<![CDATA[Schizophrenia symptom severity estimation provides quantitative information that is useful at both the detection and treatment stages of the mental disorder, as the information helps in decision-making and improves the management of the illness. Very limited studies have been recorded for estimating the symptom severity as a regression task with machine learning, especially from speech recordings, which is the aim of this study coupled with detection. Acoustic features, which comprise frequency-domain and time-domain features, were extracted from 60 schizophrenia subjects and 59 healthy controls enrolled in this research. The acoustic features were used to train GridSearchCV-optimized XGBoost as a classifier. Three Multi-Layer Perceptron (MLP) networks, hyper-parameter-tuned by Bayesian Optimizer, were trained to predict the sub-type symptom severity from acoustic extracted features from the schizophrenia groups. The XGBoost classification model that discriminates between schizophrenia and healthy groups achieved a classification accuracy of 98.6%. The three MLP regression models yielded Mean Absolute Errors of 1.975, 2.856, and 1.555, as well as correlation coefficients of 0.888, 0.806, and 0.786 for predicting positive, negative, and cognitive symptom scores, respectively. Solution architecture for the deployment of the models for practical use was suggested]]>
Copyrights © 2025
