Neural Network Analysis of the Contribution of Psychotropic Prescription Sequences to the Risk of non-psychiatric Adverse Events in Bipolar and Schizophrenia Spectrum Disorders

Nathan Vidal^1,2*MOHAMMED SEDKI^1,3Nadia Younès^1,2Hugo Bottemanne^1,4,5Paul Roux^1,2,6Eric Brunet-Gouet^1,2

• ¹Centre de Recherche en Epidemiologie et Sante des Populations, Villejuif, France
• ²Centre Hospitalier de Versailles, Service Universitaire de Psychiatrie d’Adultes et d’Addictologie, Le Chesnay, France
• ³Institut Gustave Roussy Departement de Medecine Oncologique, Villejuif, France
• ⁴Hopital Bicetre, Le KremlinBictre, France
• ⁵Assistance Publique-Hôpitaux de Paris (AP-HP), Paris, France
• ⁶Service de psychiatrie du secteur 78G18, Centre Hospitalier de Plaisir, Plaisir, France

Psychotropic medications are associated with lower mortality in bipolar disorders (BD) and schizophrenia spectrum disorders (SZD) but may trigger serious adverse events requiring hospitalization. Determining the iatrogenic causes of such events can considerably help psychiatrists understand their development and adjust the prescription accordingly. We aimed to assess to what extent the psychotropic prescription sequence contributes to in-hospital non-psychiatric adverse events in BD and SZD. We conducted a case-control design including adults with BD or SZD from the French national healthcare system claims database (n = 87,182). A recurrent neural network model was trained to discriminate between adults who experienced adverse events and matched adults who did not, based only on psychotropic prescription sequences over the past 18 months and demographic data. Explainable AI combined enabled us to understand the model’s prediction. Psychotropic doses during the months preceding the adverse events were relatively more important than earlier doses to predict in-hospital urinary retention and thyroid disorders, but it was not the case to predict movement or cardiac disorders. The doses of certain benzodiazepines, tropatepine, quetiapine, clozapine, loxapine, lithium salts, and valproate were significant risk factors for adverse events. A recurrent neural network combined with explainable AI identified key psychotropic prescription features and duration associated with non-psychiatric adverse events among a large number of features. Yet, it was unable to predict events with high accuracy. Such a model could only be used retrospectively to generate hypotheses about iatrogenic risk factors for adverse events, offering limited value for integration into prescription softwares.