Predicting clinical outcomes in Employee Assistance Programmes: A machine learning analysis of 30,000 clients

The increasing availability of routine mental health data presents substantial opportunities for improving health service planning, yet analysis is often challenged by noise, data access, data wrangling/cleaning, heterogeneity, and real-world complexity. Machine learning (ML) approaches offer a means of identifying clinically meaningful patterns that support mental health service optimisation and outcome prediction. This study develops and evaluates ML models to predict clinical outcomes and treatment intensity in a large real-world Employee Assistance Programme (EAP) dataset. The analytical sample comprised 30,821 clients who accessed counselling services between 2012 and December 2024. Clinical outcomes were operationalised using CORE10 (a validated measure of psychological distress). Logistic Regression, Random Forest, and XGBoost were applied to predict two binary clinical outcomes (reliable change and recovery) and session count, using eight predictors/features including baseline psychological distress, demographic characteristics, occupational sector, primary presenting issue, and temporal contextual factors. Model development and validation are currently ongoing, and comparative performance, feature importance, and predictive accuracy for clinical outcomes and session count will be presented at the conference. Findings will be discussed in terms of their implications for early risk detection, demand forecasting, and resource planning in mental health services.