Prediction of neurological outcome after pediatric cardiac arrest using heart rate variability and machine learning.

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 Resuscitation . Author manuscript; available in PMC: 2026 Apr 13. Published in final edited form as: Resuscitation. 2026 Jan 5;219:110961. doi: 10.1016/j.resuscitation.2026.110961 Search in PMC Search in PubMed View in NLM Catalog Add to search Prediction of neurological outcome after pediatric cardiac arrest using heart rate variability and machine learning Luiz EV Silva Luiz EV Silva a Tsui Laboratory, Department of Biomedical and Health Informatics, Children’s Hospital of Philadelphia, Perelman School of Medicine at the University of Pennsylvania, USA Find articles by Luiz EV Silva a , Daniel Balcarcel Daniel Balcarcel b Department of Anesthesiology and Critical Care Medicine, Children’s Hospital of Philadelphia, Perelman School of Medicine at the University of Pennsylvania, USA Find articles by Daniel Balcarcel b , Tiffany S Ko Tiffany S Ko b Department of Anesthesiology and Critical Care Medicine, Children’s Hospital of Philadelphia, Perelman School of Medicine at the University of Pennsylvania, USA Find articles by Tiffany S Ko b , Ryan W Morgan Ryan W Morgan b Department of Anesthesiology and Critical Care Medicine, Children’s Hospital of Philadelphia, Perelman School of Medicine at the University of Pennsylvania, USA Find articles by Ryan W Morgan b , Robert A Berg Robert A Berg b Department of Anesthesiology and Critical Care Medicine, Children’s Hospital of Philadelphia, Perelman School of Medicine at the University of Pennsylvania, USA Find articles by Robert A Berg b , Alexis Topjian Alexis Topjian b Department of Anesthesiology and Critical Care Medicine, Children’s Hospital of Philadelphia, Perelman School of Medicine at the University of Pennsylvania, USA Find articles by Alexis Topjian b , Fuchiang (Rich) Tsui Fuchiang (Rich) Tsui a Tsui Laboratory, Department of Biomedical and Health Informatics, Children’s Hospital of Philadelphia, Perelman School of Medicine at the University of Pennsylvania, USA b Department of Anesthesiology and Critical Care Medicine, Children’s Hospital of Philadelphia, Perelman School of Medicine at the University of Pennsylvania, USA c Department of Biostatistics, Epidemiology, and Informatics, Perelman School of Medicine at the University of Pennsylvania, USA Find articles by Fuchiang (Rich) Tsui a, b, c , Matthew P Kirschen Matthew P Kirschen b Department of Anesthesiology and Critical Care Medicine, Children’s Hospital of Philadelphia, Perelman School of Medicine at the University of Pennsylvania, USA Find articles by Matthew P Kirschen b, * Author information Article notes Copyright and License information a Tsui Laboratory, Department of Biomedical and Health Informatics, Children’s Hospital of Philadelphia, Perelman School of Medicine at the University of Pennsylvania, USA b Department of Anesthesiology and Critical Care Medicine, Children’s Hospital of Philadelphia, Perelman School of Medicine at the University of Pennsylvania, USA c Department of Biostatistics, Epidemiology, and Informatics, Perelman School of Medicine at the University of Pennsylvania, USA * Corresponding author at: Department of Anesthesiology and Critical Care Medicine, Perelman School of Medicine at the University of Pennsylvania, The Children’s Hospital of Philadelphia, 3401 Civic Center Blvd, Philadelphia, PA 19104, USA. [email protected] , [email protected] (M.P. Kirschen). Issue date 2026 Feb. This is an open access article under the CC BY-NC-ND license ( http://creativecommons.org/licenses/by-nc-nd/4.0/ ). PMC Copyright notice PMCID: PMC13070295  NIHMSID: NIHMS2163412  PMID: 41500411 The publisher's version of this article is available at Resuscitation Abstract Aims: Heart rate variability (HRV), a non-invasive measure of autonomic function, may offer prognostic value after pediatric cardiac arrest. We used machine learning models to determine whether HRV features within the first 24 h after return of spontaneous circulation can predict outcomes in children following cardiac arrest, and whether adding clinical cardiac arrest characteristics improves model performance. Methods: Retrospective study of children who received post-arrest care in the PICU at the Children’s Hospital of Philadelphia from 2020 to 2023. Thirty-six HRV features were extracted from ECG recordings and Extreme Gradient Boosting (XGB) models were trained to predict unfavorable neurological outcome, defined as Pediatric Cerebral Performance Category 4–6 and an increase >1 from baseline. Models were evaluated by cross-validation across the entire 24-h period and within sequential 6-h epochs. Additional models included clinical arrest characteristics. Performance was assessed by area under the receiver operating characteristic curve (AUROC). Results: Of the 75 patients who met inclusion criteria (median age 6.8 [IQR 10.4] years), 51% had an unfavorable outcome. Model considering HRV features and age achieved an AUROC of 0.80 (95% CI: 0.68–0.88). Top HRV predictors included standard deviation (SDNN), power at very low and low frequency bands, entropy, and fractal scaling. Performance was similar across the 6-h epochs (p’s > 0.1). Adding cardiac arrest characteristics did not improve model performance (AUROC 0.83 [0.73–0.92], p > 0.41). Conclusion: Using machine learning, HRV features within 24 h after pediatric cardiac arrest predict unfavorable outcome with AUROC 0.8. Adding clinical variables did not improve model performance. Keywords: Artificial intelligence, Time series, Children, Brain, Injury, Neuroprognostication Introduction Hypoxic-ischemic brain injury is the leading cause of morbidity and mortality following pediatric cardiac arrest. 1 , 2 Accurate and timely neuroprognostication is essential for guiding clinical decision-making and informing meaningful discussions with families. 3 Several clinical features routinely assessed during post-cardiac arrest care, including the neurological examination, electroencephalography (EEG), neuroimaging, and biomarkers, are associated with patient outcomes. However, no single modality provides sufficient predictive accuracy in the first 24 h post arrest. Thus, the American Heart Association recommends clinicians use a multimodal approach to neuroprognosticate, and highlights that predictions are not sufficient until 72-h post arrest. 4 To further enhance the accuracy of prognostic models, additional tools that reflect the severity of post-cardiac arrest brain injury in the first 24 h post-arrest are needed. Technologies already integrated into routine clinical care offer the advantage of being more cost effective and potentially implementable in resource limited settings. Heart rate variability (HRV) reflects beat-to-beat fluctuations in heart rate mediated by the autonomic nervous system. Higher HRV reflects greater autonomic flexibility and is generally associated with better health, whereas lower HRV indicates reduced autonomic activity and can be observed in patients with acute brain injury. 5 – 8 In children, abnormal HRV is associated with poor outcomes following traumatic brain injury and stroke. 9 – 11 The association between HRV impairment and outcomes after pediatric cardiac arrest remains understudied. 12 HRV metrics can be derived from standard electrocardiogram (ECG) recordings and are typically categorized into time-domain, frequency domain, and non-linear measures, each capturing different aspects of autonomic regulation. We aimed to determine whether HRV features predict neurologic outcomes in the first 24 h after pediatric cardiac arrest using machine learning models, and whether adding clinical cardiac arrest characteristics improved models’ predictive performance. We hypothesized that machine learning models can accurately predict outcomes using HRV features from the first 24 h after return of spontaneous circulation (ROSC) and that model performance would not improve after adding clinical variables. Methods This is a retrospective, observational study of children <18 years old who received post-arrest care in the pediatric intensive care unit (PICU) at the Children’s Hospital of Philadelphia (CHOP) between March 2020 and May 2023. Post arrest patients from our neuromonitoring database who were reported in a previous EEG study 13 were included if they had available raw ECG waveform data from the first 24 h after ROSC. Clinical cardiac arrest characteristics were abstracted from an institutional cardiac arrest database. The study was approved by CHOP’s Institutional Review Board (IRB 19–016118) with a waiver of consent. Neurologic outcome Neurological outcome was determined using the Pediatric Cerebral Performance Category (PCPC), a 6-point scale of neurologic function ranging from normal (1) to death (6). 14 , 15 PCPC scores were computed pre-arrest and post-arrest at hospital discharge or 30 days following arrest, whichever was earlier. Trained reviewers determined PCPC scores using information from the medical record. Unfavorable outcome was defined as a post-arrest PCPC of 4, 5, or 6, and a change of >1 from pre-arrest baseline. ECG processing Lead II ECG waveforms were recorded by a GE Carescape module at 300 Hz and stored using a Moberg CNS Monitor (Natus Medical Inc., USA). The waveforms underwent notch (60 Hz) and high-pass (0.5 Hz) filtering to remove noise ( Fig. 1 ). We selected a single 15-min segment from each hour within the first 24 h after ROSC for HRV analysis. We used a 15-min segment to limit the presence of artifacts while allowing sufficient time to capture slow heart rate fluctuations. 16 To identify the highest quality segment from each hour for inclusion in the analyses, we constructed overlapping 15-min segments by applying a 10-second sliding window. The quality of each segment was determined by (1) QRS template matching and (2) identifying the presence of specific elements in the ECG waveform. 17 , 18 We included the highest quality segment from each hour in the analysis ( Supplemental Material ). Fig. 1 – Open in a new tab Overview of the study design. Step 1: Lead II ECG waveforms within the first 24 h after ROSC were filtered to remove noise (notch and band-pass filters) and the highest-quality 15-min ECG segment every hour was obtained. Step 2: RR series were created as the time differences between consecutive ECG R-peaks and artifacts (including R-peak misdetections and ectopic beats) were corrected by linear interpolation. Step 3: Thirty-six heart rate variability (HRV) features were extracted from all hourly 15-min segments, along with subject’s age, to train extreme gradient boosting (XGB) machine learning models. Step 4: Feature vectors from epochs 0–6, 6–12, 12–18, and 18–24 h were used to train separate models, as well as all feature vectors obtained in the full 24 h after resuscitation. RR series generation We next generated a R-to-R interval time series (RR series) from each 15-min ECG segment and used that series to extract the HRV features that were ultimately included in the machine learning model. First, the R-peaks were detected using the “neurokit” peak detector from Neurokit2 package, which has shown the best overall performance over other common R-peak detectors. 18 We then computed the RR interval series as the time between successive R-peaks. Because the automated R-peak detector may occasionally misidentify non-sinus beats or miss true R-peaks, the calculated RR intervals can become spuriously short or long. Such non-sinus-beat intervals represent artifacts in the RR interval series, so we applied a procedure to identify and adjust these cases. For each RR series, we first generated a reference series using a sliding window median over 12 consecutive RR intervals. We then calculated the standard deviation of the RR intervals and constructed upper and lower tolerance thresholds by shifting the baseline up and down by 2.5 times this value. RR intervals falling outside these thresholds were considered artifacts and replaced by linear interpolation. 19 HRV feature extraction We extracted 36 HRV features from the RR series that were categorized into three groups: time-domain, frequency-domain, and nonlinear dynamics. Time-domain features included mean RR interval, standard deviation of normal-to-normal intervals (SDNN), root mean square of successive differences (RMSSD), percentage of RR intervals greater than 20 ms (pNN20), triangular index, and triangular interpolation of interval histogram (TINN). 16 Frequency-domain features included spectral power in the very low frequency (VLF: 0–0.04 Hz), low-frequency (LF: 0.04–0.15 Hz), and high-frequency (HF: 0.15–0.4 Hz) bands, as well as the LF/HF ratio. 16 Nonlinear dynamics features included heart rate acceleration and deceleration capacity, 20 three asymmetry indices (Porta’s, Guzik’s, and Ehlers’ 21 ) and five entropy measures including sample entropy, 22 distribution entropy, 23 permutation entropy, 24 dispersion entropy, 25 and attention entropy. 26 In addition, we computed the detrended fluctuation analysis scaling exponent using window sizes from 5 to 100. 27 Heart rate fragmentation was characterized using eight measures: the percentage of inflection points (PIP); the frequency of words with 0, 1, 2, or 3 inflection points (W0, W1, W2, W3); and the frequency of words composed by only hard (WH), only soft (WS), and mixed (WM) types of inflection types. 28 Symbolic dynamics patterns were estimated using two approaches: the equal probability method and the binary method. 29 In the former, we quantified the frequency of words with zero (0V), one (1V), two-like (2LV), or two-unlike (2UV) variations. In the latter, the binary nature of patterns only allows estimation of words with 0V, 1V, or two variations (2V). Machine learning training and evaluation We used eXtreme Gradient Boosting (XGB) algorithm to train classification models for predicting unfavorable outcome. HRV features were analyzed across the entire 24-h period as well as within sequential 6-h epochs (i.e., 0–6 h, 6–12 h, 12–18 h, 18–24 h). Models were trained such that each hourly 15-min ECG segment was treated as an independent sample. Each patient could have a maximum of twenty-four 15-min segments analyzed. To avoid performance overestimation, the median predicted probability from all samples from each patient was used for the final model performance estimation. We first trained a model to predict unfavorable outcome using HRV features and age. Age was included as developmental-related changes in the autonomic nervous system influence HRV throughout childhood and adolescence. 30 , 31 To examine the impact of clinical cardiac arrest characteristics on model performance, we subsequently trained a model using HRV features, age, and clinical cardiac arrest characteristics including CPR duration, first lactate level within six hours after ROSC, and whether the cardiac arrest was witnessed. Model performance was evaluated using a nested cross-validation with 3 inner folds for model selection and 5 outer folds for model evaluation ( Supplement Material ). Model performance was quantified by the area under the receiver operating characteristic (AUROC) and precision-recall curves (AUPRC). Bootstrapping with 2000 repetitions was used to calculate AUROC and AUPRC 95% confidence intervals. 32 Calibration plots were estimated by dividing the patients into tertiles based on the probability of unfavorable outcome. Expected calibration error (ECE) 33 was used to quantify model calibration, and a paired bootstrap procedure on the ECE was employed to assess whether the calibration differed significantly between models with and without clinical cardiac arrest characteristics. The relative importance of features was evaluated using gain, which reflects how much a feature improves the model’s decision-making when used to split data, averaged across all trees generated by the XGB model. We reported the relative contribution of the top 10 features as a percentage of total model gain across all features, obtained from the final XGB model (i.e., a model trained using the full sample). As sensitivity analyses, we compared model performance between sex and race categories, as well as for patients with and without baseline neurodevelopmental disabilities (pre-arrest PCPC 3 or 4) since patients with chronic brain injury may have abnormal HRV at baseline. Data processing and analyses were performed using Python 3.11. Data were reported according to the TRIPOD+AI Statement. 34 Results A total of 75 patients (median age 6.8 [1.2, 11.6], 48% female) were analyzed. Of these, 51% (38/75) had an unfavorable outcome ( Table 1 ). Median age and pre-arrest PCPC scores were similar between outcome groups (p = 0.45 and 0.29, respectively). Patients with unfavorable outcome had longer duration CPR, a greater proportion of unwitnessed arrests, and higher initial post-arrest lactate than patients with favorable outcome. All patients were mechanically ventilated and the proportion of patients receiving targeted temperature management to 33°C or an epinephrine or norepinephrine infusion were similar between favorable and unfavorable outcome groups (p = 0.77 and 0.68, respectively). Among all patients, median time from ROSC to initiation of ECG recording in the Moberg neuromonitoring system was 4.6 [3.2, 7.5] h and ECG was available for 18.6 [14.1, 20.6] h within the first 24 h after ROSC, with no differences between outcome groups ( p ’s > 0.9). Table 1 – Clinical characteristics. Clinical characteristics All ( N = 75) Favorable ( N = 37) Unfavorable ( N = 38) P -value Age (years) 6.8 [1.2, 11.6] 7.3 [1.3, 12.8] 6.3 [1.1, 11.4] 0.45 Female 36 (48) 19 (51) 17 (45) 0.73 Race African American 21 (28) 9 (24) 12 (32) 0.90 White 28 (37) 14 (38) 14 (37) Other 22 (30) 12 (33) 10 (26) Unknown 4 (5) 2 (5) 2 (5) Past medical history None 20 (27) 4 (11) 16 (42) 0.03 Respiratory 19 (25) 10 (27) 9 (24) Cardiac 5 (7) 4 (11) 1 (2) Neurologic 24 (32) 15 (40) 9 (24) Other 7 (9) 4 (11) 3 (8) Pre-arrest PCPC 1 51 (68) 23 (62) 28 (74) 0.29 2 6 (8) 2 (5.5) 4 (10.5) 3 4 (5) 2 (5.5) 2 (5) 4 14 (19) 10 (27) 4 (10.5) Out-hospital cardiac arrest 43 (57) 14 (38) 29 (76) Witnessed cardiac arrest 50 (67) 33 (89) 17 (45) <0.01 Cause of arrest Arrhythmia 9 (12) 6 (16) 3 (8) <0.01 Hypotension/shock 17 (23) 14 (38) 3 (8) Respiratory decompensation 40 (53) 16 (43) 24 (63) Trauma 3 (4) 1 (3) 2 (5) Unknown 6 (8) — 6 (16) Epinephrine doses during CPR 0 21 (28) 16 (43) 5 (13) <0.01 1 14 (19) 7 (19) 7 (19) 2 9 (12) 6 (16) 3 (8) 3+ 25 (33) 7 (19) 18 (47) Unknown 6 (8) 1 (3) 5 (13) CPR duration (minutes) a 19.9 [5.0, 26.8] 9.5 [3.0, 11.0] 30.4 [10.0, 45.0] <0.01 Post-arrest care Initial lactate (mmol/L) b 7.8 [2.9, 11.4] 6.2 [2.0, 9.3] 9.4 [4.9, 13.3] 0.01 Epinephrine or norepinephrine infusion 67 (89) 32 (86) 35 (92) 0.68 Mechanical ventilation 75 (100) 37 (100) 38 (100) 1.00 Targeted temperature management to 33°C 10 (13) 4 (11) 6 (16) 0.77 Hospital discharge PCPC 1 10 (13) 10 (27) – <0.01 2 7 (10) 7 (19) – 3 10 (13) 10 (27) – 4 20 (27) 10 (27) 10 (26) 5 – – – 6 28 (37) – 28 (74) Open in a new tab Numerical variables are median [1st, 3rd quartiles]. Categorical variables are the number of subjects (% of total). CA: cardiac arrest; CPR: cardiopulmonary resuscitation; PCPC: pediatric cerebral performance category. Lactate was measured within 6 h after resuscitation. a Missing one value in Unfavorable group. b Missing six values in Favorable and eight in Unfavorable group. The machine learning model trained with HRV features and age showed an AUROC of 0.80 (0.68–0.88) and AUPRC of 0.80 (0.70–0.88) ( Fig. 2 ). When cardiac arrest characteristics were added to the model, AUROC was 0.84 (0.73–0.92) and AUPRC was 0.84 (0.73–0.92). AUROCs were similar between models with and without clinical cardiac arrest characteristics ( p = 0.41). Expected calibration error was 0.09 (0.03–0.18) and 0.13 (0.04–0.24) for models with and without clinical cardiac arrest characteristics, respectively, with no difference between them ( Fig. S1 ). Model performance did not differ across the four 6-h epochs ( Fig. 2 ). Fig. 2 – Open in a new tab AUROC (A) and AUPRC (B) [95% confidence interval] for XGB classification models of cerebral outcome. Performance of hypothetical random models (0.5 for AUROC and unfavorable prevalence for AUPRC) is shown as red dots. CA: cardiac arrest; AUROC: area under the receiver operating characteristic curve; AUPRC: area under the precision-recall curve. The most important features in the full 24-h XGB model were SDNN, TINN, and triangular index (time-domain), power at LF and VLF bands (frequency-domain), and entropy, heart rate fragmentation, and fractal scaling (nonlinear dynamics) ( Fig. 3 , Table S1 ). Fig. 3 – Open in a new tab Top 10 features assigned by the XGB model trained using the full 24 h of recording. Importance of features are represented by the percentage of gain, considering the sum of gains from all features. Gain represents the average contribution of a feature when it is used to split data in a decision tree created by XGB. HRV: heart rate variability; CA: cardiac arrest. For description of HRV feature acronyms, see the text. In sensitivity analyzes, we observed no difference between sex and race categories when the full epoch (0–24 h) was considered ( p ’s > 0.1) ( Fig. S2 ). In addition, model performance for patients with baseline neurodevelopmental disabilities (i.e., pre-arrest PCPC of 3 or 4) were similar compared to patients with pre-arrest PCPC of 1 or 2 ( Fig. S3 ). Discussion We demonstrated that HRV features during the first 24 h following pediatric cardiac arrest predicted unfavorable outcome with an AUROC of 0.8 and that prediction accuracy did not change over the first 24 h. Incorporating clinical features of the cardiac arrest into the model did not improve prediction accuracy. These results are similar to prior pediatric studies which demonstrated that EEG-based outcome prediction models have AUROCs between 0.8 and 0.9. 13 , 35 , 36 The favorable AUROC and calibration of HRV-based models suggest that HRV may augment multimodal prognostic models that integrate EEG, neuroimaging, and other clinical data within the first 24 h after cardiac arrest and improve early risk stratification. Our findings support that hypoxic-ischemic brain injury after pediatric cardiac arrest can impair autonomic nervous system function. Similar associations have been reported in neonates with hypoxic-ischemic encephalopathy, where HRV features correlated not only with outcomes but also with other markers of brain injury derived from magnetic resonance imaging and EEG. 37 – 40 Reduced HRV has also been observed in patients with severe brain injury who progress to death by neurologic criteria. 10 , 11 These data suggest that HRV features may have value as a component of multimodal prognostication algorithms. Since HRV can be obtained early after ROSC, it could contribute to risk stratification models and potentially identify patients most likely to benefit from neuroprotective interventions. Given its dynamic nature, HRV could also be used to monitor severity of autonomic dysfunction over time, assess response to therapeutic interventions, and guide post-arrest clinical management. Additional work is needed to determine which specific time-domain, frequency-domain, or non-linear dynamic HRV features are most informative for these applications. In our cohort, very low frequency (VLF) and low frequency (LF) spectral features were reduced in children with unfavorable outcomes, suggesting that impaired cardiac sympathetic modulation may be associated with more severe post-cardiac arrest brain injury. VLF power is associated with slow physiological mechanisms, such as thermoregulation, renin-angiotensin-aldosterone, and other humoral effects, while the LF component is associated with cardiac sympathetic modulation. 16 , 41 , 42 Lower absolute spectral power in the VLF and LF bands, as well as a reduced LF/HF ratio have been associated with poor outcome after adult cardiac arrest as well. 43 – 49 While time- and frequency-domain HRV metrics assume linearity in the underlying heart rate regulation, nonlinear HRV metrics capture more complex dynamics that linear methods are not able to detect. 50 – 52 In adults, reduced DFA scaling exponent has been observed in patients with poor outcomes after cardiac arrest. 47 Two heart rate fragmentation features (W1 and W3) were selected by our XGB model as important features. Fragmentation of heart rate is a more contemporary approach that quantifies the transition patterns between acceleration and deceleration of heart rate. Higher fragmentation (i.e., increased number of transitions between heart rate acceleration and deceleration), has been associated with ageing, cardiovascular, and vascular brain injury, likely due to a combined neuroautonomic–electrophysiologic mechanism controlling the heart’s sinus node. 28 , 53 – 55 Approximately 20% of patients in our cohort had chronic neurodevelopmental disabilities prior to their cardiac arrest. Ten of these 14 patients had no change in their neurologic status after their arrest and were classified as having a favorable outcome. These patients may have had altered HRV at baseline due to impaired autonomic function from their chronic brain injury. 56 It was possible that the model classified these patients as unfavorable given their potential baseline HRV impairment, leading to overall decreased prediction accuracy. To address the potential influence of this confounder, we stratified the model results by pre-arrest PCPC and did not appreciate a prediction bias for patients with a pre-arrest PCPC 3 or 4. Although sample sizes in each group were small, these data suggest that the models were able to accurately delineate outcome for patients with pre-existing neurodevelopmental disabilities. This study has several limitations. First, as a retrospective, observational, and single-center study, there is inherent risk of selection bias and unmeasured confounders that could have affected both feature extraction and outcome assessment. Second, sedative medications, temperature, vasopressors and inotropes, and mechanical ventilation can affect HRV, the latter due to its impact on vagally-mediated sinus arrhythmia (a component of HRV) observed with natural respiration. 57 However, the proportion of patients receiving these therapies was similar between outcome groups, mitigating potential biases related to them. Third, due to the relatively small sample, we evaluated model performance using internal cross-validation without external validation. Finally, we were unable to evaluate the association between HRV features and other measures of brain injury severity (e.g., EEG, neuroimaging, or biomarkers). Conclusion Metrics of HRV from the first 24 h after pediatric cardiac arrest can predict unfavorable neurologic outcome with AUROC of 0.8. Adding clinical cardiac arrest characteristics did not improve prediction performance. HRV metrics should be considered as components of future multimodal prognostic and risk-stratification tools for pediatric cardiac arrest. Supplementary Material 1 NIHMS2163412-supplement-1.pdf (1,002.8KB, pdf) Source of funding This study was funded by NINDS [grant number K23NS116120], Russell C. Raphaely Endowed Chair in Critical Care Medicine (Topjian), and the CHOP Resuscitation Science Center. The funder had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication. Appendix A. Supplementary data Supplementary data to this article can be found online at https://doi.org/10.1016/j.resuscitation.2026.110961 . Footnotes CRediT authorship contribution statement Luiz E.V. Silva: Writing – review & editing, Writing – original draft, Validation, Software, Methodology, Formal analysis, Data curation, Conceptualization. Daniel Balcarcel: Writing – review & editing, Writing – original draft, Methodology, Conceptualization. Tiffany S. Ko: Writing – review & editing, Validation. Ryan W. Morgan: Writing – review & editing, Validation, Resources. Robert A. Berg: Writing – review & editing, Validation, Resources. Alexis Topjian: Writing – review & editing, Validation, Supervision, Resources, Methodology, Data curation. Fuchiang (Rich) Tsui: Writing – review & editing, Supervision, Resources. Matthew P. Kirschen: Writing – review & editing, Writing – original draft, Validation, Supervision, Resources, Methodology, Data curation, Conceptualization. Declaration of competing interest The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Dr. Alexis Topjian currently serves as an Editor for the journal Resuscitation. Data availability The datasets analyzed during the current study are not publicly available due to patient privacy concerns. Sharing of study data are subject to the establishment of data-sharing agreements between institutions, approved by the Children’s Hospital of Philadelphia. Code availability The underlying code for this study is not publicly available but may be made available to qualified researchers on reasonable request to the corresponding author. REFERENCES 1. Topjian AA, de Caen A, Wainwright MS, Abella BS, Abend NS, Atkins DL, et al. Pediatric post–cardiac arrest care: a scientific statement from the American Heart Association. Circulation 2019;140:e194–233. 10.1161/CIR.0000000000000697. [ DOI ] [ PubMed ] [ Google Scholar ] 2. Sekhon MS, Ainslie PN, Griesdale DE. Clinical pathophysiology of hypoxic ischemic brain injury after cardiac arrest: a “two-hit” model. Crit Care 2017;21:90. 10.1186/s13054-017-1670-9. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 3. Fischer D, Edlow BL. Coma prognostication after acute brain injury: a review. JAMA Neurol 2024. 10.1001/jamaneurol.2023.5634. [ DOI ] [ PubMed ] [ Google Scholar ] 4. Lasa JJ, Dhillon GS, Duff JP, Hayes J, Kamath-Rayne BD, Levy A, et al. Part 8: pediatric advanced life support: 2025 American Heart Association and American Academy of pediatrics guidelines for cardiopulmonary resuscitation and emergency cardiovascular care. Circulation 2025;152:S479–537. 10.1161/CIR.0000000000001368. [ DOI ] [ PubMed ] [ Google Scholar ] 5. Gang Y, Malik M. Heart rate variability in critical care medicine. Curr Opin Crit Care 2002;8:371–5. 10.1097/00075198-200210000-00002. [ DOI ] [ PubMed ] [ Google Scholar ] 6. Karmali SN, Sciusco A, May SM, Ackland GL. Heart rate variability in critical care medicine: a systematic review. Intensive Care Med Exp 2017;5:33. 10.1186/s40635-017-0146-1. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Johnston BW, Barrett-Jolley R, Krige A, Welters ID. Heart rate variability: measurement and emerging use in critical care medicine. J Intensive Care Soc 2020;21:148–57. 10.1177/1751143719853744. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Bodenes L, N’Guyen Q-T, Le Mao R, Ferrière N, Pateau V, Lellouche F, et al. Early heart rate variability evaluation enables to predict ICU patients’ outcome. Sci Rep 2022;12:2498. 10.1038/s41598-022-06301-9. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Goldstein B, Fiser DH, Kelly MM, Mickelsen D, Ruttimann U, Pollack MM. Decomplexification in critical illness and injury: relationship between heart rate variability, severity of illness, and outcome. Crit Care Med 1998;26:352–7. 10.1097/00003246-199802000-00040. [ DOI ] [ PubMed ] [ Google Scholar ] 10. Piantino JA, Lin A, Crowder D, Williams CN, Perez-Alday E, Tereshchenko LG, et al. Early heart rate variability and electroencephalographic abnormalities in acutely brain-injured children who progress to brain death. Pediatr Crit Care Med 2019;20:38–46. 10.1097/PCC.0000000000001759. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 11. LaRovere KL, Luchette M, Akhondi-Asl A, DeSouza BJ, Tasker RC, Mehta NM, et al. Heart rate change as a potential digital biomarker of brain death in critically ill children with acute catastrophic brain injury. Crit Care Explor 2023;5:e0908. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 12. Kwon SB, Megjhani M, Nametz D, Agarwal S, Park S. Heart rate and heart rate variability as a prognosticating feature for functional outcome after cardiac arrest: a scoping review. Resusc Plus 2023;15:100450. 10.1016/j.resplu.2023.100450. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Silva LEV, Chen C-C, Press CA, Graham K, Shi L, Abend NS, et al. Pediatric cardiac arrest outcome prediction using data-driven machine learning of early quantitative electroencephalogram (qEEG) features. Resuscitation 2025;216. 10.1016/j.resuscitation.2025.110854. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Fiser DH, Long N, Roberson PK, Hefley G, Zolten K, Brodie-Fowler M. Relationship of pediatric overall performance category and pediatric cerebral performance category scores at pediatric intensive care unit discharge with outcome measures collected at hospital discharge and 1- and 6-month follow-up assessments. Crit Care Med 2000;28:2616–20. 10.1097/00003246-200007000-00072. [ DOI ] [ PubMed ] [ Google Scholar ] 15. Topjian AA, Scholefield BR, Pinto NP, Fink EL, Buysse CMP, Haywood K, et al. P-COSCA (Pediatric Core Outcome Set for Cardiac Arrest) in children: an advisory statement from the International Liaison Committee on Resuscitation. Circulation 2020;142:e246–61. 10.1161/CIR.0000000000000911. [ DOI ] [ PubMed ] [ Google Scholar ] 16. Task Force of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology. Heart rate variability. Standards of measurement, physiological interpretation, and clinical use. Circulation 1996;93:1043. 10.1161/01.CIR.93.5.1043. [ DOI ] [ PubMed ] [ Google Scholar ] 17. Zhao Z, Zhang Y. SQI quality evaluation mechanism of single-lead ECG signal based on simple heuristic fusion and fuzzy comprehensive evaluation. Front Physiol 2018;9. 10.3389/fphys.2018.00727. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 18. Makowski D, Pham T, Lau ZJ, Brammer JC, Lespinasse F, Pham H, et al. NeuroKit2: a Python toolbox for neurophysiological signal processing. Behav Res 2021;53:1689–96. 10.3758/s13428-020-01516-y. [ DOI ] [ PubMed ] [ Google Scholar ] 19. Silva LEV, Fazan R, Marin-Neto JA. PyBioS: a freeware computer software for analysis of cardiovascular signals. Comput Methods Programs Biomed 2020;197:105718. 10.1016/j.cmpb.2020.105718. [ DOI ] [ PubMed ] [ Google Scholar ] 20. Bauer A, Kantelhardt JW, Barthel P, Schneider R, Mäkikallio T, Ulm K, et al. Deceleration capacity of heart rate as a predictor of mortality after myocardial infarction: cohort study. Lancet 2006;367:1674–81. 10.1016/S0140-6736(06)68735-7. [ DOI ] [ PubMed ] [ Google Scholar ] 21. Porta A, Casali KR, Casali AG, Gnecchi-Ruscone T, Tobaldini E, Montano N, et al. Temporal asymmetries of short-term heart period variability are linked to autonomic regulation. Am J Physiol Regul Integr Comp Physiol 2008;295:R550–7. 10.1152/ajpregu.00129.2008. [ DOI ] [ PubMed ] [ Google Scholar ] 22. Richman JS, Moorman JR. Physiological time-series analysis using approximate entropy and sample entropy. Am J Physiol Heart Circ Physiol 2000;278:H2039–49. [ DOI ] [ PubMed ] [ Google Scholar ] 23. Li P, Liu C, Li K, Zheng D, Liu C, Hou Y. Assessing the complexity of short-term heartbeat interval series by distribution entropy. Med Biol Eng Comput 2015;53:77–87. 10.1007/s11517-014-1216-0. [ DOI ] [ PubMed ] [ Google Scholar ] 24. Bandt C, Pompe B. Permutation entropy: a natural complexity measure for time series. Phys Rev Lett 2002;88:174102. [ DOI ] [ PubMed ] [ Google Scholar ] 25. Rostaghi M, Azami H. Dispersion entropy: a measure for time-series analysis. IEEE Signal Process Lett 2016;23:610–4. 10.1109/LSP.2016.2542881. [ DOI ] [ Google Scholar ] 26. Yang J, Choudhary GI, Rahardja S, Franti P. Classification of interbeat interval time-series using attention entropy. IEEE Trans Affect Comput 2020:1. 10.1109/TAFFC.2020.3031004. [ DOI ] [ Google Scholar ] 27. Peng CK, Havlin S, Stanley HE, Goldberger AL. Quantification of scaling exponents and crossover phenomena in nonstationary heartbeat time-series. Chaos 1995;5:82–7. [ DOI ] [ PubMed ] [ Google Scholar ] 28. Costa MD, Davis RB, Goldberger AL. Heart rate fragmentation: a symbolic dynamical approach. Front Physiol 2017;8:827. 10.3389/fphys.2017.00827. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 29. Cysarz D, Edelhäuser F, Javorka M, Montano N, Porta A. On the relevance of symbolizing heart rate variability by means of a percentile-based coarse graining approach. Physiol Meas 2018;39:105010. 10.1088/1361-6579/aae302. [ DOI ] [ PubMed ] [ Google Scholar ] 30. Silvetti MS, Drago F, Ragonese P. Heart rate variability in healthy children and adolescents is partially related to age and gender. Int J Cardiol 2001;81:169–74. 10.1016/S0167-5273(01)00537-X. [ DOI ] [ PubMed ] [ Google Scholar ] 31. Bobkowski W, Stefaniak ME, Krauze T, Gendera K, Wykretowicz A, Piskorski J, et al. Measures of heart rate variability in 24-h ECGs depend on age but not gender of healthy children. Front Physiol 2017;8:311. 10.3389/fphys.2017.00311. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 32. Carpenter J, Bithell J. Bootstrap confidence intervals: when, which, what? A practical guide for medical statisticians. Stat Med 2000;19:1141–64. [ DOI ] [ PubMed ] [ Google Scholar ] 33. Guo C, Pleiss G, Sun Y, Weinberger KQ. On calibration of modern neural networks. Proceedings of the 34th international conference on machine learning, vol. 70. Sydney, NSW, Australia: JMLR.org ; 2017. p. 1321–30. [ Google Scholar ] 34. Collins GS, Moons KGM, Dhiman P, Riley RD, Beam AL, Calster BV, et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ 2024;385:e078378. 10.1136/bmj-2023-078378. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 35. Lee S, Zhao X, Davis KA, Topjian AA, Litt B, Abend NS. Quantitative EEG predicts outcomes in children after cardiac arrest. Neurology 2019;92:e2329–38. 10.1212/WNL.0000000000007504. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 36. Hunfeld M, Verboom M, Josemans S, van Ravensberg A, Straver D, Lückerath F, et al. Prediction of survival after pediatric cardiac arrest using quantitative EEG and machine learning techniques. Neurology 2024;103:e210043. 10.1212/WNL.0000000000210043. [ DOI ] [ PubMed ] [ Google Scholar ] 37. Bersani I, Piersigilli F, Gazzolo D, Campi F, Savarese I, Dotta A, et al. Heart rate variability as possible marker of brain damage in neonates with hypoxic ischemic encephalopathy: a systematic review. Eur J Pediatr 2021;180:1335–45. 10.1007/s00431-020-03882-3. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 38. Andersen M, Andelius TCK, Pedersen MV, Kyng KJ, Henriksen TB. Severity of hypoxic ischemic encephalopathy and heart rate variability in neonates: a systematic review. BMC Pediatr 2019;19:242. 10.1186/s12887-019-1603-7. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 39. Goswami I, Kamino D, Widjaja E, Paniccia M, Mitsakakis N, Moore A, et al. Early neonatal heart rate variability patterns in different subtypes of perinatal hypoxic-ischemic brain injury. Pediatr Res 2022;92:1630–9. 10.1038/s41390-022-02016-9. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 40. Pavel AM, Mathieson SR, Livingstone V, O’Toole JM, Pressler RM, de Vries LS, et al. Heart rate variability analysis for the prediction of EEG grade in infants with hypoxic ischaemic encephalopathy within the first 12 h of birth. Front Pediatr 2023;10:1016211. 10.3389/fped.2022.1016211. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 41. Stauss HM. Heart rate variability. Am J Physiol Regul Integr Comp Physiol 2003;285:R927–31. 10.1152/ajpregu.00452.2003. [ DOI ] [ PubMed ] [ Google Scholar ] 42. Stauss HM. Identification of blood pressure control mechanisms by power spectral analysis. Clin Exp Pharmacol Physiol 2007;34:362–8. 10.1111/j.1440-1681.2007.04588.x. [ DOI ] [ PubMed ] [ Google Scholar ] 43. Endoh H, Kamimura N, Honda H, Nitta M. Early prognostication of neurological outcome by heart rate variability in adult patients with out-of-hospital sudden cardiac arrest. Crit Care 2019;23:323. 10.1186/s13054-019-2603-6. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 44. Riganello F, Zubler F, Haenggi M, De Lucia M. Heart rate complexity: an early prognostic marker of patient outcome after cardiac arrest. Clin Neurophysiol 2022;134:27–33. 10.1016/j.clinph.2021.10.019. [ DOI ] [ PubMed ] [ Google Scholar ] 45. Tiainen M, Parikka HJ, Mäkijärvi MA, Takkunen OS, Sarna SJ, Roine RO. Arrhythmias and heart rate variability during and after therapeutic hypothermia for cardiac arrest. Crit Care Med 2009;37:403–9. 10.1097/CCM.0b013e31819572c4. [ DOI ] [ PubMed ] [ Google Scholar ] 46. Chen W-L, Tsai T-H, Huang C-C, Chen J-H, Kuo C-D. Heart rate variability predicts short-term outcome for successfully resuscitated patients with out-of-hospital cardiac arrest. Resuscitation 2009;80:1114–8. 10.1016/j.resuscitation.2009.06.020. [ DOI ] [ PubMed ] [ Google Scholar ] 47. Benghanem S, Sharshar T, Gavaret M, Dumas F, Diehl J-L, Brechot N, et al. Heart rate variability for neuro-prognostication after CA: insight from the Parisian registry. Resuscitation 2024;202:110294. 10.1016/j.resuscitation.2024.110294. [ DOI ] [ PubMed ] [ Google Scholar ] 48. Estévez-Báez M, Machado C, García-Sánchez B, Rodríguez V, Alvarez-Santana R, Leisman G, et al. Autonomic impairment of patients in coma with different Glasgow coma score assessed with heart rate variability. Brain Inj 2019;33:496–516. 10.1080/02699052.2018.1553312. [ DOI ] [ PubMed ] [ Google Scholar ] 49. Zhang P, Roberts T, Richards B, Haseler LJ. Utilizing heart rate variability to predict ICU patient outcome in traumatic brain injury. BMC Bioinf 2020;21:481. 10.1186/s12859-020-03814-w. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 50. Sassi R, Cerutti S, Lombardi F, Malik M, Huikuri HV, Peng C-K, et al. Advances in heart rate variability signal analysis: joint position statement by the e-Cardiology ESC Working Group and the European Heart Rhythm Association co-endorsed by the Asia Pacific Heart Rhythm Society. Europace 2015;17:1341–53. 10.1093/europace/euv015. [ DOI ] [ PubMed ] [ Google Scholar ] 51. Silva LEV, Lataro RM, Castania JA, Silva CAA, Salgado HC, Fazan R, et al. Nonlinearities of heart rate variability in animal models of impaired cardiac control: contribution of different time scales. J Appl Physiol 2017;123:344–51. 10.1152/japplphysiol.00059.2017. [ DOI ] [ PubMed ] [ Google Scholar ] 52. Voss A, Schulz S, Schroeder R, Baumert M, Caminal P. Methods derived from nonlinear dynamics for analysing heart rate variability. Phil Trans R Soc A 2009;367:277–96. [ DOI ] [ PubMed ] [ Google Scholar ] 53. Heckbert SR, Jensen PN, Erus G, Nasrallah IM, Rashid T, Habes M, et al. Heart rate fragmentation and brain MRI markers of small vessel disease in MESA. Alzheimer’s Dement 2024;20:1397–405. 10.1002/alz.13554. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 54. Lensen IS, Monfredi OJ, Andris RT, Lake DE, Moorman JR. Heart rate fragmentation gives novel insights into non-autonomic mechanisms governing beat-to-beat control of the heart’s rhythm. JRSM Cardiovasc Dis 2020;9:2048004020948732. 10.1177/2048004020948732. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 55. da Silva TM, Silva CAA, Salgado HC, Fazan R, Silva LEV. The role of the autonomic nervous system in the patterns of heart rate fragmentation. Biomed Signal Process Control 2021;67:102526. 10.1016/j.bspc.2021.102526. [ DOI ] [ Google Scholar ] 56. Gaôsior JS, Zamunér AR, Silva LEV, Williams CA, Baranowski R, Sacha J, et al. Heart rate variability in children and adolescents with cerebral palsy-a systematic literature review. J Clin Med 2020;9:1141. 10.3390/jcm9041141. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 57. Skytioti M, Elstad M. Respiratory sinus arrhythmia is mainly driven by central feedforward mechanisms in healthy humans. Front Physiol 2022;13:768465. 10.3389/fphys.2022.768465. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials 1 NIHMS2163412-supplement-1.pdf (1,002.8KB, pdf) Data Availability Statement The datasets analyzed during the current study are not publicly available due to patient privacy concerns. Sharing of study data are subject to the establishment of data-sharing agreements between institutions, approved by the Children’s Hospital of Philadelphia. The underlying code for this study is not publicly available but may be made available to qualified researchers on reasonable request to the corresponding author. ACTIONS View on publisher site PDF (649.2 KB) 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