Bipolar disorder (BD) is frequently misdiagnosed as major depressive disorder (MDD), with diagnostic delays averaging 6–10 years leading to suboptimal treatment and poorer outcomes. Early identification of patients at risk for diagnostic conversion remains a critical clinical priority. We aimed to identify clinical predictors of BD conversion in a large population-based cohort and develop a clinically applicable prediction model.

We conducted a retrospective cohort study using electronic health records from the Catalan PADRIS-PRESTO database (2010–2019). Patients aged ≥18 years with incident MDD diagnosed in specialized mental health services (2015–2019) were followed until BD diagnosis, death, or study end. Time-varying Cox regression addressed proportional hazards violations, with competing risks analysis accounting for mortality.

Among 52,001 patients (median follow-up 4.0 years), 3,958 (7.61%) converted to BD. Psychotic features emerged as the strongest predictor (HR = 3.18, 95% CI: 2.92–3.45, E-value = 5.82), with consistent effects across follow-up. Anxiety disorders showed time-varying protection (baseline HR = 0.41, 95% CI: 0.35–0.48) that attenuated over time (interaction HR = 1.30 per period). The time-varying model achieved good discrimination (C-index = 0.745). Machine learning validation confirmed predictor importance rankings (Spearman r = 0.87).

Psychotic features during depressive episodes robustly predict BD conversion, warranting enhanced clinical monitoring. Our time-varying prediction model using routinely collected clinical data enables early risk stratification to guide diagnostic vigilance and treatment decisions.