DEEPTALD: A SYSTEM FOR SUPPORTING SCHIZOPHRENIA-RELATED FORMAL THOUGHT DISORDERS DETECTION WITH ARTIFICIAL INTELLIGENCE THROUGH NLP MODELS

Abstract Background Schizophrenia (SCZ) is a severe mental disorder characterized by positive, negative1, and cognitive symptoms2. The most severe forms of this disorder include relevant formal thought disturbances, which are collectively referred to as the disorganization domain. Notably, a recent study demonstrated that the disorganization domain is the most predictive domain to diagnose the treatment resistant subtype of SCZ3. Formal thought disorders can be defined as a fragmentation of the logical thought process and can be reconstructed by skilled clinicians through different evaluation tools, such as the Thought and Language Disorder (TALD) scale4. However, this process is time-consuming and scoring these scales may be biased and poorly objective as it depends on clinician’ s opinion and skills. Artificial Intelligence (AI) methods and techniques may be helpful both for the standardization of procedures, eliminating the subjective operator-dependent element, and for speeding up evaluation processes. Aims and Objectives To support the clinician in language and thought disorder detection by developing an easy-to-use system providing objective measures and scores related to the TALD scale items, together with a detailed explanation of the results. Clinicians may modify the parameters of the metrics with a Human-in-the-loop approach. Method We used three simulated clinical interviews for this study. The interviews were structured considering TALD item definitions and clinicians’ experience. The transcriptions were uploaded to the DeepTald System. Four TALD items were selected: perseverance, rumination, logorrhea, and rumination. We adopted Natural Language Processing models for determining data similarity useful for detecting perseverance, sentiment analysis for rumination of negative thoughts, and temporal speech parameters for accelerated (logorrhea) or slowed thinking disorders. Objective metrics for quantifying the disturbs are defined to generate the TALD scores and detailed reports. The DeepTald system has been assessed in an exploratory study by three clinicians. Results There exists a strong positive correlation, ρ = 0.72, between the DeepTald and the clinician scores of the DeepTald items on height synthetic speech. First results were encouraging: participants' perception of the tool reliability, technical competence, and understandability related to their trust in the system were positive, as well as their satisfaction. The report features, agreed with the proposed metrics, were considered useful by participants. Discussion and Conclusions Our preliminary results suggest that the DeepTald System is a reliable tool, which could be advantageous for supporting the detection of language and formal thought disorders with an objective measure. Our future purpose is to extend the use of the DeepTald System to other TALD items, using authentic recorded clinical interviews between clinicians and patients.