Francesco Bartoli, Mario Luciano, Cristina Crocamo, Carlo Bassetti, Daniele Cavaleri, Andrea Aguglia, Anna-Rita Atti, Isabella Berardelli, Angela Chiara Cecere, Salvatore Cipolla, Pasquale De Fazio, Renato de Filippis, Daniele De Francesco, Carla Gramaglia, Valeria Latorre, Giulia Lucca, Massimo Pasquini, Nicola Poloni, Maurizio Pompili, Alessandro Rodolico, Gianluca Rosso, Gianluca Serafini, Maria Salvina Signorelli, Lorenzo Tarsitani, Antonio Ventriglio, Umberto Volpe, Patrizia Zeppegno, DEEP READ Study Group, Andrea Amerio, Flavia Boccardi, Carlo Bommartini, Vincenzo Borrelli, Elisa Briasco, Ruggero Cavallaro, Martina Citton, Giorgio Cucchi, Giacomo De Curtis, Stefano De Santo, Beatrice Defilippi, Denise Erbuto, Luigi Gadaleta, Eleonora Gambaro, Gaspare Liparota, Giulio Longo, Annalisa Maraone, Federica Marcolini, Monica Migliorati, Laura Orsolini, Silvia Paparesta, Francesco Piarulli, Maria Luigina Pignataro, Cecilia Quitadamo, Alessia Scordo, Giuseppe Carrà, Andrea Fiorillo
A ML model based on routinely collected data showed moderate discrimination for 90-day readmission in MDD. These findings suggest that standard clinical variables may hold predictive value for early readmission. However, external validation, assessment of model calibration, and clinical utility are needed prior to clinical implementation.
BACKGROUND: Major depressive disorder (MDD) is associated with high relapse rates, with around 20% of inpatients experiencing readmission within 90 days after discharge. Routinely collected clinical information may help identify individuals at higher risk of readmission. Machine learning (ML) models could complement more traditional statistical approaches by exploring complex relationships among multiple vulnerability factors.
METHODS: The DEEP READ study is a prospective, multicentre cohort study conducted across 13 Italian provinces, enrolling adults aged 18-65 years with a DSM-5-TR MDD diagnosis between January 2024 and December 2025. The primary outcome was unplanned psychiatric readmission within 90 days. Sociodemographic and clinical features (including Hamilton Depression Rating Scale scores and DSM-5 specifiers), comorbidities, and pharmacological treatment at discharge were recorded. A Random Forest classifier was trained on 22 predictors and internally evaluated using stratified 5-fold cross-validation.
RESULTS: The sample included 322 individuals (mean age 43.0 years; 38.5% men), of whom 15.5% (n = 50) were readmitted within 90 days. The model achieved a mean test AUC of 0.74 (range: 0.59-0.87). Suicide attempts, prior hospitalisations, anticonvulsant prescription, cluster B personality disorder, female sex, and anxious distress ranked among the leading predictors. Although exploratory, lithium prescription showed a negative SHapley Additive exPlanations (SHAP) pattern.
CONCLUSIONS: A ML model based on routinely collected data showed moderate discrimination for 90-day readmission in MDD. These findings suggest that standard clinical variables may hold predictive value for early readmission. However, external validation, assessment of model calibration, and clinical utility are needed prior to clinical implementation.