Characterising symptom trajectories using electronic health records and natural language processing

Understanding the prodrome of severe mental disorders (SMD) (including unipolar [UMD], bipolar [BMD], and psychotic disorders [PSY]) requires modelling how symptoms and substance use evolve and interact over time. Temporal network analysis enables identification of directional relationships between prodromal features by representing them as nodes and their temporal associations as edges. This study aimed to construct a global transdiagnostic network of prodromal features and compare subgroup-specific networks for UMD, BMD and PSY.

Electronic health records from South London and Maudsley NHS Foundation Trust were analysed for 6,462 individuals with SMD diagnoses (UMD: 2,066; BMD: 740; PSY: 3,656). Validated natural language processing algorithms identified 61 prodromal features at three-month intervals from two years to six months prior to onset. Temporal networks were estimated using generalised vector autoregression panel models, adjusting for covariates. Edge weights (partial directed correlations, z) captured autocorrelative, unidirectional, and bidirectional relationships. Centrality was assessed via outgoing (cout) and incoming (cin) connections.

The global SMD network showed consistent autocorrelations (0.04 ≤ z ≤ 0.10) and predominantly positive associations. Aggression (cout = 0.103) and tearfulness (cin = 0.134) emerged as the most central features. Sub-network comparisons revealed minimal differences, with 3.5% of edges differing between UMD and PSY, 0.8% between UMD and BMD, and 0.4% between BMD and PSY. Community detection identified a psychotic cluster (delusional thinking, hallucinations, paranoia) and two behavioural clusters involving aggression, hostility, agitation, and substance use.

These findings support transdiagnostic early detection approaches, highlight shared prodromal structures across SMD and identify central features as potential intervention targets.