Classification of Physiological States Through Machine Learning Algorithms Applied to Ultra-Short-Term Heart Rate and Pulse Rate Variability Indices on a Single-Feature Basis

Abstract

This study investigates the feasibility of classifying physiological stress states usingMachine Learning (ML) algorithms on short-term (ST,∼5min) and ultra-short-term (UST, < 5 min, down to 10 heartbeats) heart rate (HRV) or pulse rate variability (PRV) features computed from inter-beat interval time series. Three widely employed ML algorithms were used, i.e. Naive Bayes Classifier, Support Vector Machines, and Neural Networks, on various time-, frequency and information domain HRV/PRV indices on a single-feature basis. Data were collected from healthy individuals during different physiological states including rest, postural and mental stress. Results highlighted comparable values using either HRV or PRV indices, and higher accuracy (>65% for most features and all classifiers) when classifying postural than mental stress. While decreasing the time series length, time-domain indices resulted still reliable down to ∼10 s, contrary to UST frequency-domain features which reported lower accuracy below 60 heartbeats.