Vocal acoustic analysis and machine learning for the identification of schizophrenia

Abstract

Psychiatry still needs objective biomarkers. In the context of schizophrenia, there are speech abnormalities such as tangentiality, derailment, alogia, neologisms, poverty of speech, and aprosodia. There is a growing interest in speech signals features as possible indicators of schizophrenia. This article aims to develop an intelligent tool for detection of schizophrenia using vocal patterns and machine learning techniques. The main advantages of this type of solution are the low cost, high performance, and for being non-invasive. Thirty-one individuals over 18 years old were selected, 20 with previous diagnosis of schizophrenia, and 11 healthy controls. Their speech was audio-recorded in naturalistic settings, during a routine medical assessment for psychiatric patients. In the case of healthy patients, the recordings were made in different environments. Recordings were pre-processed, excluding non-participant voices. We extracted 33 features. We used the particle swarm optimization algorithm for feature selection. The classifiers’ performance was analyzed with four metrics: accuracy, sensibility, specificity, and kappa index. Best results were achieved when considering all 33 extracted features. Within machine models, support vector machines (SVM) models provided the greatest classification performance, with mean accuracy of 91.76% for PUK kernel. Our results outperform those from most studies published so far for the detection of schizophrenia based on acoustic patterns. The use of machine learning classifiers using vocal parameters, in particular SVM, has shown to be very promising for the detection of schizophrenia. Nevertheless, further experiments with a larger sample will be necessary to validate our findings.