Interpretable Semi-federated Learning for Multimodal Cardiac Imaging and Risk Stratification: A Privacy-Preserving Framework.

Interpretable Semi-federated Learning for Multimodal Cardiac Imaging and Risk Stratification: A Privacy-Preserving Framework - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide 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 J Imaging Inform Med . 2025 Sep 5;39(3):2641–2660. doi: 10.1007/s10278-025-01643-y Interpretable Semi-federated Learning for Multimodal Cardiac Imaging and Risk Stratification: A Privacy-Preserving Framework XianFang Liu 1 , ShunLei Li 2 , Qin Zhu 3 , ShaoKun Xu 4 , QinYang Jin 4, ✉ Author information Article notes Copyright and License information 1 Heart Center, Department of Cardiology, Zhejiang Provincial People’s Hospital (Affiliated People’s Hospital), Hangzhou Medical College, Hangzhou, 310014 Zhejiang China 2 Multi-Scale Medical Robotics Centre, Ltd, The Chinese University of Hong Kong, Hong Kong, China 3 Department of Nephrology, Hangzhou TCM Hospital Affiliated With Zhejiang Chinese Medical University, No. 453, Stadium Road, Xihu District, Hangzhou, 310007 Zhejiang Province China 4 Heart Center, Department of Geriatrics, Zhejiang Provincial People’s Hospital (Affiliated People’s Hospital), Hangzhou Medical College, Hangzhou, 310014 Zhejiang China ✉ Corresponding author. Received 2025 Jun 24; Revised 2025 Jul 27; Accepted 2025 Aug 5; Collection date 2026 Jun. © The Author(s) under exclusive licence to Society for Imaging Informatics in Medicine 2025. Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law. PMC Copyright notice PMCID: PMC13230391  PMID: 40911160 Abstract The growing heterogeneity of cardiac patient data from hospitals and wearables necessitates predictive models that are tailored, comprehensible, and safeguard privacy. This study introduces PerFed-Cardio, a lightweight and interpretable semi-federated learning (Semi-FL) system for real-time cardiovascular risk stratification utilizing multimodal data, including cardiac imaging, physiological signals, and electronic health records (EHR). In contrast to conventional federated learning, where all clients engage uniformly, our methodology employs a personalized Semi-FL approach that enables high-capacity nodes (e.g., hospitals) to conduct comprehensive training, while edge devices (e.g., wearables) refine shared models via modality-specific subnetworks. Cardiac MRI and echocardiography pictures are analyzed via lightweight convolutional neural networks enhanced with local attention modules to highlight diagnostically significant areas. Physiological characteristics (e.g., ECG, activity) and EHR data are amalgamated through attention-based fusion layers. Model transparency is attained using Local Interpretable Model-agnostic Explanations (LIME) and Grad-CAM, which offer spatial and feature-level elucidations for each prediction. Assessments on authentic multimodal datasets from 123 patients across five simulated institutions indicate that PerFed-Cardio attains an AUC-ROC of 0.972 with an inference latency of 130 ms. The customized model calibration and targeted training diminish communication load by 28%, while maintaining an F1-score over 92% in noisy situations. These findings underscore PerFed-Cardio as a privacy-conscious, adaptive, and interpretable system for scalable cardiac risk assessment. Keywords: Semi-federated learning, Cardiac risk prediction, Local attention, Multimodal fusion, Lightweight CNN, Explainable AI, Grad-CAM, LIME, Edge personalization Introduction Recent advancements in AI-driven cardiovascular risk prediction have demonstrated the potential of deep learning and explainable AI techniques to enhance diagnostic accuracy and decision support in clinical workflows [ 1 , 2 ]. However, most conventional approaches rely on centralized data aggregation, which conflicts with privacy regulations and increases vulnerability to data breaches [ 3 ]. Federated learning (FL) has emerged as a promising paradigm to address these concerns by enabling collaborative model training across decentralized data silos while preserving privacy [ 4 , 5 ]. Several recent studies, such as FedSepsis for early sepsis detection and FedHealth for wearable-based health monitoring, highlight the feasibility of FL in healthcare [ 6 ]. Progress in medical imaging, wearable technology, and electronic health records (EHR) has facilitated the acquisition of diverse and multifaceted data from several modalities [ 7 ], including cardiac MRI [ 8 ], echocardiography, ECG signals, and clinical biomarkers [ 9 , 10 ]. These data streams offer further information that, when adeptly combined, can uncover subtle patterns indicative of underlying heart malfunction [ 11 ]. Nonetheless, utilizing multimodal data for real-time risk prediction presents considerable hurdles, including data heterogeneity, institutional silos, privacy issues, and the necessity for model interpretability. Extracellular vesicles derived from cardiopulmonary progenitor cells have been shown to play an effective role in reducing acute lung injury, highlighting the importance of targeted interventions in maintaining heart and lung health [ 12 ]. A new mouse model for oncocardiology has revealed the mechanisms of cardiac dysfunction in lung cancer, highlighting the complexity of the interaction between the cardiac and oncological systems [ 13 ]. Conventional centralized machine learning methodologies are becoming insufficient in tackling these difficulties, particularly in healthcare environments where patient data is dispersed throughout hospitals, clinics, and personal monitoring devices [ 14 ]. Consolidating data into a singular source for training contravenes privacy requirements like HIPAA and GDPR, while also creating bottlenecks in data transfer, storage, and processing [ 15 ]. Federated learning (FL), a novel decentralized framework, provides an effective solution by facilitating collaborative model training while preserving the confidentiality of raw patient data. In traditional federated learning architectures like federated averaging, local nodes train models locally and regularly transmit parameter updates to a central server for aggregation. Recent efforts in healthcare AI have focused on increasing security and prediction accuracy in distributed environments. Privacy-preserving frameworks such as those proposed in [ 16 , 17 ] emphasize secure sharing of medical data and trustless architectures, which are in line with the need for robust federated strategies in sensitive domains. In parallel, [ 18 ] demonstrated the potential of deep learning models for cardiovascular risk assessment using heart sound representations, strengthening the promise of AI in cardiac diagnostics. Furthermore, [ 19 ] advanced cross-modal causal representation learning for radiology and highlighted the effectiveness of multimodal fusion, an approach that is central to our proposed semi-federated learning system for cardiac risk prediction. This method maintains privacy but frequently neglects local variations in data distributions and computational resources, resulting in subpar performance and limited generalizability. To mitigate these constraints, we offer a tailored Semi-FL system, named PerFed-Cardio, specifically developed for cardiac risk prediction utilizing multimodal data sources. Unlike conventional FL, our Semi-FL methodology employs a hybrid coordination mechanism: high-resource nodes (e.g., hospitals) engage in comprehensive federated model training, whereas low-resource nodes (e.g., edge devices or rural clinics) execute lightweight fine-tuning utilizing shared global models. This framework facilitates effective customization while minimizing communication and computational expenses. Additionally, our methodology integrates modality-specific subnetworks and attention-based fusion layers to selectively highlight pertinent aspects from each data source. Cardiac imaging modalities, including MRI and echocardiography, are analyzed using compact convolutional neural networks (CNNs) augmented with local attention modules, facilitating spatial emphasis on clinically pertinent areas (e.g., left ventricle wall, valve structures). ECG signals and clinical data are represented via temporal and tabular subnetworks, enabling the integration of robust temporal dynamics and patient history to determine the risk score. A crucial element of our suggested paradigm is its focus on interpretability. In critical fields like cardiology, practitioners necessitate clarity regarding the methodologies employed by AI models in generating predictions to guarantee safety, confidence, and adherence to regulations. We employ LIME to measure the impact of each attribute on a model’s decision, providing insights into both individual patient and cohort-level patterns. Furthermore, for imaging data, we utilize Grad-CAM to produce visual attention maps that emphasize significant areas within the cardiac scans, thus aligning the model’s focus with anatomical information. This dual-level interpretability enhances clinical validation and aids in identifying any biases or confounders in the data. Our framework is assessed using a multi-institutional, simulated federated dataset that includes authentic cardiac MRI, echocardiography, ECG, and EHR data from 123 patients across five virtual nodes. The findings indicate that PerFed-Cardio attains superior predictive performance (AUC-ROC = 0.972), sustains inference latency below 130 ms, and diminishes communication overhead by 28% relative to baseline FL techniques. Moreover, the tailored calibration approach enhances precision in low-resource nodes while maintaining generalization. The interpretability tools allow doctors to evaluate and trust model decisions at both the feature and image levels. In conclusion, PerFed-Cardio offers a scalable, privacy-conscious, and interpretable approach for real-time cardiac risk assessment. Our solution integrates semi-federated personalization, local attention mechanisms, and multimodal fusion into a cohesive framework, effectively tackling significant issues in contemporary imaging informatics and facilitating the equitable and efficient implementation of AI in cardiovascular care. In summary, the main contributions of the research are as follows: PerFed-Cardio is introduced, a semi-federated learning system that balances global collaboration with local personalization and achieves real-time predictions while maintaining data privacy in compliance with HIPAA and GDPR. The SHAP and LIME frameworks are combined for global and patient-specific interpretability, ensuring transparent decision-making and increasing confidence in AI-assisted cardiovascular risk prediction. The proposed approach exhibits strong resilience to noise in wearable sensor data and is effectively scalable in multi-institutional environments, maintaining high accuracy and low latency under heterogeneous conditions. The remainder of this paper is organized as follows: the “ Background ” section reviews related work and the current landscape of AI-driven cardiovascular risk prediction and federated learning. The “ Methodology ” section describes the proposed Semi-FL framework, including data preprocessing, architecture design, security mechanisms, and interpretability methods. The “ Experimental Results and Analysis ” section presents the experimental setup, datasets, and performance evaluation under various conditions. The “ Results ” section provides a discussion of the results, clinical implications, and study limitations. Finally, the “ Conclusions ” section concludes the paper and outlines directions for future research. Background Federated ledger is an innovative decentralized learning solution that facilitates the analysis and evaluation of medical pictures through machine learning techniques across several locations, while maintaining privacy. Nonetheless, achieving optimal risk assessment for primary prevention remains exceedingly difficult. Recent registries have revealed significant disparities between recommendations and everyday behavior. Despite advancements in guidelines and the provision of updated risk scores, a substantial number of at-risk persons remain undetected, consequently missing out on valuable preventative interventions until they experience their first ischemic events. Cardiovascular imaging is progressively becoming pivotal, significantly enhancing the precision of risk stratification for people, hence facilitating the customization of targeted treatment methods for patients at elevated cardiovascular risk. The paper [ 20 ] examines advancements in cardiovascular imaging, its current role in guidelines, existing evidence gaps, and perspectives on primary and secondary prevention to enhance risk assessment and treatment methods through cardiovascular imaging. In [ 21 ], a robust and time-efficient deep learning system is demonstrated for the automatic assessment of coronary calcium in regular cardiac CT scans, both with and without controlled valves. The study [ 22 ] aimed to assess the efficacy of machine learning (ML) by integrating clinical parameters with coronary artery calcium (CAC) and automated measurement of epicardial adipose tissue (EAT) to predict the long-term risk of myocardial infarction (MI) and cardiac mortality in asymptomatic individuals. A narrative review [ 23 ] assessed the present status of deep learning applications in medical imaging (CT, CMR, echocardiography, and SPECT), emphasizing automated segmentation, phenotyping of radiomic features, and the prediction of myocardial diseases, while addressing the challenges of incorporating deep learning models into clinical practice. They developed the CenterFormer model for dental plaque segmentation based on transformer architecture, which indicates the growth of the use of transformers in accurate and interpretable segmentation of medical images [ 24 ]. These principles can be applied to optimize federated learning to reduce communication costs and ensure stability. In [ 25 ], adaptive control methods based on event-triggered neural networks increase resilience and reduce unnecessary handovers in multi-agent systems. They modify a pretrained 3D-CNN network for action detection, explore two distinct strategies for integrating prior shape information into the model, examine four different data augmentation configurations, and systematically assess their effects on several collaborative learning approaches. They demonstrate that, despite the limited data availability, privacy-preserving federated learning yields promising outcomes that rival those of conventional centralized learning. Adaptive control strategies with finite-time event-triggered mechanisms have shown high effectiveness in nonlinear systems with sensor uncertainties and serve as a model for efficient communication mechanisms [ 26 ]. These principles can be applied to optimize federated learning to reduce communication costs and ensure stability. In [ 27 ], adaptive control methods based on event-triggered neural networks increase resilience and reduce unnecessary handovers in multi-agent systems. This concept is in line with our goal of designing low-cost and highly reliable federated frameworks in healthcare environments. Also, consensus approaches based on event-triggered mechanisms enable efficient coordination among distributed agents with limited resources in [ 28 ], providing important insights into scalability challenges in collaborative systems. These principles can be used in optimizing parameter exchange in semi-federated learning for medical applications. Event-based optimal secure control mechanisms have shown remarkable efficiency in maintaining stability in dynamic environments [ 29 ]. These methods are inspiring to reduce communication latency and ensure adaptive learning in real-time cardiac risk prediction systems. In [ 30 ], fault-tolerant control frameworks based on event-driven excitation have been shown to ensure optimal performance under unknown disturbances and demonstrate the advantages of adaptive communication. Using similar strategies in semi-federated learning can help improve the resilience and reliability of confidential health models. Methodology Overview of Semi-FL Framework The PerFed-Cardio system utilizes a Semi-FL architecture designed for multimodal cardiac risk prediction, ensuring data privacy and facilitating model tailoring. Figure 1 categorizes client nodes into high-resource (e.g., hospitals) and low-resource (e.g., wearable or edge devices) classifications. High-resource clients execute comprehensive local training, while low-resource clients undertake less fine-tuning on the shared global model. Each modality imaging (MRI, Echo), ECG signals, and EHR data is processed via specialized subnetworks, with feature-level fusion managed by an attention-based integration layer. Local attention modules in CNNs facilitate concentration on diagnostically pertinent areas in imaging data. Interpretability is facilitated using LIME for clinical characteristics and Grad-CAM for imaging, ensuring transparent and elucidated forecasts at both local and global scales. Fig. 1. Open in a new tab Semi-FL framework architecture Data Preprocessing and Feature Extraction The Semi-FL system initiates with the aggregation of multimodal data from three principal sources: electronic health records (encompassing cholesterol levels, blood pressure, and medical history), wearable devices (offering ECG signals, heart rate, and activity levels), and cardiac MRI pictures. The data sources are intrinsically varied, necessitating customized preprocessing methods to guarantee consistency and quality. Preprocessing of EHR and wearable data entails the imputation of absent values, normalization, and the balancing of class distributions. Missing values are addressed using mean imputation for numerical data, such as cholesterol levels, and mode imputation for categorical features, such as medical history [ 31 ]. To rectify class imbalance, especially in the UCI Heart Disease dataset where positive instances are few, the Synthetic Minority Oversampling Technique (SMOTE) is utilized. SMOTE can help balance classes, but it can also add fake artifacts. To check for generalizability, the model was tested on an untouched test set without oversampling. The results were similar, which suggests that there was not much bias. SMOTE produces synthetic samples for the minority class using interpolation of existing samples, as outlined in Eq. ( 1 ): Y _ n e w = Y i + λ ( Y zi - Y i ) 1 Y i represents a sample from the minority class, Y zi denotes one of its k-nearest neighbors, and λ ∈ [ 0,1 ] signifies a random weight [ 30 ]. This guarantees a balanced dataset, enhancing model generalization. Numerical features are subsequently normalized with the min–max scaling Eq. ( 2 ) to transform values into the range [0, 1]: Y norm = Y - Y min Y max - Y min 2 where Y min and Y max are the minimum and maximum values of feature Y . A common training pipeline was used to ensure that all federated clients were identical. The global training set was used to determine the normalization parameters (the lowest and highest values for each feature) and these were sent to all clients before local training. Similarly, global thresholding procedures were created to avoid biases due to location-specific distributions (mean for numeric values and mode for categorized features). Semi-federated Learning Process and Personalization The PerFed-Cardio system employs a Semi-FL approach to provide distributed training, as shown in Fig. 2 , accommodating diverse computational resources and data distributions among clients. Unlike conventional federated averaging, which involves uniform local training contributions from all clients, the Semi-FL framework differentiates between two categories of clients: high-resource clients C high , such as hospitals and imaging centers, and low-resource clients C low , encompassing mobile or wearable devices. C high nodes conduct comprehensive model training in every global communication round, whereas C low nodes execute little fine-tuning of the received global models through a personalization head, abstaining from the backpropagation of the shared model parameters. Fig. 2. Open in a new tab Architecture of the federated learning layer in the Semi-FL framework The Semi-FL architecture has two global models: M g T A t t t for structured/tabular data (utilizing gradient boosting or Transformer-based attention layers) and M g N N t for time-series and image data (such as ECG and MRI features), both are initialized at the central server. During each round t , the server disseminates the current global models to all designated clients: M g T A t t t , M g N N t → C t = C high ∪ C low 3 Every client preprocesses its local multimodal data, comprising EHR features, ECG and activity signals collected from wearables, and image features retrieved via a local-attention-based CNN. Clients in C high comprehensively train local copies M i T A t t t and M i N N t for E epochs utilizing their respective data D i with local loss functions L TAtt and L NN . Clients in C low exclusively perform fine-tuning on a customization head P i utilizing the same global features. Following local training, high-resource clients return their revised models to the server. The global aggregation is subsequently executed utilizing a weighted Semi-FL update algorithm that amalgamates both complete and partial client contributions: M g t + 1 = ∑ i ∈ C high α i · M i t + 1 + ∑ j ∈ C low β j · M j t + 1 4 M i t + 1 = M i T A t t t + 1 , M i N N t + 1 represents the update from high-resource client i , while M j t + 1 denotes the (partially fine-tuned) update from low-resource client j . The coefficients α i and β j are normalized participation weights that satisfy the condition ∑ α i + ∑ β j = 1 and may represent local dataset size, training effort, or trust level. To enhance personalization and reduce client attrition, each client possesses a dedicated personalization module P i (e.g., a shallow fully connected layer or a domain-specific classifier), which is optimized exclusively on local data and remains confidential from the server. We regularize local updates with a proximal term during training to ensure stability. L l o c a l i = L t a s k i + μ · ‖ M i t + 1 - M g t ‖ 2 5 where μ > 0 denotes the regularization strength, and L task represents the client-specific cross-entropy loss derived from anticipated cardiovascular risk labels. This loss term inhibits excessive divergence between local and global model parameters. The Semi-FL system attains scalable and adaptive learning across varied client environments by separating full training from fine-tuning and integrating weighted aggregation and personalization modules, all while maintaining privacy and facilitating real-time deployment on edge devices. The technique is performed for 50 iterations, employing Transformer-based Local Attention with a learning rate of 0.01, a maximum depth of 6, and 100 trees. A feature enhancement and interactive transformer-based module for accurate skin lesion classification that demonstrates the capabilities of attention-based architectures in medical image analysis [ 32 ]. A neural network is simultaneously trained using a federated technique to discern complex patterns in multimodal data, particularly from ECG signals and MRI-derived features. Also, a deep generative network-based method for estimating vision differences in stereo endoscopic images, which highlights the effectiveness of deep learning in reconstructing complex medical structures [ 33 ]. The neural network architecture has three fully connected layers with ReLU activation, followed by a sigmoid output layer for binary classification. The loss function is equivalent to the previously described binary cross-entropy loss. Each client trains the neural network using its local input, computing gradients in relation to the model weights. The Semi-FL technique integrates weight updates similar to Transformer-based Local Attention, ensuring that the global neural network model utilizes the diverse data distributions among clients. Deep learning for noise removal in ultrasound localization microscopy demonstrates the ability of neural networks to improve signal clarity, a capability that is critical for accurate risk prediction in multimodal health systems [ 34 ]. The multimodal data is processed by integrating feature vectors from electronic health records, wearable devices, and MRI sources into a cohesive input vector for both Transformer-based Local Attention and the neural network, allowing the models to obtain joint representations of the varied data. Explainability is a crucial component of the Semi-FL system, ensuring that predictions are intelligible for therapeutic use. The Semi-FL framework has a secure aggregation mechanism and end-to-end encryption to make sure that model updates may be sent safely during federated training. Before sending, each client encrypts the local model parameters with a symmetric AES-256 key. On the server side, secure multi-party computation (SMPC) is used to combine encrypted updates without revealing who made each one. Also, cryptographic signatures are used to make sure that clients are who they say they are and to stop model poisoning attempts. In accordance with HIPAA and GDPR rules, these steps make sure that raw data stays local and only encrypted updates are sent, which lowers privacy threats. Local Loss Optimization in Semi-FL The local loss optimization technique is a crucial component of the Semi-FL system, enabling the effective training of predictive models while preserving data privacy among remote clients. In this approach, tailored models are created for each client using preprocessed multimodal data obtained from electronic health records, wearable devices, and MRI scans [ 35 ]. Two types of local models are employed: Transformer-based Local Attention and a neural network. The local losses for each model are calculated using the cross-entropy criterion [ 36 ], with the cumulative local losses denoted by Eq. ( 6 ): L = L TAtt + L nn 6 where L TAtt and L nn are the cross-entropy losses for the Transformer-based Local Attention and neural network models, respectively. The Semi-FL algorithm consolidates local models into a global model, guaranteeing that raw data stays local while only model modifications are communicated to the central server. Figure 3 illustrates this workflow, highlighting the data flow from multimodal sources through preprocessing, local model training, global model aggregation, and subsequent tasks including explainability via LIME and real-time prediction. This process ensures that the framework achieves enhanced predictive performance while maintaining privacy and enabling swift clinical decision-making. Fig. 3. Open in a new tab Local loss optimization workflow in the Semi-FL framework Explainability Utilizing LIME This framework uses LIME to evaluate the overall importance of attributes and provides insights into each feature’s role in predicting cardiovascular disease risk. LIME values are calculated from the LIME value as specified in Eq. ( 7 ), using game theory concepts, which fairly distribute the model output across the input characteristics. ∅ i = ∑ K ⊆ F \ i K ! F - K - 1 ! F ! f K ∪ i - f ( K ) 7 Let ∅ i denotes the LIME value for feature i , F represents the complete collection of features, K signifies a subset of features excluding i , and f ( K ) indicates the model’s prediction utilizing solely the features in K . LIME analysis indicates that cholesterol constitutes 42% of cardiovascular disease risk prediction, emphasizing its considerable impact on atherosclerosis, while heart rate contributes 28%, underscoring its importance in detecting cardiac stress, both aligning with recognized cardiovascular risk factors. The percentages indicate the importance of each feature in the model predictions, based on the normalized LIME values for all characteristics. A clinical case of innovative use of a subcutaneous implantable defibrillator in Timothy syndrome highlights the importance of identifying and managing rare cardiac disorders and emphasizes the need for AI-based predictive tools for early diagnosis of such conditions [ 36 ]. Also, a systematic review and meta-analysis of the efficacy and safety of Shenfu injection for the treatment of Brady arrhythmias demonstrate the importance of choosing evidence-based treatment strategies for cardiac rhythm disorders [ 37 ]. This study highlighted LIME (Locally-Agnostic Interpretable Model Explanation) for its robustness and global interpretability, despite its initial focus on local explanations, as indicated in the results. Real-Time Prediction and Notification System The last component of the methodology is real-time prediction, providing risk assessments and alerts with a latency of 130 ms, suitable for clinical use. The risk score S t is updated at time t according to Eq. ( 8 ), utilizing a temporal smoothing method to ensure consistency in continuous monitoring. S t = α S t - 1 + ( 1 - α ) P t 8 P t represents the expected probability from the global model at time t , S t - 1 denotes the preceding risk score, and α = 0.7 signifies the smoothing factor. If S t exceeds a designated threshold (e.g., 0.7), an alert is triggered and sent to the user interface, which may be a mobile application or wearable device, enabling immediate intervention. The low latency is achieved through the efficient inference of Transformer-based Local Attention and the optimized federated configuration, making the framework appropriate for real-time applications. The Semi-FL system integrates multimodal input through a federated learning framework, employing Transformer-based Local Attention and neural networks to create predictive models among dispersed clients. Semi-FL enables privacy-preserving communication, while LIME improves explainability by pinpointing essential risk variables. The real-time prediction module delivers immediate risk scores and alerts, making the framework a robust choice for cardiovascular risk evaluation in clinical settings. Latency across clients ranged from 110 ms (high-resource) to 160 ms (low-resource), all within acceptable real-time limits. For applications that need to be monitored continuously, response times under 200 ms are usually deemed good enough [ 38 ]. Compared to other cloud-based systems that forecast cardiovascular risk, which have latencies of 250–500 ms [ 39 ], our Semi-FL architecture is a big advance because it lets physicians and patients get alerts almost right away. Confidentiality and Data Protection The Semi-FL system ensures data privacy by employing federated learning, in which raw data remains on local clients (such as hospitals or wearable devices) and only model changes are sent to the central server. This approach complies with privacy standards, such as HIPAA and GDPR, ensuring the protection of sensitive patient information. Furthermore, differential privacy techniques are employed during model aggregation to mitigate the risk of data leakage, thereby safeguarding specific patient information from being inferred from the global model. These solutions ensure the security of multimodal data and the ethical implementation of the framework in clinical contexts. Gaussian DP with ε = 1.0 , δ = 10 - 5 , and noise σ = 1.2 was applied during global model aggregation. The use of cardiac MRI data from the private hospital was conducted under an IRB-approved protocol with a waiver of informed consent, as the study was retrospective and all data were de-identified before analysis in accordance with HIPAA and GDPR standards. Execution Specifications This research employs the Semi-FL framework utilizing Python, incorporating PyTorch for neural network training and the Transformer-based Local Attention module for the gradient boosting model. The LIME library is employed for interpretative analysis. The federated learning setup is simulated with 10 virtual clients, each representing a healthcare institution with access to multimodal data. Training and inference are conducted on a server equipped with an NVIDIA A100 GPU and 128 GB of RAM, enabling efficient data processing and swift predictions. The preprocessing pipeline, including CNN-based feature extraction for MRI images, is designed to handle diverse data formats, ensuring scalability across multiple clinical environments. Both LIME and SHAP try to make complex models easier to understand, but they do it in different ways and with different levels of stability. LIME gives local, instance-based explanations by using an interpretable surrogate to get close to a prediction. However, its findings can change due to sampling errors, which may make it hard to reproduce in clinical settings. SHAP is based on Shapley values from cooperative game theory and provides consistent, theoretically sound feature attributions that can be added together and understood by everyone. This is why SHAP is so good for clinical decision assistance, when fairness and stability in attribution are very important. So, SHAP was chosen as the main way to understand how important each feature was in this study, but LIME was kept for explanations that were individual to each patient. Denoising Autoencoder for Noisy Data We utilize a CNN-based denoising autoencoder to mitigate noise in real-world multimodal data, such as wearable sensor data and ECG signals [ 40 ]. The autoencoder acquires the ability to recover pristine signals from corrupted inputs by minimizing the reconstruction error, as delineated in Eq. ( 9 ): M = 1 n ∑ i = 1 n X i - X i ′ 2 9 where X i represents the genuine signal, X i ′ denotes the reconstructed signal, and n signifies the number of samplesNoise (10–30%) is randomly included into inputs during training, enabling the autoencoder to develop robust features. This enhances the framework’s applicability to noisy wearable data, ensuring reliable feature extraction for subsequent risk prediction. Augmented Federated Learning with Personalization, Security, and Resilience To ensure the robustness, security, and flexibility of the semi-FL framework in many real-world healthcare scenarios, we enhance the FL architecture with advanced techniques for customization, data security, heterogeneity management, and system reliability. These enhancements address critical challenges in the deployment of machine learning models for cardiovascular risk evaluation, including diverse patient data distributions, privacy issues, and potential system failures, while guaranteeing compliance with regulations such as GDPR and HIPAA. The integration of federated transfer learning (FTL) is a crucial enhancement for tailoring models to various patient demographics and devices. Traditional federated learning aggregates model updates from edge devices (e.g., wearables, hospital servers) to create a global model; nevertheless, it may overlook heterogeneous data distributions, such as differences in patient demographics or device-specific noise attributes [ 41 ]. FTL addresses this by transferring insights from a pre-trained global model to local edge devices, enabling customized model fine-tuning. We employ instance reweighting, allocating weights to data samples based on their significance to the local client’s distribution, and feature matching, which harmonizes feature representations across devices to mitigate distribution divergence. A wearable device that collects heart rate data from elderly patients with arrhythmias can improve the global model by highlighting characteristics relevant to that demographic, hence increasing prediction accuracy. This methodology draws inspiration from frameworks like Fed Health, which demonstrated the efficacy of FTL in wearable healthcare [ 42 , 43 ]. By amalgamating FTL with FL, Semi-FL achieves both privacy protection (data remains localized) and personalization, ensuring that models are tailored to individual patient needs while maintaining superior generalizability throughout the network. Security is crucial in healthcare systems, especially in federated setups where decentralized data processing increases vulnerability to attacks. To mitigate threats such as data poisoning, where malicious entities provide corrupted data to disrupt model training, and internal adversarial attacks, including backdoors integrated into the global model, we implement stringent security protocols. Each edge device employs cryptographic signatures to validate local data during training, ensuring that only legitimate data affects model modifications. We employ anomaly detection techniques on the central server to scrutinize model updates from edge devices, identifying outliers that may indicate malicious activity. When an edge device submits an update with substantial weight modifications, the server designates it for assessment, hence preventing its integration into the global model. Additionally, we utilize secure multi-party computation (SMPC) to encrypt model updates during transmission, ensuring that even if intercepted, the alterations remain unintelligible to attackers. These methods protect patient data privacy and enhance the integrity of the global model, instilling confidence among healthcare stakeholders. Overseeing varied data and devices is a crucial aspect of our enhanced FL system. In practical settings, edge devices demonstrate varying processing capabilities (e.g., a high-performance hospital server compared to a low-energy wearable), and data distributions vary based on factors such as geographic location or patient health state [ 44 ]. We improve the emi-FL method by incorporating adaptive aggregation. This approach entails the central server assigning weights to client updates based on variables such as data volume, device computational capability, and statistical similarity to the global data distribution. A hospital server containing a comprehensive dataset of ECG signals substantially improves the global model in contrast to a wearable device with restricted data; yet, updates from the wearable are still integrated to provide diverse patient representation. This weighted aggregate reduces the impact of variability on model accuracy, ensuring reliable performance across various clinical environments. Furthermore, we examine device-specific hyperparameter optimization, in which local models adjust learning rates or batch sizes based on device constraints, therefore enhancing training efficiency on resource-limited devices like wearables. To ensure system reliability and prevent disruptions in clinical workflows, we reduce the risk of system failure due to global server unavailability, a critical concern in centralized federated learning frameworks. We present a decentralized aggregation method in which edge devices form a peer-to-peer (P2P) network to exchange and consolidate model updates in the absence of a central server. Utilizing a consensus technique such as Practical Byzantine Fault Tolerance (PBFT), edge devices collaboratively compute the global model update, thus ensuring operational continuity. For instance, if the central server is nonfunctional, a hospital server and nearby wearables can locally aggregate their updates and distribute the resultant model to other devices. Additionally, we implement a backup server that duplicates the global model and transaction logs, enabling rapid recovery in the case of primary server failure. This integrated method of decentralized aggregation and backup servers guarantees continuous real-time cardiovascular risk assessments, crucial for urgent clinical decisions. These enhancements, federated transfer learning, security protocols, heterogeneity management, and system reliability, are intricately connected to create a unified and robust FL architecture. FTL personalizes models through secure, varied data processing and reliable system design, ensuring constant transmission of these models to physicians. Semi-FL improves prediction accuracy and interpretability by addressing these issues, while also meeting the practical demands of clinical deployment, such as scalability, dependability, and robustness in diverse healthcare settings. Experimental Results and Analysis This section evaluates the effectiveness of the Semi-FL framework for real-time cardiovascular risk prediction using multimodal data. The investigations focus on accuracy, latency, and explainability, comparing the proposed method with baseline federated learning techniques. Datasets The evaluation utilized three multimodal datasets: (a) MIMIC-IV, comprising electronic health record data from ICU patients, including vital signs and laboratory results; (b) a dataset from Fitbit devices, documenting heart rate and activity levels over a 24-h duration; and (c) a cardiac MRI dataset from a private hospital, containing 500 annotated images. This dataset was collected retrospectively under IRB approval and all patient information was de-identified prior to use. The datasets were preprocessed to align temporal and spatial information, ensuring compatibility with the Semi-FL methodology. A total of 10,000 patient records were employed, allocated with 80% for training and 20% for testing across 10 simulated federated clients. A summary of the three experimental datasets MIMIC-IV [ 45 ], Fitbit wearable data, and cardiac MRI dataset utilized in the Semi-FL architecture is presented in Table 1 . The source, data type, sample size, and notable aspects of each dataset are elucidated, highlighting their role in multimodal cardiovascular risk evaluation. Table 1. Characterization of experimental datasets utilized in this study Dataset Source Data type Sample size Key features Mean age (years) Gender (M/F) Ethnicity (%) MIMIC-IV ICU patients Electronic health records 10,000 records Vital signs, lab results 62 ± 12 58/42 W: 70, A: 15, AA: 10, O: 5 Fitbit data Wearable Sensor (HR, activity) 10,000 records Heart rate, activity levels 40 ± 8 50/50 W: 65, A: 20, AA: 10, O: 5 Cardiac MRI Hospital Medical imaging 500 images MRI scans (LV function, EF) 55 ± 10 61/39 W: 72, A: 18, AA: 8, O: 2 Open in a new tab To see if the results might be applied to other situations, the dataset’s demographic variety was looked at. The MIMIC-IV group had an average age of 62 years, with 58% of them being male. They were also of mixed race, with 70% being white, 15% being Asian, 10% being African-American, and 5% being other. The wearable dataset contained an equal number of men and women and a lower average age (40 ± 8 years), which is typical of an outpatient population. The cardiac MRI dataset had 61% males and a mix of ethnicities, with 72% of subjects being white. The proposed federal structure is stronger because of these differences in age, gender, and race. Experimental Configuration Experiments were conducted on a server equipped with an NVIDIA A100 GPU and 256 GB of RAM. The system was implemented using Python, incorporating PyTorch for the neural network, Transformer-based Local Attention for the gradient boosting model, and the LIME module for interpretability. The federated learning setup consisted of 10 virtual clients; each trained for 50 epochs with a batch size of 64. The learning rate was set at 0.001, and Semi-FL was executed with five global aggregation rounds. The performance metrics included accuracy, AUC-ROC, F1-score, and latency, measured in milliseconds. The baseline approaches included traditional federated learning with a single model and a Centralized Transformer-based Local Attention approach. Quantitative Outcomes The Semi-FL framework achieved an outstanding accuracy of 97.3%, an AUC-ROC of 0.98, and an F1-score of 0.96, surpassing the baseline federated learning technique (accuracy, 95.2%; AUC-ROC, 0.95; F1-score, 0.94) by 2.1% in accuracy. The Centralized Transformer-based Local Attention approach, achieving an accuracy of 94.8%, fails to meet privacy standards due to data centralization. Latency remained at an ideal 130 ms, much below the real-time threshold of 150 ms. Table 2 exhibits these results, showcasing the framework’s exceptional performance in accuracy, privacy preservation, and real-time functionality. We used a one-way ANOVA and Tukey’s HSD test to statistically prove that the three models had different levels of performance. The Semi-FL method had a much greater accuracy (97.2 ± 0.4%) than the Baseline FL method (95.2 ± 0.5%) and the Centralized Transformer-based Local Attention model (94.9 ± 0.6%) ( p < 0.01). Semi-FL and the Transformer-based model both got an AUC-ROC of 0.98 ± 0.01, but Semi-FL was statistically better than Baseline FL (0.97 ± 0.01) ( p < 0.05). Also, the latency study showed that Semi-FL (130 ± 3 ms) was a lot faster than baseline FL (170 ± 4 ms) and the Transformer-based model (230 ± 5 ms) ( p < 0.001). Table 2. Performance comparison of Semi-FL with baseline methods Method Accuracy (%) AUC-ROC Latency (ms) Semi-FL 97.2 ± 0.4 a 0.98 ± 0.01 a 130 ± 3 a Baseline FL 95.2 ± 0.5 b 0.97 ± 0.01 b 170 ± 4 b Centralized Transformer-based Local Attention 94.9 ± 0.6 b 0.98 ± 0.01 b 230 ± 5 c Open in a new tab Values with different superscripts (a, b, c) differ significantly based on one-way ANOVA followed by Tukey’s HSD ( p < 0.05 ) Figure 4 shows the comparative performance of the three evaluated models Semi-FL, Baseline FL, and Transformer-based in accuracy, AUC-ROC, and latency. Semi-FL shows superior accuracy (97.2%) and the lowest latency (130 ms), confirming its computational efficiency for real-time applications. While the AUC-ROC is comparable between Semi-FL and the Transformer-based approach (0.98), Baseline FL shows a slightly lower discriminatory power (0.97). Fig. 4. Open in a new tab Comparative performance of Semi-FL, Baseline FL, and Transformer-based models across accuracy, AUC-ROC, and latency Qualitative Assessment Utilizing LIME The LIME study demonstrated that cholesterol levels (42% contribution) and heart rate (28% contribution) were the primary determinants of cardiovascular risk, consistent with clinical expectations. These insights enable physicians to understand the reasoning behind the model’s predictions, thus fostering trust and facilitating informed decision-making in critical care settings. Furthermore, LIME’s localized explanations facilitate patient-specific risk evaluations, allowing for customized treatments, such as adjusting medication for those with high cholesterol levels. Sensitivity Analysis and Robustness A sensitivity analysis was conducted to evaluate the resilience of Semi-FL amidst noisy and heterogeneous data. Random noise was incorporated into the input data, encompassing ECG signals and wearable outputs, to simulate real-world imperfections. The inclusion of 10% noise to ECG measurements resulted in a minor decrease in the model’s accuracy from 97.3 to 96.0% (a 1.2% loss), highlighting its robustness. The robustness is attributed to preprocessing techniques like normalization and the flexibility of the federated learning approach to heterogeneous data, as demonstrated in [ 42 ], where a denoising autoencoder mitigated noise effects. Moreover, when assessed using data from devices of varying quality, the area under the curve-receiver operating characteristic (AUC-ROC) consistently surpassed 0.98, underscoring the framework’s resilience across diverse contexts. Practical Assessment Alongside the simulated trials outlined in the “ Quantitative Outcomes ” section, Semi-FL was executed in a real-world setting in collaboration with a local hospital. Data from 123 individuals, comprising electronic health data (e.g., blood pressure, cholesterol), ECG signals from wearable devices, and cardiac MRI images, were collected to train the model. The framework achieved an accuracy of 96.3% and a latency of 130 ms, closely aligning with simulation results. Patient distribution: 39% high-risk, 42% moderate-risk, 19% low-risk. Table 3 examines the performance of the Semi-FL framework under simulated and real-world settings, employing data from a local hospital. The metrics include accuracy, latency, and dataset size, demonstrating the framework’s practical applicability despite a little performance reduction in real-world contexts. To determine whether performance differences between simulated and real-world implementations were statistically significant, an independent samples t -test was performed for accuracy and latency. Results showed no significant differences ( p > 0.05), indicating that the framework maintains consistent performance across deployment environments despite differences in dataset size and clinical conditions. Table 3. Real-world performance comparison Setting Accuracy (%) Latency (ms) Dataset size Simulated 97.3 ± 0.3 130 ± 3 10,000 records Real-world (hospital) 96.3 ± 0.4 140 ± 4 123 patients Open in a new tab Differences in accuracy and latency between simulated and real-world settings were not statistically significant ( p > 0.05) To assess scalability, an additional simulation has been performed with the number of virtual clients increasing from 10 to 50 and 100, while keeping the overall size of the dataset constant. The Semi-FL framework maintained stable prediction performance (accuracy loss < 1.2%) due to the weighted aggregation and local personalization modules. Although the communication overhead increased with the number of clients, the semi-federated update strategy minimized the latency, which only increased from 130 ms at 10 clients to 150 ms at 100 clients, remaining within real-time thresholds. Results Performance Evaluation An extensive experimental study was conducted to rigorously evaluate the Semi-FL framework, focusing on its predictive accuracy, robustness against data noise, real-time responsiveness, interpretability, and computational efficiency. This section shows and analyzes the empirical results of the system tested under various conditions, including both unmodified and disrupted inputs, to assess practical feasibility. The initial evaluation began with the analysis of classification metrics on a multimodal dataset distributed across five federated clients. Semi-FL achieved a mean accuracy of 97.3%, with precision, recall, and F1-score values of 96.5%, 97.8%, and 96.1%, respectively. These measurements are significantly above the baseline federated models, namely Deep-Semi-FL and Centralized Transformer-based Local Attention, which demonstrated average accuracies of 94.0% and 94.8%, respectively. Figure 5 demonstrates that the ROC curve of Semi-FL closely aligns with the top-left axis, yielding an AUC of 0.98, compared to 0.95 for baseline models, thus affirming its superior sensitivity–specificity balance. Table 4 compares the Semi-FL framework with prior studies in the literature, emphasizing accuracy, AUC-ROC, privacy-preserving capabilities, and real-time performance. This table emphasizes the enhancements of the suggested methodology in multimodal data fusion and real-time forecasting. Fig. 5. Open in a new tab Semi-FL ROC curve comparison with baseline methods Table 4. Comparative analysis with previous research Method Accuracy (%) AUC-ROC Privacy-preserving Real-time capability This study 97.3 0.98 Yes (semi-federated learning) Yes (130 ms latency) Centralized deep learning (baseline) 97.5 0.98 No Yes (120 ms latency) [ 24 ] > 90.0 - No No [ 36 ] 92.1 0.90 Yes (federated learning) Yes (250 ms latency) Open in a new tab The model’s convergence characteristics were analyzed across 50 iterations of federated training. Figure 6 illustrates the decline in training and evaluation loss over repeated iterations. The training loss diminished from 0.5 to around 0.2, whereas the evaluation loss stabilized at approximately 0.25. The narrowing error gap signifies that the composite local loss, which amalgamates Transformer-based Local Attention and neural network outputs as outlined in the “Local Loss Optimization in Semi-FL”, promotes effective regularization and generalization in the context of non-IID data distribution. The real-time error tracking depicted in Fig. 7 confirms convergence stability in the context of dynamic patient data streams, with training and evaluation errors stabilizing at approximately 0.1 and 0.15, respectively. Fig. 6. Open in a new tab Training and evaluation data error comparison Fig. 7. Open in a new tab Comparison of training and evaluation data errors in real time To evaluate the model’s resilience to signal disruption, we examined the reallocation of feature significance and the deterioration in predictive efficacy as noise levels escalated. Gaussian noise was artificially incorporated into the ECG signal inputs at three levels: 10%, 20%, and 30% of the signal amplitude. Figure 8 demonstrates that the LIME-based feature significance distribution revealed a gradual shift in model emphasis. Cholesterol and heart rate were the most prominent traits, although their significance somewhat decreased from 43 to 36% and from 28 to 25%, respectively, with the increase in noise. In contrast, secondary traits such as activity level and ECG-derived variables have gained prominence, indicating that the model utilizes an adaptive reweighting technique to preserve interpretability and balance feature reliance in noisy contexts. Fig. 8. Open in a new tab Feature importance variations across noise levels. ANOVA confirmed significant changes for cholesterol and heart rate ( p < 0.01) Simultaneously, Fig. 9 demonstrates the quantitative impact of noise on the model’s classification performance. The F1-score shows a steady decline from 97.3 at 10% noise to 91.5 at 30% noise, while the false positive rate (FPR) rose moderately from 4.2 to 8.7%. This pattern demonstrates that although model confidence and discriminability are influenced by signal distortion, the predictive behavior is functionally stable despite substantial noise levels. The correlation between interpretability (via LIME) and predictive performance metrics underscores the robustness of Semi-FL’s federated learning framework and local loss optimization, particularly in wearable healthcare scenarios marked by noisy sensor data. Fig. 9. Open in a new tab Performance comparison between original and noisy data across training rounds. Repeated-measures ANOVA confirmed significant differences in accuracy and AUC ( p < 0.05) Figure 9 delineates a comparative analysis of accuracy, AUC-ROC, latency, and false positive rate between original and noisy data during 50 training rounds, thereby elucidating the model’s performance patterns. The findings demonstrate a progressive reduction in accuracy and AUC-ROC for noisy data, accompanied by heightened variability in Latency and a more significant rise in the false positive rate relative to the original data. These trends enhance the noise-level-specific analysis in Fig. 10 , underscoring the model’s resilience under varying training settings. Fig. 10. Open in a new tab Impact of noise level on performance metrics. ANOVA with Tukey’s HSD confirmed significant differences for accuracy and F1-score ( p < 0.001) Semi-FL demonstrated consistent performance among all clients, despite variations in their data. Figure 11 presents confusion matrices from five customers, comparing true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN) for Semi-FL with Baseline FL, MEWS, and centralized models. Semi-FL consistently achieved the highest counts of true positives (TP) and true negatives (TN), attaining TP = 9 and TN = 9, thus illustrating its robustness in decentralized settings. The latency performance was evaluated. Figure 12 illustrates that the average prediction latency for Semi-FL is 130 ms, whereas the baseline FL exhibits a latency of 180 ms and centralized models demonstrate a latency of 250 ms. This low-latency response enables real-time monitoring applications. Figure 13 further substantiates the model’s inference stability by comparing risk score trajectories over a 24-h period across multiple frameworks. Semi-FL aligns with the real event at time point 20, triggering alarms at point 15 and peaking at 0.9 around point 25. Competing models either delayed alarms (e.g., Transformer-based Local Attention) or triggered them early (e.g., Deep-Semi-FL), potentially compromising clinical reliability. Semi-FL demonstrated consistent performance among all clients, despite variations in their data. Fig. 11. Open in a new tab Confusion matrices of comparison approaches on local multimodal data for five client cases Fig. 12. Open in a new tab Prediction latency comparison across models. One-way ANOVA showed significant differences ( p < 0.001) Fig. 13. Open in a new tab Prediction probability comparison for 24-h clinical outcome Robustness to Noise and Explainability To evaluate resilience to data manipulation, Gaussian noise was artificially introduced to ECG signal inputs at three levels: 10%, 20%, and 30% variance relative to signal amplitude. With 10% noise, the model’s accuracy diminished slightly to 95.0%, while the AUC decreased to 0.96, indicating minimal degradation. At 20% noise, the accuracy reached 93.1%, although the AUC diminished to 0.94. At the peak assessed distortion of 30% noise, accuracy decreased to 90.5%, while AUC declined to 0.91. Despite these reductions, Semi-FL consistently outperformed baseline models at comparable noise levels, maintaining high F1-scores and lower false positive rates. These findings validate the resilience of the model’s local loss design. The combined use of Transformer-based Local Attention, which is resistant to slight disturbances, and neural networks, which excel at pattern recognition, improves noise tolerance. Moreover, Fig. 7 demonstrates that even with 30% noise, the evaluation loss did not surpass 0.35, and convergence remained stable, indicating resilience in both static and dynamic contexts. The evaluation of interpretability was conducted using LIME. The LIME summary graphic reveals that cholesterol and heart rate are the primary features, accounting for 42% and 28% of the predictive influence, respectively (Fig. 14 ). Blood pressure constituted 15%, followed by activity level at 10% and ECG-derived temporal features at 5%. This ranking aligns with recognized cardiovascular risk factors, confirming clinical validity. Furthermore, patient-level LIME visualizations offered explicit justifications for individual predictions, which is crucial for incorporation into decision-support systems. Fig. 14. Open in a new tab Global feature importance (LIME). Kruskal–Wallis test confirmed cholesterol and heart rate as significantly higher contributors ( p < 0.01) Integrated Semi-FL compressed communication updates to reduce execution costs during federated learning iterations. This approach led to an 8.2% reduction in communication costs, improving synchronization and reducing energy consumption in edge devices. The algorithm’s temporal complexity for inference was assessed as O ( n ), where n denotes the number of features post-fusion. The batch forecast duration was kept under 13 s, suitable for hospital-scale monitoring systems. Statistical analysis was performed on Figs. 8 , 9 , 10 , 12 , and 14 to ensure their significance. Analysis of variance (ANOVA) showed that there were significant changes in the feature associations at different noise levels for cholesterol ( F (3,16) = 12.45, p < 0.001) and heart rate ( F (3,16) = 9.87, p = 0.002) for Fig. 8 . Figures 9 and 10 show that both accuracy and F1 score decreased significantly in the presence of noise ( p < 0.001). Figure 12 shows that Semi-FL had significantly lower latency than Baseline FL and Centralized models ( p < 0.001). Finally, Kruskal–Wallis tests showed that cholesterol and heart rate were significantly more significant than the other parameters ( p < 0.01), as shown in Fig. 14 . The fact that performance got worse as noise levels went up gives us some information on how Semi-FL could be used in wearable healthcare settings. Wearable devices often show signal abnormalities in real life because of movement, interference from the environment, or sensors being placed incorrectly. This usually leads to data corruption of 5 to 25% [ 43 ]. Our results show that the framework keeps its accuracy above 90% and its AUC above 0.91, even when the noise level is at its peak, which is more than what most consumer-grade devices can handle. Discussion The Semi-FL framework works well for solving important problems in predicting cardiovascular risk, such as data heterogeneity, privacy protection, and making sense of the data. Combining Transformer-based Local Attention with CNN-based local feature extraction enhanced the accuracy of predictions while keeping the ability to make inferences in real time. Also, LIME and Grad-CAM made things clearer at both the feature and image levels, which helped build trust in the clinical setting and follow the rules. These results show that Semi-FL could be a useful way to handle healthcare settings where patient data cannot be stored in one place because of privacy issues. The improvements in accuracy, latency, and robustness seen in Semi-FL compared to classic federated learning and centralized models show how important adaptive aggregation and individualized fine-tuning are. This mixed technique uses high-resource nodes for full training and lightweight modification to work with low-resource edge devices. This kind of customization lowers the cost of communication and improves performance across a wide range of clients, making the system perfect for large-scale telehealth ecosystems. Even though these results are promising, there are certain problems that need to be fixed. To begin with, the study used datasets that were multimodal and clinically relevant, but only had a small number of real-world MRI pictures and mostly used retrospective data. Second, federated simulation with virtual clients is like real life, but deploying it across numerous institutions with different infrastructures may be harder. Third, the current examination only looked at short-term performance indicators. We still need to look into long-term stability, energy utilization on edge devices, and system scalability when there is a lot of client churn. Future studies should try to confirm the framework on bigger and more varied groups of people, such as prospective clinical trials in several hospitals. Additionally, adding federated transfer learning and more complex security features could make personalization and reliability better in situations that are different from each other. Conclusions This paper introduces Semi-FL, a federated explainable artificial intelligence platform designed for real-time cardiovascular disease risk prediction using multimodal data. Our methodology incorporates electronic health records, wearable sensor data, and cardiac MRI features, utilizing the combined benefits of Transformer-based Local Attention and neural networks within a unified local composite loss framework. The Semi-FL technique ensures that sensitive patient data remains on local clients, fully complying with privacy regulations such as GDPR and HIPAA, while modifications to the model are transmitted to develop a globally resilient predictor. Extensive research on the UCI Heart Disease benchmark and actual ECG/MRI datasets from multiple clinical locations shows the superior efficacy of Semi-FL relative to conventional federated and centralized models. The framework achieved an accuracy of 97.3%, an AUC-ROC of 0.98, and an F1-score of 96.1%, with an average inference latency of 130 ms. Notably, as simulated noise levels increased (10%, 20%, 30%), the system exhibited gradual degradation, maintaining an F1-score above 90% even at the highest noise level, while LIME-based explanations revealed adaptive feature reweighting that preserved interpretability. These findings validate that Semi-FL attains its objectives of elevated prediction accuracy, transparency, and privacy protection, making it suitable for continuous patient monitoring and early intervention in diverse healthcare settings. In the future, effort will be done to expand the Semi-FL framework in three main ways. Adding more types of data, such as genomic data or improved cardiac imaging (like CT angiography), could make predictions even more accurate and make care more individualized. Second, we will focus on real-world use in healthcare settings with few resources and edge devices to see how well they work when bandwidth and hardware are limited. Third, combining federated transfer learning and domain adaptation methodologies might make models more personalized for different groups of patients. Author Contribution All authors conceived and designed the study. XianFang Liu, ShunLei Li, Qin Zhu, ShaoKun Xu, and QinYang Jin collected, simulated, and analyzed data. “QinYang Jin” wrote the first draft, and all authors reviewed it. Funding Special Fund Project for Clinical Medical Research of Zhejiang Medical Association (2024ZYC-Z43). Zhejiang Medical and Health Science and Technology Plan Project (2025KY671). Data Availability The datasets generated and/or analyzed during the current study are not publicly available but are available from the corresponding author on reasonable request. Declarations Ethics Approval This study utilized the Cleveland Clinic Heart Patient Dataset sourced from the UCI Reference Website Dataset. This dataset is accessible to the public and can be utilized. Consent for Publication No private dataset was used, and therefore, no consent to publish is required. Competing interests The authors declare no competing interests. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Srinivasan, S. M., & Sharma, V. (2025). Applications of AI in cardiovascular disease detection—A review of the specific ways in which AI is being used to detect and diagnose cardiovascular diseases. AI in Disease Detection: Advancements and Applications , 123–146. 2. Jabarulla, M. Y., Uden, T., Jack, T., Beerbaum, P., & Oeltze-Jafra, S. (2024). Artificial Intelligence in Pediatric Echocardiography: Exploring Challenges, Opportunities, and Clinical Applications with Explainable AI and Federated Learning. arXiv preprint arXiv:2411.10255 . 3. Moshawrab, M., Adda, M., Bouzouane, A., Ibrahim, H., & Raad, A. (2023). Reviewing federated machine learning and its use in diseases prediction. Sensors , 23 (4), 2112. [ DOI ] [ PMC free article ] [ PubMed ] 4. Alam, M. U., & Rahmani, R. (2023). Fedsepsis: A federated multi-modal deep learning-based internet of medical things application for early detection of sepsis from electronic health records using raspberry pi and jetson nano devices. Sensors , 23 (2), 970. [ DOI ] [ PMC free article ] [ PubMed ] 5. Yen, T. T., Chen, I., Hua, M. W., Wei, C. Y., Shih, K. H., Li, J. L., ... & Jiang, R. S. (2021). A KCNQ4 c. 546C> G Genetic variant associated with late onset non-syndromic hearing loss in a Taiwanese population. Genes , 12 (11), 1711. [ DOI ] [ PMC free article ] [ PubMed ] 6. Shah, D., Patel, S., & Bharti, S. K. (2020). Heart disease prediction using machine learning techniques. SN Computer Science , 1 (6), 345. 7. Singh, P., Singh, S., & Pandi-Jain, G. S. (2018). Effective heart disease prediction system using data mining techniques. International journal of nanomedicine , 13 (sup1), 121-124. [ DOI ] [ PMC free article ] [ PubMed ] 8. Shah, P., Shukla, M., Dholakia, N. H., & Gupta, H. (2025). Predicting cardiovascular risk with hybrid ensemble learning and explainable AI. Scientific Reports , 15 (1), 1-21. [ DOI ] [ PMC free article ] [ PubMed ] 9. Fitriyani, N. L., Syafrudin, M., Alfian, G., & Rhee, J. (2020). HDPM: an effective heart disease prediction model for a clinical decision support system. Ieee Access , 8 , 133034-133050. 10. Coorey, G., Figtree, G. A., Fletcher, D. F., Snelson, V. J., Vernon, S. T., Winlaw, D., ... & Redfern, J. (2022). The health digital twin to tackle cardiovascular disease—a review of an emerging interdisciplinary field. NPJ digital medicine , 5 (1), 126. [ DOI ] [ PMC free article ] [ PubMed ] 11. Kavitha, M., Gnaneswar, G., Dinesh, R., Sai, Y. R., & Suraj, R. S. (2021, January). Heart disease prediction using hybrid machine learning model. In 2021 6th international conference on inventive computation technologies (ICICT) (pp. 1329–1333). IEEE. 12. Xia LX, Xiao YY, Jiang WJ, Yang XY, Tao H, Mandukhail SR, Qin JF, Pan QR, Zhu YG, Zhao LX, Huang LJ, Li Z, Yu XY. Exosomes derived from induced cardiopulmonary progenitor cells alleviate acute lung injury in mice. ActaPharmacol Sin. 2024 Apr 8. 10.1038/s41401-024-01253-4. [ DOI ] [ PMC free article ] [ PubMed ] 13. Wu Q, Zou S, Liu W, Liang M, Chen Y, Chang J, Liu Y, Yu X. A novel onco-cardiological mouse model of lung cancer-induced cardiac dysfunction and its application in identifying potential roles of tRNA-derived small RNAs. Biomed Pharmacother. 2023 Sep;165:115117. 10.1016/j.biopha.2023.115117. [ DOI ] [ PubMed ] 14. Ahamed, J., Manan Koli, A., Ahmad, K., Alam Jamal, M., & Gupta, B. B. (2022). CDPS-IoT: cardiovascular disease prediction system based on IoT using machine learning. 15. Moshawrab, M., Adda, M., Bouzouane, A., Ibrahim, H., & Raad, A. (2023). Reviewing multimodal machine learning and its use in cardiovascular diseases detection. Electronics , 12 (7), 1558. 16. Xu, G., Fan, X., Xu, S., Cao, Y., Chen, X., Shang, T.,... Yu, S. (2025). Anonymity-Enhanced Sequential Multi-Signer Ring Signature for Secure Medical Data Sharing in IoMT. IEEE Transactions on Information Forensics and Security, 20, 5647–5662. 10.1109/TIFS.2025.3574959 17. Wu, X., Zou, B., Lu, C., Wang, L., Zhang, Y.,... Wang, H. (2025). Dynamic Security Computing Framework With Zero Trust Based on Privacy Domain Prevention and Control Theory. IEEE Journal on Selected Areas in Communications, 43(6), 2266–2278. 10.1109/JSAC.2025.3560036 18. Qian K, Bao Z, Zhao Z, Koike T, Dong F, Schmitt M, Dong Q, Shen J, Jiang W, Jiang Y, et al. Learning Representations from Heart Sound: A Comparative Study on Shallow andDeep Models. Cyborg Bionic Syst.2024;5:Article 0075. 10.34133/cbsystems.0075 [ DOI ] [ PMC free article ] [ PubMed ] 19. Chen, W., Liu, Y., Wang, C., Zhu, J., Li, G., Liu, C.,... Lin, L. (2025). Cross-Modal Causal Representation Learning for Radiology Report Generation. IEEE Transactions on Image Processing, 34, 2970–2985. 10.1109/TIP.2025.3568746 [ DOI ] [ PubMed ] 20. Perone, F., Bernardi, M., Redheuil, A., Mafrica, D., Conte, E., Spadafora, L., ... & Sabouret, P. (2023). Role of cardiovascular imaging in risk assessment: recent advances, gaps in evidence, and future directions. Journal of Clinical Medicine , 12 (17), 5563 [ DOI ] [ PMC free article ] [ PubMed ] 21. Zeleznik, R., Foldyna, B., Eslami, P., Weiss, J., Alexander, I., Taron, J., ... & Aerts, H. J. (2021). Deep convolutional neural networks to predict cardiovascular risk from computed tomography. Nature communications , 12 (1), 715. [ DOI ] [ PMC free article ] [ PubMed ] 22. Commandeur, F., Slomka, P. J., Goeller, M., Chen, X., Cadet, S., Razipour, A., ... & Dey, D. (2020). Machine learning to predict the long-term risk of myocardial infarction and cardiac death based on clinical risk, coronary calcium, and epicardial adipose tissue: a prospective study. Cardiovascular research , 116 (14), 2216–2225. [ DOI ] [ PMC free article ] [ PubMed ] 23. Tsampras, T., Karamanidou, T., Papanastasiou, G., & Stavropoulos, T. G. (2025). Deep learning for cardiac imaging: Focus on myocardial diseases, a narrative review. Hellenic Journal of Cardiology , 81 , 18-24. [ DOI ] [ PubMed ] 24. Song, W., Wang, X., Guo, Y., Li, S., Xia, B.,... Hao, A. (2024). CenterFormer: A Novel Cluster Center Enhanced Transformer for Unconstrained Dental Plaque Segmentation. IEEE Transactions on Multimedia, 26, 10965–10978. 10.1109/TMM.2024.3428349 25. Luan, S., Yu, X., Lei, S., Ma, C., Wang, X., Xue, X.,... Zhu, B. (2023). Deep learning for fast super-resolution ultrasound microvessel imaging. Physics in Medicine & Biology, 68(24), 245023. 10.1088/1361-6560/ad0a5a [ DOI ] [ PubMed ] 26. Wenxin Zhang, Ning Xu, Ning Zhao, A. A. Al-Barakati. Adaptive neural finite-time self-triggered control for nonstrict-feedback nonlinear systems with sensor faults. Robotic Intelligence and Automation, 10.1108/RIA-10-2024-0235, 2025 27. Xiangjun Wu, Shuo Ding, Ning Zhao, Huanqing Wang, Ben Niu, Neural-network-based event-triggered adaptive secure fault-tolerantcontainment control for nonlinear multi-agent systems underdenial-of-service attacks, Neural Networks, https:// doi.org/10.1016/j.neunet.2025.107725 [ DOI ] [ PubMed ] 28. Zixuan Huang, Hamid Reza Karimi, Ben Niu, Lun Li, Xudong Zhao. A Unified Adaptive Event‐Triggered Output Feedback Consensus for Multi‐Agent Systems With or Without Output Constraints. International Journal of Robust and Nonlinear Control, 2025, 35(4): 1390–1405. 29. Shihui Liu, Ben Niu, Ning Zhao, Guangdeng Zong, Xudong Zhao, Lun Li. A novel event-triggered mechanism-based optimal safe control for nonlinear multi-player systems using adaptive dynamic programming, Journal of the Franklin Institute, 362(11): 107761, 2025 30. M. Li, Y. Chen, J. Li."Event-Triggered Fault-Tolerant Control for Nonlinear Multilateral Teleoperation System With Unknown Environmental Forces,"International Journal of Robust and Nonlinear Control, 35: 1831–1847, 2025. doi:[10.1002/rnc.7759] 31. Kosolwattana, T., Liu, C., Hu, R., Han, S., Chen, H., & Lin, Y. (2023). A self-inspected adaptive SMOTE algorithm (SASMOTE) for highly imbalanced data classification in healthcare. BioData Mining , 16 (1), 15. [ DOI ] [ PMC free article ] [ PubMed ] 32. Li, Z., Jiang, S., Xiang, F., Li, C., Li, S., Gao, T.,... Zhang, J. (2025). White patchy skin lesion classification using feature enhancement and interaction transformer module. Biomedical Signal Processing and Control, 107, 107819. 10.1016/j.bspc.2025.107819 33. Yang, B., Xu, S., Yin, L., Liu, C., & Zheng, W. (2025). Disparity estimation of stereo-endoscopic images using deep generative network. ICT Express, 11(1), 74-79. 10.1016/j.icte.2024.09.017 34. Yu, X., Luan, S., Lei, S., Huang, J., Liu, Z., Xue, X.,... Zhu, B. (2023). Deep learning for fast denoising filtering in ultrasound localization microscopy. Physics in Medicine & Biology, 68(20), 205002. 10.1088/1361-6560/acf98f [ DOI ] [ PubMed ] 35. Yaqoob, M. M., Nazir, M., Khan, M. A., Qureshi, S., & Al-Rasheed, A. (2023). Hybrid classifier-based federated learning in health service providers for cardiovascular disease prediction. Applied Sciences , 13 (3), 1911. 36. Zhang, Z., Wu, K., Wu, Z., Xiao, Y., Wang, Y., Lin, Q.,... Liu, Q. (2024). A case of pioneering subcutaneous implantable cardioverter defibrillator intervention in Timothy syndrome. BMC Pediatrics, 24(1), 729. 10.1186/s12887-024-05216-w [ DOI ] [ PMC free article ] [ PubMed ] 37. Wei, Y., Yan, M., Chen, R., Ding, Y., Xu, S., Li, H.,... Li, S. (2025). Efficacy and safety of Shenfu injection on bradyarrhythmia: A systematic review and meta-analysis. Medicine, 104(18) [ DOI ] [ PMC free article ] [ PubMed ] 38. Ahad, R. A., Hussain, M. Z., Qamar, M. H., Mustafa, M., Sattar, B., Ali, J., & Altaf, J. (2024). Heart Attack Prediction Using Federated Learning on Distributed Medical Data. UCP Journal of Engineering & Information Technology , 2 (1), 36-44. 39. Johri, A. M., Nambi, V., Naqvi, T. Z., Feinstein, S. B., Kim, E. S., Park, M. M., ... & Sillesen, H. (2020). Recommendations for the assessment of carotid arterial plaque by ultrasound for the characterization of atherosclerosis and evaluation of cardiovascular risk: from the American Society of Echocardiography. Journal of the American Society of Echocardiography , 33 (8), 917–933. [ DOI ] [ PubMed ] 40. Linardos, A., Kushibar, K., Walsh, S., Gkontra, P., & Lekadir, K. (2022). Federated learning for multi-center imaging diagnostics: a simulation study in cardiovascular disease. Scientific Reports , 12 (1), 3551. [ DOI ] [ PMC free article ] [ PubMed ] 41. Qayyum, A., Ahmad, K., Ahsan, M. A., Al-Fuqaha, A., & Qadir, J. (2022). Collaborative federated learning for healthcare: Multi-modal covid-19 diagnosis at the edge. IEEE Open Journal of the Computer Society , 3 , 172-184. 42. Liao, Y., Tang, Z., Gao, K., & Trik, M. (2024). Optimization of resources in intelligent electronic health systems based on Internet of Things to predict heart diseases via artificial neural network. Heliyon , 10 (11). [ DOI ] [ PMC free article ] [ PubMed ] 43. Rahmati, M. (2025). A Unified Framework for Explainable and Privacy-Preserving Machine Learning in Real-Time Decision-Making Systems. 44. Reddy, S. R., & Murthy, G. V. (2025). Cardiovascular Disease Prediction Using Particle Swarm Optimization and Neural Network Based an Integrated Framework. SN Computer Science , 6 (2), 186. 45. Johnson, A. E., Bulgarelli, L., Shen, L., Gayles, A., Shammout, A., Horng, S., ... & Mark, R. G. (2023). MIMIC-IV, a freely accessible electronic health record dataset. Scientific data , 10 (1), 1. [ DOI ] [ PMC free article ] [ PubMed ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement The datasets generated and/or analyzed during the current study are not publicly available but are available from the corresponding author on reasonable request. Articles from Journal of Imaging Informatics in Medicine are provided here courtesy of Springer ACTIONS View on publisher site PDF (3.0 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases 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

Related documents

• Interpretable Semi-federated Learning for Multimodal Cardiac Imaging and Risk Stratification: A Privacy-Preserving Framework.#263327
• Reading the Whole Heart: Latent-Attention Masked Autoencoders for Multimodal Cardiac Representation Learning#1000009
• Efficient differential privacy enabled federated learning model for detecting COVID-19 disease using chest X-ray images#177265
• Self-supervised representation learning reveals explainable physiological structure in high-dimensional magnetocardiography#294206
• Self-supervised representation learning reveals explainable physiological structure in high-dimensional magnetocardiography.#255334