RESUMEN
BACKGROUND: Speech features are essential components of psychiatric examinations, serving as important markers in the recognition and monitoring of mental illnesses. This study aims to develop a new clinical decision support system based on artificial intelligence, utilizing speech signals to distinguish between bipolar, depressive, anxiety and schizophrenia spectrum disorders. SUBJECTS AND METHODS: A total of 79 patients, who were admitted to the psychiatry clinic between 2020-2021, including 15 with schizophrenia spectrum disorders, 24 with anxiety disorders, 25 with depressive disorders, and 15 with bipolar affective disorder, alongside with 25 healthy individuals were included in the study. The speech signal dataset was created by recording participants' readings of two texts determined by the Russell emotion model. The number of speech samples was increased by using random sampling in speech signals. The sample audio signals were decomposed into time-frequency coefficients using Wavelet Packet Transform (WPT). Feature extraction was performed using each coefficient obtained from both Mel-Frequency Cepstral Coefficients (MFCC) and Gammatone Cepstral Coefficient (GTCC) methods. The disorder classification was carried out using k-Nearest Neighbor (kNN) and Support Vector Machine (SVM) classifiers. RESULTS: The success rate of the developed model in distinguishing the disorders was 96.943%. While the kNN model exhibited the highest performance in diagnosing bipolar disorder, it performed the least effectively in detecting depressive disorders. Whereas, the SVM model demonstrated close and high performance in detecting anxiety and psychosis, but its performance was low in identifying bipolar disorder. The findings support the utilization of speech analysis for distinguishing major psychiatric disorders. In this regard, the future development of artificial intelligence-based systems has the potential to enhance the psychiatric diagnosis process.