Machine Learning based Anxiety Detection using Physiological Signals and Context Features

Anxiety is a common mental disorder that affects millions of people worldwide. Anxiety can be detected using physiological signals such as heart rate, skin conductance, blood pressure, and respiration. Machine learning is the ability to learn from data and forecast future events. Physiological signals can be analyzed and classified into different anxiety levels using machine learning. This paper aims to present the contribution of context features in detection of anxiety using physiological signals via machine learning. It is an imbalanced multiclassification problem using ECG and EDA signals with context features collected from 15 healthy individuals. These signals of sampling rate 700 Hz collected using RespiBAN from the Wearable Stress and Affect Detection (WESAD) dataset are used here with the 6-STAI questionnaire and the context collected about the subject during the experiment. Scores for each subject are calculated from this questionnaire which categorized subjects into different levels of anxiety. Multiple features are extracted from these signals and scaling is done on the extracted features. With the help of different data balancing techniques and different machine learning algorithms, this study will classify anxiety into three classes, namely, low, moderate and high. Classification of anxiety with fusion of multimodal physiological signals gave better results as compared to single modal physiological signals with an accuracy of 89.8% with ECG signal, 85.9% with EDA signal and 96.7% with ECG and EDA both using Gradient Tree Boosting algorithm with LOOCV after random oversampling. While, the best results are achieved when multimodal physiological signals are fused with context features, it gives the accuracy of 97.3% with Gradient Tree Boosting and LOOCV algorithm.