422. ORGANIZING THE DISORGANIZATION: AN ARTIFICIAL INTELLIGENCE-BASED DECISION SUPPORTING SYSTEM FOR THE DETECTION OF DISORGANIZATION SYMPTOMS IN SCHIZOPHRENIA AND NON-AFFECTED PARTICIPANTS

Abstract Background Disorganization symptoms are a core feature of schizophrenia, encompassing cognitive, affective, and behavioral manifestations such as formal thought and language disorders (FTD), motor disorganization, and inappropriate affect. Notably, FTD is considered the only symptom in schizophrenia influenced by genetic factors. Disorganization symptoms are strong predictors of non-response to antipsychotic treatment and are associated with abnormal brain activity and functional dysconnectivity. While disorganization is a hallmark of schizophrenia, it may also represent a dimensional phenotype with varying expression across the population. Its neurobiological basis likely involves disrupted synaptic and network connectivity. However, detecting disorganization remains challenging for clinicians due to its subtle and often non-pathological presentation. AI methods could assist by standardizing and automatizing diagnostic procedures, reducing clinician bias, and accelerating evaluation processes. Aims & Objectives The goal of this project is to develop an AI-based Decision Support System (DSS) to automate the assessment of disorganization in schizophrenia patients and, more broadly, in the general population, according to the view that thought disorganization is a dimensional phenotype with solid neurobiological underpinnings and varying degrees of expression. Method We are testing AI-based methods to detect objective measures of thought and behavioral disorganization, including speech analysis, the Thought and Language Disorder (TALD) scale, gait and motor movement analysis, facial expression analysis, and neural responses to emotionally salient stimuli. The study recruits at least 60 schizophrenia patients, their first-degree relatives, and case-matched non-affected participants. Clinical, psychopathological, and cognitive data are collected per standard clinical procedures. Data are integrated with information from various sources: These multimodal data are used to train a Machine Learning/Deep Learning model, creating a predictive algorithm to quantify disorganization. Data are pre-processed, structured, and inputted into classifiers designed for the disorganization dimension, with outputs provided to a meta-classifier to assess disorganization severity in the individual. Results Recruitment is ongoing, with approximately 30 schizophrenia patients and 30 non-affected participants tested. Interim analyses will be conducted in February and March, with recruitment concluding in April. Preliminary data on the full sample will be available by May. AI algorithms are being trained on public databases to evaluate speech, facial expression, and movement performance metrics. The DeepTALD system has been implemented to automate scoring of the TALD scale. Preliminary results show a strong positive correlation (ρ = 0.72) between DeepTALD and clinician scores of TALD items on synthetic speech. Discussion & Conclusions Early results suggest that the DeepTALD system is a reliable tool for automating the detection of language and thought disorders. Future developments will focus on quantifying thought disorganization intensity at baseline and under stimulus. Ultimately, the goal is to create applications that monitor thought disorganization by tracking speech, movements, and facial expressions in naturalistic settings.