Characterizing Altered Cardiac Autonomic Dynamics in Psychotic Disorders Through Multidomain Heart Rate Variability Analysis and Deep Learning

Recent developments in wearable devices have opened the possibility of heart rate variability (HRV), a quantitative measure of autonomic nervous system (ANS) regulation, as a potential indicator for the early detection of psychiatric disorders. This study explores HRV for the detection of psychotic disorders by integrating multidomain HRV analysis with a hybrid deep learning technique. Polar H10 electrocardiogram (ECG) device was employed to collect RR interval data from 30 patients diagnosed with either schizophrenia (SZ) or bipolar disorder (BD), and 30 similar age healthy controls, with a duration of approximately 1.5-2 hours. HRV features, including time-domain, spectral, and nonlinear features such as the Poincaré plot and Recurrence Quantification Analysis (RQA), were extracted from this publicly available dataset. Group-wise statistical comparisons between the healthy and the psychotic disorders were performed using the Mann-Whitney U test with a significance level of p < 0.01. HRV indices such as Mean NN intervals, SDNN, RMSSD, triangular index, Cardiac Vagal Index, Poincaré plot measures (SD1 and SD2), Sample Entropy, and spectral power components (HF, LF, and Total Power), were significantly lower in the psychotic disorder group (p < 0.01), indicating attenuated autonomic modulation. All machine learning and deep learning models were conducted under a fully nested, subject-independent 5×4 cross-validation protocol in which Minimum Redundancy Maximum Relevance (mRMR) feature selection and hyper-parameter optimisation were performed exclusively inside each outer training fold. Gradient Boosting achieved the highest subject level accuracy of 88.33 ± 4.56, surpassing the best previously reported subject level accuracy of 83.33% on the same dataset. A hybrid CNN-BiLSTM-Attention network, applied to the same mRMR-selected features, achieved a competitive subject-level accuracy of 80.00±11.18%. These results demonstrate that a rigorously nested evaluation combined with multidomain HRV features yields reliable subject-level discrimination and offers both performance and interpretability.