Machine learning classification of patients after suicide attempts using demographic data, EEG connectivity and heart rate variability

OBJECTIVE: The aim of this study was to develop a way to distinguish suicidal patients based on their electrophysiologic (EEG connectivity and heart rate variability) and demographic data. Various machine learning algorithms were compared to find the best models and features for this task. METHODS: Using a dataset of 140 subjects (or 87 subjects for HR models) from previous studies, different machine learning classifiers (KNN, SVM, Logistic Regression and Random Forest) were trained (using Monte Carlo cross validation and L2 regularization) and tested. Different data subsets were compared (HR, EEG, Demographic data) calculating the importance of individual features. RESULTS: 84 machine learning classifiers were trained. The models based only on EEG connectivity and heart rate variability (HRV) data achieved an accuracy of 71.1 % (SVM) and 68.3 % (KNN) and AUC-ROC of 75.4 % and 78.8 % respectively. Adding the demographic variables, the results further improved to Accuracy of 76.9 % and AUC-ROC of 81.9 %. EEG connectivity alone failed to provide sufficient discriminatory value, HRV data showed promising results in this regard. CONCLUSIONS: Pure physiological data, used in combination, might be sufficient in classifying suicide attempters with good prediction power. Adding sociodemographic data, the results improved, underlying the importance of these variables. SIGNIFICANCE: Machine Learning classifiers using combined sociocultural and physiological parameters might aid the identification of suicidal patients.