Speech signal analysis as an aid to clinical diagnosis and assessment of mental health disorders

The study of the human voice is of great applicative interest. The acoustic analysis of the voice is characterized as a totally non-invasive instrument of support and comparison with other methods of clinical investigation. Thanks to technological and IT development, the study and implementation of parameters characterizing biomedical signals and voice images has become increasingly interesting for doctors, speech therapists and patients, as it can provide an aid in many applications: support for diagnosis, evaluation of the effectiveness of treatments and surgical interventions, monitoring of the patient during rehabilitation, including through the use of portable devices. In particular, in this thesis we focused on signal processing methods to aid clinicians in the diagnosis and monitoring of neurological speech disorders such as bipolar disorder (BD) and Attention Deficit Hyperactivity Disorder (ADHD). Starting from voice signals recorded during verbal fluency tests (VFT), speech features were extracted and investigated. VFT seem thus to be a sensitive means of measuring symptoms found in psychiatric disorders. In VFT, patients are instructed to produce words according to specified criterions, e.g. phonemic or semantic criteria, the continuous association of words following a cue word or simply free word generation in absence of a specified criterion. The present study aims at assessing the classification accuracy of BD patients and ADHD patients status via the analysis of vocal parameters obtained via different VFT conditions. For this, we adopted two different machine learning approaches. A first supervised machine learning method, using a model trained on previously clinically classified subjects and the speech derived features. For this method, a SVM-RFE (Support Vector Machine – Recursive Features Selection) has been used which selects the features set that gives the best classification of the subjects in terms of accuracy. The other approach was an unsupervised machine learning method, K-means clustering which divides the dataset into groups without knowing subjects labels a priori. The results of these two methods were then compared with each other. A greater accuracy was expected in the supervised approach being "guided" by the a priori knowledge of the subjects' labels