1. Background
Employment outcomes for individuals with severe mental illness remain poor, with meta-analysis showing only 32.5% of individuals with first-episode psychosis maintain long-term employment (Ajnakina et al., 2021). Despite employment’s importance for recovery, a significant methodological gap exists in how we predict these trajectories. While machine learning (ML) models show promise in other healthcare domains, whether they meaningfully outperform classical survival models like Cox proportional hazards for employment outcomes remains an open question (Spreafico, et al. 2024). This project will benchmark whether, and when, ML approaches offer meaningful improvements.
2. Novelty & Importance
This is the first systematic comparison of ML versus classical survival methods for predicting employment trajectories in severe mental illness. It leverages two world-class administrative datasets: the Danish DREAM database with weekly employment tracking for 5.16 million residents, and the UK CRIS-DWP linkage at SLaM connecting electronic mental health records with benefits data. The novelty lies in rigorous benchmarking of state-of-the-art ML survival approaches against classical methods to establish their relative advantages, combined with cross-national validation and explicit competing risk modelling. Improved prediction models could enable earlier identification of individuals at risk of employment loss and inform targeted, personalised interventions.
3. Aims & Objectives
The primary aim is to systematically compare machine learning and classical survival methods for predicting employment and benefits outcomes, initially focusing on psychosis spectrum disorders. Site-specific models will evaluate ML survival approaches (Random Survival Forests, Gradient-Boosted Survival, DeepSurv, Dynamic-DeepHit, SurvTRACE) against Cox proportional hazards and Royston-Parmar flexible parametric models. Primary outcomes include time to first sustained employment and time off benefits, with competing risks explicitly modelled. The project will culminate in cross-national validation and development of a prototype decision support tool for clinical implementation.