A novel ECG-based approach for classifying psychiatric disorders: Leveraging wavelet scattering networks

Abstract

Individuals with neuropsychiatric disorders experience both physical and mental difficulties, hindering their ability to live healthy lives and participate in daily activities. It is challenging to diagnose these disorders due to a lack of reliable diagnostic tests and the complex symptoms and treatments for various disorders. Generally, psychiatric disorders are identified manually by doctors using questionnaires, which may be prone to subjectivity and human errors. A few automated systems have recently been developed to identify these disorders using physiological signals, including electroencephalogram (EEG) and electrocardiogram (ECG) signals. Often, EEG signals are used to identify psychiatric disorders, but the EEG signals are nonlinear and non-stationary in nature and hence are relatively complex to analyze when compared to the ECG signals. The ECG signals in psychiatric patients are used due to the connection between the heart and brain. The proposed study is aimed at investigating the use of ECG signals for the automated identification of neuropsychiatric disorders, including bipolar disorder (BD), depression (DP), and schizophrenia (SZ). Generally, convolution neural networks (CNNs) have proven to be effective in accurately identifying psychological conditions. However, their application requires a large amount of data and technical expertise. The wavelet scattering network (WSN), a variant of CNNs, was introduced to overcome these limitations. The WSN is a network capable of accurately detecting unique patterns in the signal. The proposed research incorporated the WSN network and was conducted using a Psychiatric ECG Beat Dataset with a population of 233 subjects, of whom 198 were diagnosed with multiple psychiatric disorders, and 35 were control subjects. ECG signals from 3570 heartbeats were collected and analyzed using wavelet scattering-based feature extraction and machine learning techniques. The Fine K-Nearest Neighbor (FKNN) algorithm produced the best results with an average classification accuracy of 99.8% and a Kappa value of 0.996 using a ten-fold cross-validation. The model yielded an accuracy of 99.78%, 99.94%, 99.98%, and 100% for automated identification of BD, DP, SZ, and control subjects, respectively, with F1 scores and precision values close to 1. The proposed method could also help in the automated clinical detection of different psychiatric disorders.