F. Coutts, S. Mena, Alkomiet Hasan, I. Winter-van Rossum, W. W. Fleischhacker, R. Kahn, M. Davidson, I. Sommer, N. Sirbu, G. Jacobs, L. Bryant, L. Moles, N. Koutsouleris, P. A. Lalousis
- Antipsychotic medication is the primary treatment for psychotic disorders, yet nearly half of individuals experiencing a first episode of psychosis discontinue treatment within one year against clinical guidance [1]. Discontinuation may result from poor insight, substance use, negative attitudes toward medication, side effects, cognitive difficulties, or the belief that treatment is no longer needed [2, 3]. Understanding which patients are most at risk of stopping medication, and why, is essential for informing clinical decisions and developing targeted interventions. However, most clinical trials rely on "all-cause discontinuation" (ACD) as a primary endpoint, which aggregates diverse reasons for stopping treatment or leaving the trial and lacks explanatory power. To address this gap, we applied an artificial intelligence (AI)-based clustering approach to identify subgroups of patients with shared characteristics in terms of symptom severity, side effect burden, and medicationAntipsychotic medication is the primary treatment for psychotic disorders, yet nearly half of individuals experiencing a first episode of psychosis discontinue treatment within one year against clinical guidance [1]. Discontinuation may result from poor insight, substance use, negative attitudes toward medication, side effects, cognitive difficulties, or the belief that treatment is no longer needed [2, 3]. Understanding which patients are most at risk of stopping medication, and why, is essential for informing clinical decisions and developing targeted interventions. However, most clinical trials rely on "all-cause discontinuation" (ACD) as a primary endpoint, which aggregates diverse reasons for stopping treatment or leaving the trial and lacks explanatory power. To address this gap, we applied an artificial intelligence (AI)-based clustering approach to identify subgroups of patients with shared characteristics in terms of symptom severity, side effect burden, and medication attitudes at the point of discontinuation. We then trained predictive models using baseline clinical and sociodemographic data to explore whether these subgroups could be identified earlier in care.
Data were drawn from 280 individuals with schizophrenia enrolled in the European Long-Acting Antipsychotics in Schizophrenia Trial (EULAST; NCT02146547) trial [4]. For patients labelled as discontinuing their medication or the trial (ACD), data from the nearest visit within two months of discontinuation were used. Measures included the Clinical Global Impressions (CGI) scale, Positive and Negative Syndrome Scale (PANSS), Medication Adherence Report Scale (MARS), Systematic Monitoring of Adverse events Related to Treatment System (SMARTS), Subjective Wellbeing under Neuroleptic Treatment Scale (SWN), and the digit symbol task from the Wechsler Adult Intelligence Scale (WAIS). We used orthogonal non-negative matrix factorization to identify patient clusters and applied multiclass Support Vector Machine models to predict cluster membership compared to no discontinuation based on baseline clinical and sociodemographic data. We also modelled two standard trial outcomes, ACD and symptomatic remission criteria [5], for comparison.
Of the total sample, 44.2% (n=124) discontinued their medication. Clustering revealed two distinct groups: a "Less Impaired" cluster (n=86), defined by more positive views of medication and better functioning, and a "More Impaired" cluster (n=38), characterized by greater illness severity, more side effects, and more negative attitudes toward medication (p < 0.0001). Baseline predictive models distinguished the More Impaired cluster from the Less Impaired (AUC = 0.64) and from the Non-Discontinuation group (AUC = 0.65). The Less Impaired cluster versus Non-Discontinuation comparison was less accurate (AUC = 0.60). All cluster-based predictions outperformed the ACD prediction (AUC = 0.59), but none exceeded the performance of remission prediction (AUC = 0.73).
This study identified two clinically meaningful subgroups of patients who discontinued treatment, with the More Impaired cluster showing worse symptoms and side effects and being more reliably predicted from baseline characteristics. These findings highlight the potential of AI-driven approaches to move beyond traditional trial endpoints and identify individuals at risk for discontinuation, opening the door to proactive, targeted interventions in early psychosis care.…
