1.00) ( Fig. 5(a) ) while an AF patient exhibit significant misalignment (mean ZNCC of 0.47) ( Fig. 5(b) ) in (I.C2). The single-step translation model cannot account for these temporal inconsistencies. Consequently, the baseline fails to reconstruct HPWs for AF patients. FIGURE 4. Open in a new tab Radargram of heart reflection patterns in CIR matrix Illustrating spatial instability. (a) Healthy subject reflection samples are concentrated into a single, stable range bin exhibiting mostly high magnitude. (b) AF patient reflection samples are spatially dispersed across multiple unstable range bins exhibiting varying magnitude. FIGURE 5. Open in a new tab Temporal alignment of heart reflection samples and ECG. (a) Healthy subject reflections are temporally aligned with ECG, resulting in a strong cross-correlation. (b) AF patient reflections are temporally misaligned with the ECG, resulting in weak cross-correlation. Without handling spatially dispersed locations and temporally misaligned reflection samples, existing methods are bounded to fail. For instance, methods using encoder-decoder [37] , knowledge transfer [39] , and cross-domain diffusion [38] models fail to process the complex, multi-source reflection samples characteristic of AF. To address these challenges, we propose a PTL technique to locate reflection samples of varying magnitude dispersed across multiple unstable range bins. We then developed an HPR-Network model with GAT to interpret these reflection samples to reconstruct the HPWs for AF monitoring and diagnosis. III. MCardiacDx System Implementation In this section, we detail the implementation of mCardiacDx, a contactless mmWave radar-driven system to reconstruct HPW for monitoring and diagnosis of AF. We first provide a concise system-level overview, followed by detailed descriptions of data preprocessing, signal processing components, learning-based component, and heart health analysis. A. MCardiacDx Overview mCardiacDx uses Frequency Modulated Continuous Wave (FMCW) radar to transmit signals and capture raw reflections from the subject’s thorax wall. As illustrated in the system architecture ( Fig. 6 ), these reflections are processed through a sequence of signal processing and learning-based components to reconstruct HPWs. The system is designed to address two core challenges of spatial dispersion (C1) and temporal misalignment (C2) by employing: (i) Precise Target Localization (PTL) to dynamically localize reflection samples dispersed across multiple unstable range bins, which are then refined through cardiac feature extraction; and (ii) a learning-based HPR-Net to reconstruct HPWs from radar-derived cardiac motion features, mitigating temporal misalignment. The reconstructed HPWs are subsequently used for heart rate (HR), RR interval estimation, and heart rate variability (HRV)-based AF diagnosis. FIGURE 6. Open in a new tab Schematic overview of mCardiacDx, which captures heart activities through mmWave radar and then transform them into HPWs through PTL and HPR-Network for heart health analysis. B. Data Preprocessing The objective of the preprocessing stage is to synchronize and refine the multimodal data streams to ensure a precise temporal correspondence between the radar reflection samples and the ECG samples. To ensure a unified time reference, data synchronization is achieved by calibrating both the radar and ECG devices via the central host PC prior to recording. Both systems log their respective capture start and end times in milliseconds directly onto the data files. While the radar data is streamed and stored on the host PC and the ECG data is logged locally on the device, these logged timestamps serve as a common temporal reference to precisely align the radar reflection samples with the ECG samples. To further address drift, we perform cross-correlation between the radar samples and the ECG samples to identify and compensate for any remaining temporal misalignment. This synchronized data serves as an input for PTL to accurately track reflection samples across range bins. Following alignment, the ECG samples are denoised to remove baseline wander, and the processed samples are subsequently transformed into an HPW pulse train. We accomplish this by detecting ECG R-peak timestamps and replacing each peak with a Gaussian pulse, providing a high-fidelity ground-truth reference for the subsequent learning-based models. C. Signal Processing Component This component addresses spatial dispersion through dynamic localization and extracts cardiac motion features from located reflection samples. 1). Precise Target Localization (PTL) Accurate localization of varying magnitude reflection samples from dispersed across multiple unstable range bins is crucial for accurate HPW reconstruction. The PTL algorithm operates through a sequential two-step process, as detailed in Algorithm 1 . In the initial step, the algorithm identifies the target range bin corresponding to the highest magnitude reflection samples. It does this by detecting the high magnitude samples across chirps using the Most Common Bin (MCB) algorithm [34] , [61] (lines 1-2). This process enables PTL to locate high magnitude reflection samples consistently concentrated in a single stable range bin. However, this initial step is inadequate for precisely locating reflection samples of varying magnitudes dispersed across multiple unstable range bins, as such samples may dynamically shift between different range bins over time. Algorithm 1 PTL Algorithm Require: CIR matrix , Time window size (in number of chirps) , Range bin window size Ensure: Selected range bins at each chirp 1: 2: 3: , 4: 5: 6: 7: while do 8: 9: 10: 11: end while To address this limitation, PTL enhances the target selection by implementing a dynamic processing strategy in the second step. It first defines a neighboring window around the target range bin by computing the first and last bins based on a predefined window breadth (line 3). The is a parameter that determines how many adjacent range bins are considered. This step extracts a sub-matrix that includes surrounding range bins (line 4), enabling the algorithm to capture reflection samples of varying magnitude that dynamically shift across neighboring range bins over time, thereby addressing the limitation of the initial step. However, merely expanding the range bin selection does not fully account for the dynamic spatial shifts of these reflection samples over time. As these shifts can evolve, it is crucial to continually monitor and track them for precise localization. To address this, PTL iteratively processes chirps within a sliding time window (line 7), extracts data for the current window (line 8), reapplies the MCB algorithm to refine the target range bin (line 9), and tracks these spatial variations in the identified range bin over time (lines 9–10). This iterative adjustment allows PTL to successfully locate reflection samples with varying magnitude dispersed across multiple unstable range bins. PTL’s continual tracking of these spatial variations over time provides a localized and dynamically tracked stream of reflection samples for the subsequent processing stages. 2). Validation of PTL We validate PTL in two steps. First, we confirm that PTL correctly locates reflection samples across range bins by identifying a continuous PTL Path (white dashed line). As shown in the heatmaps ( Fig. 7 ), PTL successfully locates reflection samples within a single stable bin for a healthy subject as shown through the PTL Path 7(a) In contrast, for an AF patient, the PTL Path dynamically locates reflections of varying magnitude dispersed across multiple unstable bins ( Fig. 7(b) ). Crucially, the algorithm’s iterative adjustment through a sliding time window ( Algorithm 1 , line 7) not only addresses localization challenges arising from spatial dispersion but also maintains a temporally consistent path even in regions of low reflection magnitude (indicated by light blue in Fig. 7(b) ). This demonstrates its ability to distinguish weakened cardiac reflections from surrounding noise by enforcing spatial-temporal continuity. FIGURE 7. Open in a new tab Validation of PTL path for reflection localization across range bins. The PTL Path (white dashed line) demonstrates: (a) consistent localization within a single stable range bin for healthy subjects, and (b) dynamic localization across multiple unstable bins for AF patients. Second, we validate that these samples correspond to true cardiac reflections by analyzing the beat-to-beat temporal alignment between ECG R-peaks and PTL waveform peaks as shown in Fig. 8 . The PTL waveform is generated by extracting the second-order phase from the reflection samples, followed by bandpass filtering (0.8–2.0 Hz). ECG R-peaks are detected using a standard algorithm [62] and treated as discrete ventricular depolarization timestamps. For each R-peak, the nearest PTL peak within a window of −50 to + 200 ms is identified. We specifically used this window to capture the delay between ECG R-peaks and PTL peaks closely, representing the mechanical response following electrical ventricular depolarization. This serves as a temporal benchmark to distinguish valid cardiac reflection samples from asynchronous noise. FIGURE 8. Open in a new tab Validation of PTL-located reflections against true cardiac events through temporal alignment between ECG R-peaks and PTL waveform peaks. (a) Healthy subjects exhibit high temporal alignment with low, consistent delay. (b) AF patients demonstrate lower alignment with higher, more variable delays. To quantify the validation of PTL, we define two metrics: (i) Delay , the temporal difference between the ECG R-peak and the corresponding PTL peak, defined as: where is the timestamp of the -th PTL peak, is the -th ECG R-peak timestamp, represents the hardware synchronization offset, and (ii) Match Ratio , the percentage of R-peaks with a detected PTL peak, defined as: where is the number of R-peaks with a corresponding PTL peak identified within the −50 to + 200 ms window, is the total number of detected ECG R-peaks. We compute these metrics across a group of 4 healthy subjects and 4 AF patients as summarized in Table 1 . PTL achieved a high Match Ratio of 0.940 for healthy subjects and 0.857 for AF patients. These high ratios validate that the samples located by PTL consistently correspond to true cardiac events rather than random noise. The Delay analysis further confirms the temporal consistency of the located samples. Healthy subjects exhibited a mean delay of 67.49 ms ( ms), while AF patients showed a significantly higher mean delay of 158.43 ms ( ms). This increased delay in AF patients is a pure representation of the physiological temporal misalignment identified in Challenge C2, which is subsequently addressed by our learning component (HPR-Net.). These results are visually supported by the representative temporal alignments in Fig. 8 . In the healthy subject 8(a) , PTL peaks (green) consistently follow R-peaks (red) with low jitter. In contrast, the AF patient 8(b) exhibits the increased and more variable delay.This alignment across both groups validates that PTL successfully identifies and extracts reflection samples corresponding to true cardiac events. TABLE 1. PTL Validation Results: Group Analysis of Healthy and AF Subjects. Subject Group (n=4 each) Mean Delay (ms) Delay STD (ms) Match Ratio(%) Total Beats Healthy 67.49 129.49 0.940 296 AF 158.43 163.92 0.857 269 Open in a new tab 3). Cardiac Motion Feature Extraction This component transforms the reflection samples selected by PTL into cardiac motion features suitable for HPW reconstruction. We first extract the phase and magnitude signals from the reflection samples, which represent the underlying chest wall vibration, respectively. However, these initial signals contain various physiological motions, including respiration, heartbeat, and other body motions. To isolate the relevant cardiac motion, we apply a bandpass filter (0.2–50 Hz) as a clutter filter to remove low-frequency baseline drift and high-frequency noise. A subsequent narrow-band filter (0.8–2.0 Hz) is then applied to the phase to target the specific cardiac frequency range. While these filters remove general noise and focus on cardiac frequencies, isolating cardiac motion from respiration remains critical as respiration harmonics can still persist. Chest wall vibration caused by respiration is slow with low acceleration, whereas cardiac activity, such as heartbeats, induces significant acceleration. We leverage this fundamental difference in acceleration to distinguish cardiac motion. Therefore, we compute the chest acceleration as the second-order derivative [63] of the chest vibration, which serves as a refined cardiac motion feature that selectively enhances the rapid, high-acceleration components indicative of heartbeats. This filter is defined by equation 4 , where represents the phase at time , and denotes the sampling interval. Finally, this component provides three distinct processed output, the bandpass-filtered phase and magnitude, and the second-order derivative-filtered phase—for the subsequent HPR-Net. D. Learning-Based Component 1). Heart Pulse Reconstruction Network (HPR-Net) HPR-Net reconstructs HPWs as a series of Gaussian pulse trains. Its end-to-end neural network utilizes phase, the second-order derivative of phase, and magnitude as inputs to extract cardiac motion features for reconstructing HPWs, as shown in Fig. 9 . Unlike the baseline that interprets features from temporally aligned reflection samples concentrated to a single stable range bin, HPR-Net is designed to interpret features from the localized multi-bin stream provided by the PTL algorithm (1). While PTL provides some stabilization by consistently selecting the strongest localized reflection at each timestep, the resultant stream still exhibits significant temporal misalignment due to AF. FIGURE 9. Open in a new tab Example input and output of heart pulse reconstruction network (HPR-Net). To intelligently process this multi-bin input, we employ an attention mechanism inspired by GAT [64] . The GAT is essential for robust feature extraction amidst the temporal misalignment of the reflection samples (C2). It adaptively weights the contribution of each range bin at every timestep, allowing the network to robustly focus on the true cardiac motion, even when the magnitude of the motion features shifts between the localized bins. This mechanism captures inter-bin correlations at each timestep, enabling the network to learn robust inter-bin cardiac motion features. These features are then transformed by the subsequent encoder-decoder architecture to derive a cardiac motion latent representation that ultimately leads to the construction of HPWs. Fig. 11 provides an overview of the HPR-Net framework and lists key hyperparameters (e.g., kernel sizes and output shapes) for each layer. The framework comprises three modules: (1) Heart Signal Extractor, (2) Encoder-Decoder, and (3) Reconstructor. Detailed descriptions of each module follow below. FIGURE 11. Open in a new tab Architectural overview of the proposed heart pulse reconstruction network (HPR-Net). The kernel size of each residual block is denoted by k, and each layer is annotated with the shape of its output. a). Heart Signal Extractor: This module is designed to extract inter-bin cardiac motion features from the inputs. We begin by employing one-dimensional convolutional operations to capture the temporal information within each range bin. To refine the extraction process, we utilize a variant of residual blocks, as depicted in Fig. 10 . These residual blocks are applied across all range bins selected through PTL, refining the initial features and improving performance and learning efficiency [65] . Specifically, each residual block contains two 1D convolution layers (kernel size , stride 1, and zero padding to preserve temporal length), each followed by normalization and a LeakyReLU nonlinearity. When the number of channels changes, the skip path uses a convolution to match dimensions. Applying these blocks sequentially to each bin yields per-bin embeddings of shape . FIGURE 10. Open in a new tab HPR-Net building blocks: (a) residual block and (b) transposed residual block. Subsequently, a GAT mechanism [66] is implemented to model correlations between the range bins. Unlike traditional convolutional networks [67] , [68] , which assume local spatial correlations, GAT effectively detects relationships between non-adjacent range bins. This capability allows for the accurate detection of patterns within the input features that indicate cardiac motion, thereby facilitating the extraction of relevant inter-bin cardiac motion features. Notably, we apply GAT independently at each time step on a graph whose nodes correspond to the PTL-selected range bins. We use a fully-connected topology with self-loops (i.e., every bin can attend to every other bin at the same ). The node feature for bin at time is the 64-D vector produced by the last residual block, and the GAT linearly projects it to channels. Specifically, at each time step, the GAT block processes a set of temporal features from the range bins, denoted as for , where is the dimension of the input features resulting from the convolutional operations and represents the number of range bins. The learned representation of each range bin’s features, , where is the length of the output features, is computed as: Here, is a learnable weight matrix, and denotes the activation function. We employ the Leaky ReLU function with a negative slope of 0.2 as the activation function. The attention coefficient of with respect to is calculated by: where is a learnable weight vector, denotes vector concatenation, and specifies the element-wise Leaky ReLU function with a negative slope of 0.2. The GAT output preserves the bin dimension ( ). We then apply a global max-pooling over the bins at each time step to obtain a single inter-bin feature vector per frame, resulting in a sequence of shape that is passed to the encoder-decoder. This pooling step makes the downstream modules invariant to the ordering of the selected bins while retaining the most salient features across bins. b). Encoder-Decoder: The encoder module employs a convolutional encoder-decoder (CED) network to construct a cardiac motion latent representation from the extracted inter-bin cardiac motion features. CED networks are known for their efficacy in time series analysis, anomaly detection, and the compression and reconstruction of physiological data, such as ECG [69] , [70] , [71] . These networks learn an efficient latent representation of data by transforming the input into a lower-dimensional latent space and then decoding this compressed representation to recover the original data. As shown in Fig. 11 , our CED operates on the sequence and uses two temporal downsampling steps followed by two symmetric upsampling steps. The encoder component of the network employs a stack of residual blocks to capture temporal information from the inter-bin cardiac motion features. These residual blocks facilitate the extraction of pertinent temporal characteristics. The first residual block uses and outputs , followed by 1D max-pooling with pool size 2 (stride 2), yielding . A second residual block (again ) increases the channel width to 64 and outputs , followed by another max-pooling layer producing . In contrast, the decoder component employs transposed residual blocks, as shown in Fig. 10 , to reconstruct the cardiac motion latent representation from the encoded latent space. Each upsampling step increases the temporal resolution by a factor of 2 (nearest-neighbor upsampling), and each transposed residual block uses convolutions to refine features at the higher resolution. This encoder-decoder design is essential because it learns a non-linear transformation that interprets the temporally inconsistent features provided by the GAT and maps them to a smooth, high-level latent representation that restores the underlying temporal order of the cardiac motion. This enables the effective formation of a high-level cardiac motion latent representation. c). Reconstructor: The reconstructor module employs a bidirectional long short-term memory (BiLSTM) network to generate the reconstructed HPW from the cardiac motion latent representation. Long short-term memory (LSTM) networks are well-suited for modeling long-range dependencies in time series data, and their bidirectional variant (BiLSTM) further enhances performance by processing the input sequence in both forward and backward directions [72] , [73] . In our approach, the cardiac motion latent representation produced by the encoder-decoder module at each time step is fed into the BiLSTM network, as depicted in Fig. 11 . The BiLSTM captures temporal dependencies and provides a comprehensive understanding of the cardiac motion dynamics, capturing both short-term and long-term relationships in cardiac motion. We set the BiLSTM output dimensionality to 32 (i.e., 16 hidden units per direction), producing a sequence of shape . We then apply time-distributed fully-connected layers with widths ; a LeakyReLU with slope 0.2 follows the first dense layer, and dropout with rate 0.2 is applied after the second dense layer to reduce overfitting. Finally, a sigmoid activation produces the normalized HPW amplitude at each time step, resulting in an output of shape , interpreted as the reconstructed HPW. 2). Heart Health Analysis This component performs both continuous monitoring and diagnosis based on the reconstructed HPWs. For monitoring, we use a peak detection algorithm [62] to identify the peaks of the HPWs to estimate HR and RR intervals. The diagnosis model utilizes heart rate variability (HRV) metrics derived from the RR intervals as its input feature set. HRV analysis serves as an effective indicator of cardiac autonomic function; studies often link changes in these metrics to arrhythmic events [74] and autonomic stress responses [75] . Our feature set focuses on six widely accepted time-domain HRV metrics, formally defined based on the RR interval series (NN intervals) [62] , [76] , [77] : mean of RR intervals (MeanNN), median of RR intervals (MedianNN), standard deviation of RR intervals (SDNN), interquartile range of RR intervals (IQRNN), median absolute deviation of RR intervals (MadNN), and the ratio of MadNN to MedianNN. These metrics are extracted from both the ground-truth ECG readings and the reconstructed HPWs. For diagnosis, we employ a random forest classifier [78] , [79] as our AF prediction model. This classifier takes the HRV metrics defined above, derived from both the actual ECG readings and the reconstructed HPWs, as its input feature set to diagnose AF. This approach enables the classifier to utilize information on cardiac variability from both real cardiac activity (ECG) and simulated contactless activity (HPWs), enhancing its ability to differentiate between healthy and AF conditions. We detail the training and performance of this diagnostic model in Sections IV and V . IV. Implementation and Experimental Details A. Hardware and Software Toolkits We used a texas instruments millimeter wave board (AWR1642BOOST) with the DCA1000 real-time data-capture adapter for precise radar sensing [45] , [80] . The board operates in the 77–81 GHz range. We configured the system with 2 transmitter (1Tx) and 4 receivers (4Rx) to enhance spatial resolution while maintaining high temporal resolution through a short chirp period ( s) and a high sampling rate (6000 ksps). Each chirp was followed by an idle time of s, with 256 samples captured per chirp. The receive gain was set to 30 dB, and data acquisition was managed through mmWave Studio [81] , which streamed the data to a host PC for processing. Simultaneously, we recorded ECG signals using the shimmer 3TM ECG development kit [82] at a sampling rate of 500 Hz. Radar signal processing was implemented using C/C++, while neural network components were developed in python with tensorflow 2.10 [83] on a workstation with an Intel i7 CPU, 32GB DDR4 RAM, and a NVIDIA GeForce RTX 2070 graphics card [84] . B. Participant We collected data from 210 participants (ages 32–86; 5 female), comprising 108 healthy subjects from the general population (ages 34–53; 6 female) and 102 patients diagnosed with atrial fibrillation (AF) (ages 40–86). Data from healthy subjects were collected at our institution (KAIST), Korea, while data from AF patients were collected in collaboration with a renowned cardiologist from Seoul National University Bundang Hospital, Korea. Clinical comorbidities among AF patients included hypertension, diabetes mellitus, and dyslipidemia, as documented in hospital records. The mean body mass index (BMI) of AF patients was kg/m 2 . The study was conducted under university Institutional Review Board (IRB) approval (IRB No. B-2210-789-303), ensuring adherence to ethical guidelines. All participants were briefed on the study objectives, provided written informed consent prior to participation, and were compensated for their time with a gift voucher (100,000 KRW, equivalent to USD). They were instructed to minimize voluntary movement during recording sessions to ensure consistent data acquisition. For AF patients, ECG recordings acquired during radar measurements were reviewed by a cardiologist to confirm the presence of AF activity during the recording session. C. Data Collection Procedure The data collection process involved the simultaneous acquisition of radar and ECG recordings from each participant. Participants were seated upright on a chair during data acquisition, facing the radar, and instructed to breathe normally. The radar was positioned in front of the subject approximately perpendicular to the subject’s anterior chest wall (near-normal incidence, ), facing the cardiac apex, at a distance of 25 to 55 cm. Small variations in angle across subjects were unavoidable due to minor individual posture differences (e.g., comfort-related adjustments); however, the radar was consistently oriented to maximize sensitivity to chest wall motion associated with cardiac activity. Data collection was conducted in a typical office environment that included standard furniture and ambient wireless signals such as WiFi, LTE, and Bluetooth, simulating real-life conditions. No large moving objects or reflective surfaces were present within the radar field of view other than the seated participant. Representative experimental setups are shown in Figure 12 . At the same time, ECG data was collected using the Shimmer ECG device, with electrodes placed on four locations: the left arm (LA), right arm (RA), left leg (LL), and right leg (RL). FIGURE 12. Open in a new tab Experimental setups for data collection. (a) Setup at KAIST. (b) Setup at SNU Hospital. A total of 210 trials were conducted, each lasting 60 seconds, resulting in a dataset of approximately 20,000 heartbeats. The recorded heart rates ranged from 63 to 86 beats per minute (bpm) for healthy subjects and from 64 to 94 bpm for AF patients. To ensure the robust performance of the baseline, baseline+PTL, and mCardiacDx systems in reconstructing HPWs and diagnosing AF, we randomly partitioned the dataset of 210 subjects (108 healthy subjects and 102 AF patients) into training, validation, and test sets, following a 60/24/24 split for healthy subjects and 56/22/24 for AF patients. D. Network and Model Training We detail the network and model training below: • HPR-Net Training: The neural networks for the baseline, baseline+PTL, and mCardiacDx for HPW reconstruction were trained on the training set for 300 epochs using an Adam optimizer. For the baseline and baseline+PTL models, training was performed using a batch size of 16 and an initial learning rate of 0.001. Similarly, the mCardiacDx model was trained using a batch size of 16 and an initial learning rate of 0.0005. • Diagnosis: The random forest model was trained on a combined feature set of HRV metrics. This set included metrics derived from ground truth ECG. These ECG-derived metrics were augmented by additional similar HRV metrics derived from the HPWs generated by the three systems (baseline, baseline+PTL, and mCardiacDx), as illustrated in Figure 13 . This approach enables the model to learn directly from both original and system-augmented metrics, which enhance its foundation for robust AF diagnosis. For testing, model extracts similar metrics derived from the HPWs of the corresponding system. FIGURE 13. Open in a new tab Overview of the dataset split for diagnosis model training and testing. V. Performance Results In this section, we evaluate the performance of baseline+PTL and mCardiacDx in comparison to the baseline using the evaluation metrics outlined in section V-A . Our analysis focuses on both healthy subjects and AF patients, emphasizing three main areas: the quality of HPWs reconstruction, the monitoring of HR and RR intervals, and the diagnosis of AF. A. Evaluation Metrics We employed the following set of quantitative metrics [46] , [47] , [85] for each evaluation stage: 1). HPW Reconstruction Quality • Dynamic Time Warping (DTW): DTW [46] measures the similarity between two-time series that may vary in speed or timing. For HPW reconstruction, we use DTW to quantify the similarity between systems HPWs and ECG-generated HPWs. Lower DTW scores indicate higher similarity and better reconstruction. 2). Heart Rate(HR) and RR Interval Estimation Accuracy • Median Absolute Percentage Error (MedAPE): MedAPE [47] quantifies the median absolute percentage difference between estimated values (HR and RR intervals from systems) and ECG. Lower MedAPE values signify higher estimation accuracy. • Pearson Correlation Coefficient ( ): It measures the strength of the linear relationship between estimated values (HR and RR intervals from systems) and ECG. Values closer to 1 indicate stronger linear agreement. • Bland–Altman Analysis: It evaluates agreement between estimated values (HR and RR intervals from systems) and ECG by quantifying the mean bias and limits of agreement. Lower bias and narrower limits indicate better consistency. 3). Atrial Fibrillation Diagnosis Performance For evaluating the diagnostic capabilities, we use standard binary classification metrics [85] derived from the confusion matrix, where the AF is considered the positive class and healthy is the negative class: • Accuracy: Proportion of all correctly classified instances, defined as . • Precision: Proportion of true positive predictions among all predictions. It measures the model’s ability to avoid false positives, defined as . • Recall (Sensitivity): Proportion of true positive predictions among all actual positive instances. It measures the model’s ability to identify all relevant cases, avoiding false negatives, defined as . • Specificity: Proportion of true negative predictions among all actual negative instances. It measures the model’s ability to identify all relevant negative cases, avoiding false positives, defined as . • F1-score: The harmonic mean of precision and recall, balancing both. It is a key metric in AF diagnosis, calculated as . • Receiver Operating Characteristic (ROC): Measures the ability of a classifier to distinguish classes across various threshold settings. A ROC value of 1.0 represents a perfect classifier, while 0.5 indicates a random chance. B. Reconstructing the Heart Pulse Waveform We evaluate the performance of our proposed methods, baseline+PTL and the mCardiacDx system, in reconstructing heart pulse waveforms (HPWs) and compare them against the baseline. The primary goal is to evaluate their ability to overcome the baseline’s previously identified failure in reconstructing HPWs, which, as shown, leads to significant errors in HR and RR interval estimation. A qualitative analysis, shown in Figure 14 , illustrates the performance of each method. For healthy subjects, all three approaches successfully reconstruct HPWs that align well with the ECG-derived HPWs ( Figure 14(a) ). However, a critical distinction appears in AF cases. As established previously, the baseline method produces distorted HPWs, reduced in number and poorly aligned with the ECG-derived HPWs (as highlighted in red). As shown in Figure 14(b) , both baseline+PTL and mCardiacDx successfully reconstruct the HPWs for AF patients. Notably, mCardiacDx achieves the highest fidelity, generating the correct number of HPWs and exhibiting the closest alignment to the ECG-derived HPWs, surpassing the improvement offered by PTL alone. FIGURE 14. Open in a new tab Performance comparison of baseline, baseline+PTL, and mCardiacDx HPW reconstruction against ECG-derived HPWs across healthy subjects and AF patients. To quantitatively validate these findings, we compare the reconstructed HPWs directly with the ECG-derived HPWs. For this, we use dynamic time warping (DTW), a metric ideal for measuring the similarity between time-series data that may have timing variations. A lower DTW score indicates that the reconstructed HPW is more similar to the ECG-derived HPWs. The results shown in Figure 15 confirm the superiority of our methods. For AF patients, the baseline exhibits a high DTW score of 5.92, reflecting poor performance. The introduction of our PTL technique (baseline+PTL) significantly reduces this score to 3.78. Our complete mCardiacDx system achieved the best score of 2.92, indicating the most accurate waveform reconstruction. This trend also holds for healthy subjects, where mCardiacDx (2.82) and baseline+PTL (3.90) outperform the baseline (4.02). Collectively, these results demonstrate that both PTL and the mCardiacDx system effectively address the baseline’s limitations, enabling the accurate reconstruction of HPWs that faithfully match the ECG-derived HPWs, even in challenging AF cases. FIGURE 15. Open in a new tab Dynamic time warping (DTW) scores of mCardiacDx and baseline+PTL compared to the baseline across healthy subjects and AF patients. C. Monitoring Atrial Fibrillation We evaluate the monitoring performance of baseline+PTL and mCardiacDx in estimating HR and RR intervals based on the reconstructed HPWs. We measure their performance using the MedAPE, calculated relative to ground truth ECG, and compare it against the baseline method. According to the association for the advancement of medical instrumentation clinical guidelines [86] , the acceptable HR threshold should not exceed ±10% or ±5 bpm, whichever is greater, when measured relative to the ground truth. Some studies adopt a more conservative threshold of a mean absolute percentage error (MAPE) of % [87] , [88] , while others report a MAPE range of 3–6% [89] . As a smaller error indicates enhanced estimation accuracy, we evaluate system performance against a MedAPE threshold of 6%. Notably, unlike HR, there is no established clinical threshold for RR interval error; thus, the same MedAPE criterion is applied for consistency. Figure 16 illustrates that both baseline+PTL and mCardiacDx consistently achieve lower MedAPE error rates for HR and RR intervals estimation across both healthy and AF subjects. For the healthy subjects, baseline+PTL achieves error rates of 2.63% and 2.70% for HR and RR intervals, showing a reduction compared to baseline errors of 2.66% and 2.73%, respectively. mCardiacDx further reduces these errors to 1.68% and 1.71% for HR and RR intervals. For the AF patients, baseline+PTL yielded error rates of 4.95% and 5.21% for HR and RR intervals, which are significantly lower than the baseline errors of 9.10% and 8.42%. mCardiacDx further reduced these errors to 2.94% and 2.95% for HR and RR intervals. These results indicate that both baseline+PTL and mCardiacDx are robust in monitoring AF, with mCardiacDx achieving the best results. FIGURE 16. Open in a new tab Monitoring performance (MedAPE) of mCardiacDx and baseline+PTL compared to the baseline across healthy subjects and AF patients. 1). Correlation Analysis We further evaluate the performance of baseline+PTL and mCardiacDx by analyzing the linear relationship between the estimated HR, RR intervals and the ECG. We measure this using Pearson correlation analysis [90] to calculate the correlation coefficient (r) and compare the results with the baseline. Figures 17 and 18 show the correlation plots of the three systems for HR and RR intervals, respectively, for both the healthy subjects and the AF patients. For the healthy subjects, the baseline method achieves correlation coefficients of for HR and for RR intervals. The integration of the PTL technique (baseline+PTL) improves the correlation to for HR and for RR intervals. The mCardiacDx system achieves the highest correlation for the healthy subjects, with coefficients of for HR and for RR intervals. FIGURE 17. Open in a new tab Pearson correlation analysis between ECG-derived HR and estimated HR for the baseline, baseline+PTL, and mCardiacDx across healthy subjects and AF patients. FIGURE 18. Open in a new tab Pearson correlation analysis between ECG-derived RR intervals and estimated RR intervals for the baseline, baseline+PTL, and mCardiacDx across healthy subjects and AF patients. For the AF subjects, the baseline shows considerably weaker correlation with the ECG, yielding coefficients of for HR and for RR intervals. The introduction of PTL (baseline+PTL) significantly improves the correlation to for HR and for RR intervals. The mCardiacDx further improves the correlation, achieving for HR and for RR intervals. These results demonstrate that baseline+PTL and mCardiacDx provide more consistent HR and RR intervals relative to the ECG in AF patients, with mCardiacDx achieving the strongest linear correlation, while maintaining high correlation in healthy subjects. 2). Agreement Analysis (Bland-Altman) While correlation measures the strength of a linear relationship, it does not evaluate the agreement between two measurement methods. To evaluate the clinical agreement of baseline+PTL and mCardiacDx with the ECG, we performed Bland-Altman [91] analysis on the estimated HR, RR intervals and compare the results against the baseline. Figures 19 and 20 present the Bland-Altman plots of the three systems for HR and RR intervals, respectively, for both the healthy subjects and the AF patients. As illustrated in these figures, both baseline+PTL and mCardiacDx demonstrate narrower 95 % limits of agreement (LoA) and lower mean bias compared to the baseline method. For the healthy subjects, the baseline method shows a mean bias of −2.34 bpm (SD: 4.09 bpm) for HR and 0.021 s (SD: 0.041 s) for RR intervals. The integration of the PTL technique (baseline+PTL) reduces the bias to −0.09 bpm (SD: 2.72 bpm) for HR and to −0.002 s (SD: 0.028 s) for RR intervals. mCardiacDx further improves these results, showing a highly stable agreement with a mean bias of −0.49 bpm (SD: 2.36 bpm) for HR and a bias of only 0.002 s (SD: 0.028 s) for RR intervals. FIGURE 19. Open in a new tab Bland-Altman agreement analysis between ECG-derived HR and estimated HR for the baseline, baseline+PTL, and mCardiacDx across healthy subjects and AF patients.The plots show the mean bias and 95% LoA. FIGURE 20. Open in a new tab Bland-Altman agreement analysis between ECG-derived RR intervals and estimated RR intervals for the baseline, baseline+PTL, and mCardiacDx across healthy subjects and AF patients. The plots show the mean bias and 95% LoA. The improvement is more profound in the AF patients. The baseline method exhibits a high mean bias of −4.89 bpm (SD: 9.21 bpm) for HR and 0.065 s (SD: 0.063 s) for RR intervals. The integration of PTL (baseline+PTL) successfully mitigates these errors, reducing the bias to 2.43 bpm (SD: 5.07 bpm) for HR and −0.029 s (SD: 0.056 s) for RR intervals. mCardiacDx achieves the highest agreement for the AF patients, yielding a negligible mean bias of 0.44 bpm (SD: 3.30 bpm) for HR and −0.008 s (SD: 0.035 s) for RR intervals. These results confirm that baseline+PTL and mCardiacDx achieve superior agreement with the ECG in AF patients, with mCardiacDx providing the lowest bias and narrowest LOA, while maintaining consistent performance in healthy subjects. D. Diagnosing Atrial Fibrillation We finally evaluate the diagnostic capabilities of baseline+ PTL and mCardiacDx, comparing their performance to the baseline in diagnosing AF. While clinicians utilize the term ‘diagnosis’ for a full medical evaluation, within our work, it refers to detection using a classifier. The diagnostic performance is visualized using confusion matrices and ROC curves ( Figs. 21 – 23 ). In these confusion matrices, the color intensity corresponds to the absolute number of subjects in each cell, while the numerical annotations denote the class-wise recall (percentage) and the raw subject counts. The findings are presented in the subsequent subsections. FIGURE 21. Open in a new tab Diagnosis performance of the baseline. (a) Confusion matrix (b) ROC curves. FIGURE 22. Open in a new tab Diagnosis performance of the baseline+PTL. (a) Confusion matrix (b) ROC curves. FIGURE 23. Open in a new tab Diagnosis performance of the mCardiacDx. (a) Confusion matrix (b) ROC curves. 1). Baseline It demonstrates considerable diagnostic performance for healthy subjects, achieving a precision of 0.94 with 1 false positive and a specificity of 0.96. However, its recall was limited to 0.7500, resulting in 6 false negatives among the 24 AF patients, indicating significant challenges in diagnosing AF cases. While it correctly identifies 23 out of 24 healthy cases and an overall accuracy of 0.8500, the F1-score of 0.83 reflects a moderate balance between precision and recall. The ROC of 0.97 (actual 0.9661) indicates good discriminative ability, though the model’s limitations in diagnosing AF patients prompt further improvements through PTL. The results are illustrated in the confusion matrix and ROC curve plots in Figure 21(a) and (b) . 2). Baseline+PTL The baseline+PTL shows improvements over the baseline. Its precision and recall increase to 0.95 and 0.83, respectively, with 1 false positive consistent with the baseline. It diagnoses 20 true positives and reduces false negatives to 4, demonstrating improved sensitivity while maintaining the 0.96 specificity of the baseline. While correctly identifying 23 out of 24 true negatives, it achieves an improved F1-score of 0.88, indicating better balance between precision and recall. The overall accuracy increases to 0.89 and the ROC reaches 0.97 (actual 0.9679), demonstrating strong discriminative capacity. Despite these advancements, the remaining false negatives indicate room for improvement, which is addressed by our system mCardiacDx. Nonetheless, baseline+PTL proves to be a more effective solution for diagnosing AF compared to the baseline. The results are shown in the confusion matrix and ROC curve plots in Figure 22(a) and (b) . 3). MCardiacDx The mCardiacDx further advances the diagnostic capabilities, achieving a precision of 0.95 with 1 false positive consistent with the baseline. Its recall improves significantly to 0.91, diagnosing 22 true positives with only 2 false negatives, thus outperforming both the baseline and baseline+PTL in sensitivity. While correctly identifying 23 out of 24 true negatives to achieve a specificity of 0.96, mCardiacDx demonstrates superior performance with an F1-score of 0.93 and overall accuracy of 0.93. The ROC improves to 0.98, affirming its outstanding discriminative capability. This performance positions mCardiacDx as a leading solution for contactless AF monitoring. The results are shown in the confusion matrix and ROC curve plots in Figure 23(a) and (b) . VI. Related Work In this section, we review existing studies on monitoring heart activity, with a particular focus on arrhythmia. We categorize these approaches into two main categories: wearable cardiac sensing and contactless cardiac sensing . A. Wearable Cardiac Sensing Wearable cardiac sensing primarily employs two key modalities: electrocardiography (ECG) [11] , [15] , [16] , [17] , [20] , [92] , [93] and photoplethysmography (PPG) [21] , [22] , [23] . ECG has traditionally been regarded as the gold standard for arrhythmia diagnosis in clinical environments such as hospitals [11] , [92] , [93] . However, its application for continuous monitoring outside these settings has been challenging due to the requirement for medical expertise. To address this challenge, recent advancements have led to the creation of portable ECG devices, including Holter monitors [13] , [14] , [15] and ECG patches [16] , [17] . While these devices enhance convenience, they require the attachment of electrodes to the skin for longer duration, which can disrupt daily activities like sleeping or showering. Beside, the electrode attachment makes them unsuitable for certain populations, such as newborns, the elderly, and patients with skin burns. Newly released single-lead ECG sensors integrated into smartwatches provide another option but require active user interaction, specifically touching an electrode button during monitoring, which prevents their use for continuous monitoring [20] . In addition to ECG, PPG-based wearables offer a convenient alternative for arrhythmia detection [21] , [22] , [23] . Despite this convenience, long-term use can cause discomfort and skin allergies. A study found that 88 % of users remove their watches before bed [24] . B. Contactless Cardiac Sensing Contactless cardiac sensing has emerged as an alternative to traditional wearable sensing by leveraging radio frequency (RF) signals, such as WiFi [25] , [26] , RFID [27] , ultra-wideband (UWB) [28] , and mmWave radar [29] , [30] , to monitor heart rate (HR) and heart rate variability (HRV). These systems detect thoracic wall vibrations to infer mechanical motions associated with the cardiac cycle. Research has extended contactless sensing to support applications like emotion recognition [30] and stress detection [94] . Notably, advancements have focused on reconstructing cardiac waveforms, such as seismocardiograms (SCG), which reflect mechanical activity of atrial and ventricular contraction. For example, radar systems enhanced with multichannel beamforming and one-dimensional 1-D CNNs have been used to extract SCG features [34] . Recently, research has shifted toward reconstructing electrocardiogram (ECG) waveforms from mmWave reflections using generative models [31] , [32] , [33] . CardiacWave [32] introduced an attention-augmented LSTM-based CaSE-ECG solver to recover ECG-like signals, whereas RF-ECG [33] used a conditional generative adversarial network (cGAN) to synthesize ECG waveforms. In parallel, other studies have explored the reconstruction of photoplethysmogram (PPG) waveforms from mmWave signals to infer vascular and respiratory dynamics [35] , [36] . However, these studies primarily focused on healthy subjects with normal heartbeats. More recently, studies have focused on patients with abnormal heartbeats exhibiting asynchronous and irregular heart contractions. Zhao et al. [37] showed that asynchronous heart contractions in atrial fibrillation (AF) create a challenge in cardiac waveform generation. Acknowledging this, the authors proposed a classification-based approach to identify arrhythmia based on radar signal patterns. Although innovative, this method produces a ‘black box’ output that fails to provide crucial temporal features required for diagnosis. In a related study, the same author focused on reconstructing cardiac waveforms, particularly ECGs, using a cross-domain diffusion model [38] . However, this method requires calibration with a traditional ECG, undermining its fully contactless claim, and the outputs lack necessary fidelity for diagnosis. Its performance declines in unseen patients, indicating limited generalization beyond training data. Similarly, Yuan et al. [39] introduced an AF monitoring system employing a knowledge transfer approach using a teacher-student model pretrained on contact-based ECG data. While innovative, the mapping between radar-based features and true cardiac activity remains ambiguous, limiting contactless feature interpretability and confidence in diagnosis. Distinct from prior work, mCardiacDx and Baseline+PTL facilitate the monitoring and diagnosis of AF by contactless creation of cardiac waveforms in the form of HPWs, which provide the temporal features required for clinical diagnosis—all without relying on calibration with a traditional ECG. More specifically, our system is designed to perform the monitoring and diagnosis of AF on real patients in practical real-world seated scenarios, a critical advancement over existing methods. VII. Limitations and Future Work While the proposed mCardiacDx system and PTL technique show promise in monitoring and diagnosing AF, the current implementation is limited to patients in quiescent or minimally-moving states (e.g., seated, as evaluated). This constraint poses a significant challenge for accurately monitoring patients during dynamic motions such as walking on the treadmill, cycling. Moreover, the system currently focuses on reconstructing HPWs, which, although effective, provide limited insight into detailed cardiac dynamics. Future work will aim to extend the system and technique to support monitoring and diagnosis under motion. This includes developing motion-robust signal processing techniques to manage distortion and variability introduced by movement (walking on the treadmill, cycling, etc). Furthermore, we plan to extend capability of mCardiacDx to reconstruct detailed cardiac cycles, enabling analysis comparable to conventional ECG in terms of atrial and ventricular activity. Another important direction involves investigating the relationship between spatially dispersed and temporally misaligned reflections characteristics and AF severity. VIII. Conclusion In this paper, we present mCardiacDx, a radar-driven contactless system designed for AF monitoring and diagnosis. The system leverages a PTL technique to analyze complex reflections and reconstruct HPWs for estimating key cardiac rhythm metrics (HR and RR interval). Evaluation results demonstrate that mCardiacDx, powered by PTL, outperforms state-of-the-art work in monitoring and diagnosing AF. mCardiacDx is a robust contactless system for pervasive and continuous cardiac care in real-world seated scenarios. Acknowledgment The authors sincerely thank the anonymous reviewers for their valuable comments, which greatly improved the quality of this manuscript. Funding Statement This work was supported in part by the Bio and Medical Technology Development Program of the National Research Foundation (NRF) funded by Korean Government through Ministry of Science and ICT (MSIT) under Grant RS-2023-00222910 and in part by the NRF funded by Korean Government through MSIT/Ministry of Education (MOE) under Grant RS-2024-00347516. References [1]. Cardiovascular Diseases (cvds), World Health Organization, Geneva, Switzerland, 2024. [ Google Scholar ] [2]. Atrial Fibrillation Infographic, World Heart Federation, Geneva, Switzerland, 2024. [ Google Scholar ] [3]. World Heart Report 2023, World Heart Federation, Geneva, Switzerland, 2023. [ Google Scholar ] [4]. Xiao Y.-F., “Cardiac arrhythmia and heart failure: From bench to bedside,” J. Geriatric Cardiol., vol. 8, no. 3, pp. 131–132, Oct. 2011. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [5]. Nesheiwat Z., Goyal A., Jagtap M., and Shammas A., “Atrial fibrillation (nursing),” StatPearls, Apr. 2023. [ PubMed ] [6]. Gorenek B., “Cardiac arrhythmias in acute coronary syndromes: Position paper from the joint EHRA, ACCA, and EAPCI task force,” EP Europace, vol. 16, no. 11, pp. 1655–1673, 2014. [ DOI ] [ PubMed ] [ Google Scholar ] [7]. Higaki T., et al. , “Asynchronous contraction of the 2 ventricles caused by ventricular pacing after a fontan-type operation in a patient with a biventricular heart,” Int. J. Cardiol., vol. 150, no. 3, pp. e116–e118, Aug. 2011. [ DOI ] [ PubMed ] [ Google Scholar ] [8]. Zhang L., Li B., and Wu L., “Heart rate variability in patients with atrial fibrillation of sinus rhythm or atrial fibrillation: Chaos or merit?,” Ann. Med., vol. 57, no. 1, Dec. 2025, Art. no. 2478474. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [9]. Arrhythmia Facts and Statistics: What You Need to Know, Arrhythmia Alliance, Stratford-upon-Avon, U.K., 2025. [ Google Scholar ] [10]. Karatela M. F. and Calkins H., “The global impact of atrial fibrillation,” Arrhythmia Electrophysiology Rev., vol. 14, p. 28, Nov. 2025. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [11]. Calkins H., et al. , “Treatment of atrial fibrillation with antiarrhythmic drugs or radiofrequency ablation: Two systematic literature reviews and meta-analyses,” Circulat., Arrhythmia Electrophysiology, vol. 2, no. 4, pp. 349–361, 2009. [ DOI ] [ PubMed ] [ Google Scholar ] [12]. López-López J. A., et al. , “Oral anticoagulants for prevention of stroke in atrial fibrillation: Systematic review, network meta-analysis, and cost effectiveness analysis,” BMJ, vol. 359, p. j5058, Nov. 2017. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [13]. Hannun A. Y., et al. , “Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network,” Nature Med., vol. 25, no. 1, pp. 65–69, Jan. 2019. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [14]. Turakhia M. P., et al. , “Diagnostic utility of a novel leadless arrhythmia monitoring device,” Amer. J. Cardiol., vol. 112, no. 4, pp. 520–524, Aug. 2013. [ DOI ] [ PubMed ] [ Google Scholar ] [15]. (2025). Holter Monitor. [Online]. Available: https://en.wikipedia.org/wiki/Holtermonitor [ Google Scholar ] [16]. Yenikomshian M., et al. , “Cardiac arrhythmia detection outcomes among patients monitored with the zio patch system: A systematic literature review,” Current Med. Res. Opinion, vol. 35, no. 10, pp. 1659–1670, Oct. 2019. [ DOI ] [ PubMed ] [ Google Scholar ] [17]. (2025). Carepulse Index—Home. [Online]. Available: https://index.carepulse.cn/home/index.html [ Google Scholar ] [18]. Almalty A., Hamed S., Jebril M., and Abdelnour H., “The effect of electrical stimulation on skin vulnerability to irritants,” Skin Res. Technol., vol. 30, no. 2, 2024, Art. no. e13591. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [19]. Nigusse A. B., Mengistie D. A., Malengier B., Tseghai G. B., and Langenhove L. V., “Wearable smart textiles for long-term electrocardiography monitoring—A review,” Sensors, vol. 21, no. 12, p. 4174, Jun. 2021. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [20]. Isakadze N. and Martin S. S., “How useful is the smartwatch ECG?,” Trends Cardiovascular Med., vol. 30, no. 7, pp. 442–448, Oct. 2020. [ DOI ] [ PubMed ] [ Google Scholar ] [21]. Guo Y., et al. , “Mobile photoplethysmographic technology to detect atrial fibrillation,” J. Amer. College Cardiol., vol. 74, no. 19, pp. 2365–2375, 2019. [ DOI ] [ PubMed ] [ Google Scholar ] [22]. Pérez M., et al. , “Large-scale assessment of a smartwatch to identify atrial fibrillation,” New England J. Med., vol. 381, no. 20, pp. 1909–1917, 2019. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [23]. Zhang H., et al. , “Towards early detection and burden estimation of atrial fibrillation in an ambulatory free-living environment,” Proc. ACM Interact., Mobile, Wearable Ubiquitous Technol., vol. 5, no. 2, pp. 1–19, Jun. 2021. [ Google Scholar ] [24]. Jeong H., Kim H., Kim R., Lee U., and Jeong Y., “Smartwatch wearing behavior analysis: A longitudinal study,” Proc. ACM Interact., Mobile, Wearable Ubiquitous Technol., vol. 1, no. 3, pp. 1–31, Sep. 2017. [ Google Scholar ] [25]. Wang X., Yang C., and Mao S., “PhaseBeat: Exploiting CSI phase data for vital sign monitoring with commodity WiFi devices,” in Proc. IEEE 37th Int. Conf. Distrib. Comput. Syst. (ICDCS), Jun. 2017, pp. 1230–1239. [ Google Scholar ] [26]. Wang X., Yang C., and Mao S., “On CSI-based vital sign monitoring using commodity WiFi,” ACM Trans. Comput. Healthcare, vol. 1, no. 3, pp. 1–27, 2020. [ Google Scholar ] [27]. Ning J., et al. , “RF-badge: Vital sign-based authentication via RFID tag array on badges,” IEEE Trans. Mobile Comput., vol. 22, no. 2, pp. 1170–1184, Feb. 2023. [ Google Scholar ] [28]. Chen Z., Zheng T., Cai C., and Luo J., “MoVi-fi: Motion-robust vital signs waveform recovery via deep interpreted RF sensing,” in Proc. 27th Annu. Int. Conf. Mobile Comput. Netw., Oct. 2021, pp. 392–405. [ Google Scholar ] [29]. Wang F., Zeng X., Wu C., Wang B., and Liu K. J. R., “MmHRV: Contactless heart rate variability monitoring using millimeter-wave radio,” IEEE Internet Things J., vol. 8, no. 22, pp. 16623–16636, Nov. 2021. [ Google Scholar ] [30]. Zhao M., Adib F., and Katabi D., “Emotion recognition using wireless signals,” in Proc. 22nd Annu. Int. Conf. Mobile Comput. Netw., Oct. 2016, pp. 95–108. [ Google Scholar ] [31]. Chen J., Zhang D., Wu Z., Zhou F., Sun Q., and Chen Y., “Contactless electrocardiogram monitoring with millimeter wave radar,” IEEE Trans. Mobile Comput., vol. 23, no. 1, pp. 270–285, Jan. 2024. [ Google Scholar ] [32]. Xu C., et al. , “CardiacWave: A mmWave-based scheme of non-contact and high-definition heart activity computing,” Proc. ACM Interact., Mobile, Wearable Ubiquitous Technol., vol. 5, no. 3, pp. 1–26, Sep. 2021. [ Google Scholar ] [33]. Wang Z., Jin B., Li S., Zhang F., and Zhang W., “ECG-grained cardiac monitoring using UWB signals,” Proc. ACM Interact., Mobile, Wearable Ubiquitous Technol., vol. 6, no. 4, pp. 1–25, Dec. 2022. [ Google Scholar ] [34]. Ha U., Assana S., and Adib F., “Contactless seismocardiography via deep learning radars,” in Proc. 26th Annu. Int. Conf. Mobile Comput. Netw., Sep. 2020, pp. 1–14. [ Google Scholar ] [35]. Zhang S., Zheng T., Chen Z., and Luo J., “Can we obtain fine-grained heartbeat waveform via contact-free RF-sensing?,” in Proc. IEEE INFOCOM Conf. Comput. Commun., May 2022, pp. 1759–1768. [ Google Scholar ] [36]. Khan U. M., Rigazio L., and Shahzad M., “Contactless monitoring of PPG using radar,” Proc. ACM Interact., Mobile, Wearable Ubiquitous Technol., vol. 6, no. 3, pp. 1–30, Sep. 2022. [ Google Scholar ] [37]. Zhao L., et al. , “MmArrhythmia: Contactless arrhythmia detection via mmWave sensing,” Proc. ACM Interact., Mobile, Wearable Ubiquitous Technol., vol. 8, no. 1, pp. 1–25, Mar. 2024. [ Google Scholar ] [38]. Zhao L., et al. , “AirECG: Contactless electrocardiogram for cardiac disease monitoring via mmWave sensing and cross-domain diffusion model,” Proc. ACM Interact., vol. 8, no. 3, pp. 1–27, 2024. [ Google Scholar ] [39]. Yuan Y., et al. , “Atrial fibrillation detection via contactless radio monitoring and knowledge transfer,” Nature Commun., vol. 16, no. 1, p. 4317, May 2025. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [40]. Khare S. and Chawala A., “Effect of change in body position on resting electrocardiogram in young healthy adults,” Nigerian J. Cardiol., vol. 13, no. 2, pp. 125–129, 2016. [ Google Scholar ] [41]. Adams M. G. and Drew B. J., “Body position effects on the ECG: Implication for ischemia monitoring,” J. Electrocardiology, vol. 30, no. 4, pp. 285–291, 1997. [ DOI ] [ PubMed ] [ Google Scholar ] [42]. Singh A., Rehman S. U., Yongchareon S., and Chong P. H. J., “Modelling of chest wall motion for cardiorespiratory activity for radar-based NCVS systems,” Sensors, vol. 20, no. 18, p. 5094, Sep. 2020. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [43]. Pänkäälä M., et al. , “Detection of atrial fibrillation with seismocardiography,” in Proc. 38th Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. (EMBC), Aug. 2016, pp. 4369–4374. [ DOI ] [ PubMed ] [ Google Scholar ] [44]. Cross-correlation, Wikipedia Contributors, San Francisco, CA, USA, 2025. [ Google Scholar ] [45]. Awr1642boost, Texas Instrum., Dallas, TX, USA, 2024. [ Google Scholar ] [46]. Senin P., “Dynamic time warping algorithm review,” Dept. Inf. Comput. Sci., Univ. Hawaii at Manoa Honolulu, Honolulu, HI, USA, Tech. Rep., 2008. [ Google Scholar ] [47]. Mdape—Median Absolute Percentage Error, Wikipedia, USA, 2025. [ Google Scholar ] [48]. Pfeiffer E. R., Tangney J. R., Omens J. H., and McCulloch A. D., “Biomechanics of cardiac electromechanical coupling and mechanoelectric feedback,” J. Biomechanical Eng., vol. 136, no. 2, Feb. 2014, Art. no. 021007. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [49]. Ripa R., George T., Shumway K. R., and Sattar Y., “Physiology, cardiac muscle,” StatPearls, USA, Tech. Rep., 2023. [ PubMed ] [ Google Scholar ] [50]. Heart Chambers, Cleveland Clinic, Cleveland, OH, USA, 2022. [ Google Scholar ] [51]. Bowman A. W. and Kovács S. J., “Assessment and consequences of the constant-volume attribute of the four-chambered heart,” Amer. J. Physiol.-Heart Circulatory Physiol., vol. 285, no. 5, pp. 2027–2033, Nov. 2003. [ DOI ] [ PubMed ] [ Google Scholar ] [52]. Maeda H., “Quantification of synchronous contraction of left ventricle in normal subjects using ECG-gated-SPECT images,” Physiological Meas., vol. 25, no. 1, pp. 71–84, Feb. 2004. [ DOI ] [ PubMed ] [ Google Scholar ] [53]. The Heart: An Owner’s Guide, National Heart, Lung, and Blood Institute, Bethesda, MD, USA, 2024. [ Google Scholar ] [54]. Prabhu S., McLellan A. J. A., Walters T. E., Sharma M., Voskoboinik A., and Kistler P. M., “Atrial structure and function and its implications for current and emerging treatments for atrial fibrillation,” Prog. Cardiovascular Diseases, vol. 58, no. 2, pp. 152–167, Sep. 2015. [ DOI ] [ PubMed ] [ Google Scholar ] [55]. Pirruccello J. P., et al. , “Deep learning of left atrial structure and function provides link to atrial fibrillation risk,” Nature Commun., vol. 15, no. 1, p. 4304, May 2024. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [56]. Li L., Zimmer V. A., Schnabel J. A., and Zhuang X., “Medical image analysis on left atrial LGE MRI for atrial fibrillation studies: A review,” Med. Image Anal., vol. 77, Apr. 2022, Art. no. 102360. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [57]. Di Biase L., et al. , “Does the left atrial appendage morphology correlate with the risk of stroke in patients with atrial fibrillation? Results from a multicenter study,” J. Amer. College Cardiol., vol. 60, no. 6, pp. 531–538, 2012. [ DOI ] [ PubMed ] [ Google Scholar ] [58]. Beinart R., et al. , “Left atrial wall thickness variability measured by CT scans in patients undergoing pulmonary vein isolation,” J. Cardiovascular Electrophysiology, vol. 22, no. 11, pp. 1232–1236, Nov. 2011. [ DOI ] [ PubMed ] [ Google Scholar ] [59]. Zhang Q., et al. , “Evaluation of left atrial volume and function using single-beat real-time three-dimensional echocardiography in atrial fibrillation patients,” BMC Med. Imag., vol. 17, no. 1, pp. 1–6, Dec. 2017. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [60]. Zheng T., Chen Z., Ding S., and Luo J., “Enhancing RF sensing with deep learning: A layered approach,” IEEE Commun. Mag., vol. 59, no. 2, pp. 70–76, Feb. 2021. [ Google Scholar ] [61]. Assana S., “Cardiovascular activity monitoring using mmWaves,” Ph.D. thesis, Massachusetts Inst. of Technol., Cambridge, MA, USA, 2020. [ Google Scholar ] [62]. Makowski D., et al. , “NeuroKit2: A Python toolbox for neurophysiological signal processing,” Behav. Res. Methods, vol. 53, no. 4, pp. 1689–1696, Aug. 2021. [ DOI ] [ PubMed ] [ Google Scholar ] [63]. Holoborodko P., “Smooth low-noise differentiators,” Independent Res. Publication, Tech. Rep., 2024. [64]. Velickovic P., Cucurull G., Casanova A., Romero A., Liò P., and Bengio Y., “Graph attention networks,” Stat, vol. 1050, no. 20, pp. 10–48550, 2017. [ Google Scholar ] [65]. He K., Zhang X., Ren S., and Sun J., “Deep residual learning for image recognition,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Jun. 2016, pp. 770–778. [ Google Scholar ] [66]. Velickovic P., Cucurull G., Casanova A., Romero A., Liò P., and Bengio Y., “Graph attention networks,” 2018. arXiv:1710.10903. [67]. Chang H.-Y., Hsu C.-H., and Chung W.-H., “Fast acquisition and accurate vital sign estimation with deep learning-aided weighted scheme using FMCW radar,” in Proc. IEEE 95th Veh. Technol. Conf., (VTC-Spring), Jun. 2022, pp. 1–6. [ Google Scholar ] [68]. Xu L., et al. , “Soli-enabled noncontact heart rate detection for sleep and meditation tracking,” Sci. Rep., vol. 13, no. 1, p. 18008, Oct. 2023. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [69]. Schlagenhauf T., Wolf J., and Puchta A., “Convolutional-based encoder–decoder network for time series anomaly detection during the milling of 16MnCr5,” Data, vol. 7, no. 12, p. 175, Dec. 2022. [ Google Scholar ] [70]. Fotiadou E., Konopczyński T., Hesser J., and Vullings R., “End-to-end trained encoder–decoder convolutional neural network for fetal electrocardiogram signal denoising,” Physiological Meas., vol. 41, no. 1, Jan. 2020, Art. no. 015005. [ DOI ] [ PubMed ] [ Google Scholar ] [71]. Llugsi R., Yacoubi S. E., Fontaine A., and Lupera P., “A novel encoder–decoder structure for time series analysis based on Bayesian uncertainty reduction,” in Proc. IEEE Latin Amer. Conf. Comput. Intell. (LA-CCI), Nov. 2021, pp. 1–6. [ Google Scholar ] [72]. Graves A. and Schmidhuber J., “Framewise phoneme classification with bidirectional LSTM networks,” in Proc. IEEE Int. Joint Conf. Neural Netw., 2005, pp. 2047–2052. [ DOI ] [ PubMed ] [ Google Scholar ] [73]. Hochreiter S. and Schmidhuber J., “Long short-term memory,” Neural Comput., vol. 9, no. 8, pp. 1735–1780, Nov. 1997. [ DOI ] [ PubMed ] [ Google Scholar ] [74]. Huikuri H. V. and Stein P. K., “Heart rate variability in risk stratification of cardiac patients,” Prog. Cardiovascular Diseases, vol. 56, no. 2, pp. 153–159, Sep. 2013. [ DOI ] [ PubMed ] [ Google Scholar ] [75]. Ma C., et al. , “Longitudinal changes and recovery in heart rate variability of young healthy subjects when exposure to a hypobaric hypoxic environment,” Frontiers Physiol., vol. 12, 2022, Art. no. 688921. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [76]. Makowski D., et al. , “Neurokit2: Heart rate variability (HRV),” GitHub, USA, Tech. Rep., 2021. [ Google Scholar ] [77]. Shaffer F. and Ginsberg J. P., “An overview of heart rate variability metrics and norms,” Frontiers Public Health, vol. 5, p. 258, Sep. 2017. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [78]. Breiman L., “Random forests,” Mach. Learn., vol. 45, no. 1, pp. 5–32, Oct. 2001. [ Google Scholar ] [79]. Yang P., Wang D., Zhao W., Fu L., Du J., and Su H., “Ensemble of kernel extreme learning machine based random forest classifiers for automatic heartbeat classification,” Biomed. Signal Process. Control, vol. 63, 2020, Art. no. 102138. [ Google Scholar ] [80]. Dca1000EVM Data Capture Card User’s Guide, Heart rhythm, Santa Clara, CA, USA, 2024. [ Google Scholar ] [81]. mmWave-studio, Texas Instrum., Dallas, TX, USA, 2024. [ Google Scholar ] [82]. Shimmer 3 ECG Unit, Shimmer Res. Ltd., Ireland, 2024. [ Google Scholar ] [83]. Abadi M., et al. , “TensorFlow: Large-scale machine learning on heterogeneous systems,” Google Inc., USA, Tech. Rep., 2015. [ Google Scholar ] [84]. NVIDIA GeForce RTX 20 Series Graphics Cards, NVIDIA Corporation, 2025. [ Google Scholar ] [85]. Novaković J. D., Veljović A., Ilić S. S., Papić Ž., and Tomović M., “Evaluation of classification models in machine learning,” Theory Appl. Math. Comput. Sci., vol. 7, no. 1, p. 39, 2017. [ Google Scholar ] [86]. Cardiac Monitors, Heart Rate Meters, and Alarms, Standard ANSI/AAMI EC13: 2002, 2002. [ Google Scholar ] [87]. Jiao Y., “Validity and challenges of optical heart rate sensors in real-time classification of exercise intensity zones,” Sensor Rev., pp. 1–14, Jan. 2026. [88]. Wang R., et al. , “Accuracy of wrist-worn heart rate monitors,” JAMA Cardiol., vol. 2, no. 1, pp. 104–106, Jan. 2017. [ DOI ] [ PubMed ] [ Google Scholar ] [89]. Nelson B. W. and Allen N. B., “Accuracy of consumer wearable heart rate measurement during an ecologically valid 24-hour period: Intraindividual validation study,” JMIR MHealth UHealth, vol. 7, no. 3, Mar. 2019, Art. no. e10828. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [90]. Schober P., Boer C., and Schwarte L. A., “Correlation coefficients: Appropriate use and interpretation,” Anesthesia Analgesia, vol. 126, no. 5, pp. 1763–1768, 2018. [ DOI ] [ PubMed ] [ Google Scholar ] [91]. Giavarina D., “Understanding bland Altman analysis,” Biochemia Medica, vol. 25, no. 2, pp. 141–151, 2015. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [92]. Electrocardiogram (ECG or EKG), Mayo Clinic, Rochester, NY, USA, 2024. [ Google Scholar ] [93]. Preejith S. P., Dhinesh R., Joseph J., and Sivaprakasam M., “Wearable ECG platform for continuous cardiac monitoring,” in Proc. 38th Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. (EMBC), Aug. 2016, pp. 623–626. [ DOI ] [ PubMed ] [ Google Scholar ] [94]. Ha U., Madani S., and Adib F., “WiStress: Contactless stress monitoring using wireless signals,” Proc. ACM Interact., Mobile, Wearable Ubiquitous Technol., vol. 5, no. 3, pp. 1–37, Sep. 2021. [ Google Scholar ] Articles from IEEE Journal of Translational Engineering in Health and Medicine are provided here courtesy of Institute of Electrical and Electronics Engineers ACTIONS View on publisher site PDF (4.1 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top