Detection of Major Depressive Disorder, Bipolar Disorder, Schizophrenia and Generalized Anxiety Disorder Using Vocal Acoustic Analysis and Machine Learning

Abstract

Purpose Current diagnosis and treatment in psychiatry are heavily depends patients’ reports and clinician judgement, thus are sensitive to memory and subjectivity biases. On the other hand, a growing effort in developing objective markers in psychiatry may be a way to face all these challenges. Vocal acoustic features have been recently studied as objective measures, with the advantages of being accessible, inexpensive, non-invasive and remotely performed. Method The main objective of this work is to propose a methodology to support the diagnosis of major depressive disorder, bipolar disorder, schizophrenia, and generalized anxiety disorder using vocal acoustic analysis and machine learning. Seventy-eight individuals over 18 years old were recruited into five groups: 28 participants with major depressive disorder; 20 patients with schizophrenia; 15 patients with bipolar disorder; 4 patients with generalized anxiety disorder; and 12 healthy controls. ResultsRecordings were submitted to a vocal feature extraction algorithm, and to experiments using different machine learning classification techniques. Random Forests with 300 trees achieved the greatest classification performance (75.27% for accuracy, 69.08% for kappa, 75.30% for sensitivity, and 93.80% for specificity) for the simultaneous detection of major depressive disorder, schizophrenia, bipolar disorder and generalized anxiety disorder.ConclusionAs changes in vocal patterns have been reported in several mental disorders and appear to correlate with illness severity, vocal acoustic features have shown to be promising markers, with the advantages of being abundant, inexpensive, non-invasive and remotely performed. The results provided by our proposed solution support the feasibility of a computational method based on vocal parameters for assisting clinicians in patient triage and diagnosis in psychiatry.As the main contribution of this work, we present the intelligent system composed of the support vector machine and the time and frequency characteristics of the audio, which works as a digital biomarker to support the diagnosis of mental disorders in the context of psychiatric emergency.