Relapse risk prediction in patients with first-episode bipolar disorder: development, external validation, and pharmacotherapy associations of a machine learning model

There are no established prognostic tools for predicting relapse risk in first-episode bipolar disorder (FEBD), limiting the use of personalized treatment approaches. We aimed to develop and validate a machine learning (ML) model to predict relapse in FEBD patients and assess whether pharmacotherapy effectiveness varies across predicted risk strata. We used nationwide registry data from Sweden (n = 30,402; follow-up = 2006-2021) and Finland (n = 13,790; follow-up = 1996-2018). The developed ML model achieved an area under the receiver operating characteristic curve (AUROC) of 0.71 (95% CI = 0.69-0.72, highest for relapse due to psychotic mania [AUROC = 0.85, 95% CI = 0.80-0.89]) in the Swedish cohort (internal validation) and 0.68 (95% CI = 0.66-0.69, highest for nonpsychotic mania [AUROC = 0.74, 95% CI = 0.69-0.78]) in the Finnish cohort (external validation). The model incorporated only seven accessible predictors, including pharmacotherapies during the first 30 days post-FEBD (lithium, combination treatments, and antipsychotics), outpatient follow-up within 30 days post-FEBD, prolonged initial hospitalization for FEBD, history of prior psychiatric hospitalizations, and sickness absence days. Among high-risk patients (N = 10,119, 31.77%), long-acting injectable (LAI) antipsychotics were associated with the greatest reduction in psychiatric rehospitalization risk (HR = 0.44, 95% CI 0.29-0.67). In the low-risk group (n = 21,736, 68.23%), only combinations of quetiapine with valproate (HR = 0.79, 95% CI = 0.65-0.95) or lamotrigine (HR = 0.86, 95% CI = 0.74-0.99) were associated with reduced rehospitalization risk. Available online for research purposes, our internally and externally validated prognostic model may inform clinical decision-making by identifying high-risk individuals who could benefit from early initiation of LAI antipsychotics, promoting a shift toward proactive, risk-guided treatment in FEBD patients.