Mental health risk stratification using heart rate variability and prefrontal electroencephalography: eXtreme gradient boosting-based classifier and latent class analysis-based subtyping.

Heart rate variability (HRV) and electroencephalography (EEG) are reflective of cardiac autonomic modulation/emotional regulation and cortical brain activity/psychopathology, respectively. This study aimed to determine the key HRV/prefrontal EEG classifiers of healthy controls (HC) versus patients with psychiatric disorders (PT) and to identify the subtypes of these features. A simultaneous 5-min-long acquisition of earlobe photoplethysmography and resting-state prefrontal EEG signals was performed in 147 HC and 123 PT [major depressive disorder (MDD; N = 107) and panic disorder (N = 16)]. A mixed graphical model (MGM) was constructed using the 9 HRV-related features, 11 prefrontal EEG-related features, depressive symptoms and anxiety, and psychiatric diagnosis by licensed psychiatrists. Top 12%-ranked nodes in regards of the betweenness centrality or expected influence were defined as hubs to be employed in the eXtreme Gradient Boosting (XGBoost) classifier model and the latent profile analysis (LPA): 2 HRV-related features [total power (TP), absolute power in low frequency band (LF)], 4 prefrontal EEG-related features [relative powers in frequencies of theta (THETA), low-alpha (L_ALPHA), low-beta (L_BETA), middle-beta (M_BETA)], and anxiety. The XGBoost model had an accuracy of 74.9 ± 5.6%. The SHapley Additive explanation values demonstrated that lower anxiety and M_BETA were indicative of a higher probability of being a HC. A lower LF indicated a greater possibility of being a PT. The LPA identified 3 PT subgroups of (1) hyperarousal [severe depression and moderate anxiety; higher M_BETA and lowered HRV], (2) relatively stable [minimal depressive symptoms and anxiety; no marked frontal EEG deviation and lowered HRV], and (3) frontal EEG slowing [moderate depression and mild anxiety; higher THETA and markedly lowered HRV in TP and LF]. Future studies to determine the results of risk indicators for patients with psychiatric disorders using HRV and prefrontal EEG features derived by the combined use of network-based feature selection, machine learning-based classifiers, and latent class subtyping in diverse populations are warranted. TRIAL REGISTRATION: N/A.