TOPLINE

In a large retrospective study, advanced machine learning (ML) models identified demographic, socioeconomic, and clinical predictors of the future onset of epilepsy in patients with depression (PWD) and the future onset of depression in patients with epilepsy (PWE).

METHODOLOGY

• Researchers conducted a retrospective observational cohort study analysing longitudinal patient-level data from 18 data sources across seven European countries (Denmark, France, Germany, Italy, Spain, Sweden, and the UK).
• Supervised ML models were trained separately within each country using demographics, socioeconomic status indicators, clinical history, Charlson Comorbidity Index (CCI) score, and healthcare utilisation features.
• The analysis evaluated two distinct cohorts comprising approximately 2.2 million PWE and 9.7 million PWD.
• Shapley Additive Explanations were implemented to identify and rank the most important predictive features of the future onset of disease within a 365-day prediction window.
• Model performance was assessed using the area under the receiver operating characteristic curve (AUROC) and relative area under the precision-recall curve (rAUPRC).

TAKEAWAY

• The models achieved their highest performance using data from the UK for predicting depression in PWE (AUROC, 80%; rAUPRC, 6) and for predicting epilepsy in PWD (AUROC, 78%; rAUPRC, 15), followed by Denmark and Sweden.
• Predictors of the onset of depression in PWE were female sex; adult age (above 20 years); low socioeconomic status; a history of alcohol use; and prescriptions for anxiolytics, antipsychotics, or antimigraine agents.
• Predictors of the onset of epilepsy in PWD were male sex, low socioeconomic status, alcohol abuse, a higher CCI score, and use of antithrombotic agents.
• Increased healthcare resource utilisation, including a higher number of medical visits and prescriptions, consistently predicted future comorbidity onset across both patient cohorts.

IN PRACTICE

"Predictors common to both depression in PWE and epilepsy in PWD — including gender, low SES [socioeconomic status], high psychiatric multimorbidity burden and extensive healthcare use — highlight the need for a multidisciplinary care model that addresses these overlapping risk factors," the authors wrote.

"These [ML] tools could help clinicians identify high-risk individuals and enable timely interventions aimed at enhanced prevention, improving prognosis and enhancing quality of life of patients," they added.

SOURCE

The study was led by Alessandro Ruggieri, Angelini Pharma SpA, Rome, Italy. It was published online on June 19, 2026, in BMJ Neurology Open.

LIMITATIONS

The study was limited by its retrospective and observational design, heterogeneous electronic medical record data coverage across countries, variations in prescription data sources, the non-uniform capture of variables such as socioeconomic status, and small sample sizes in certain nations. It was further limited by the reliance on treatment proxies for multi-indication drug classes to establish diagnosis dates, incomplete patient histories from database panel coverage, the lack of data linkage between clinical and pharmacy records, and discordant international disease coding systems.

DISCLOSURES

The study was funded by Angelini Pharma SpA. Multiple authors disclosed being employees of Angelini Pharma SpA or IQVIA, with no other competing interests reported.

This article was created using several editorial tools, including AI, as part of the process. Human editors reviewed this content before publication.