Presentation (CTI)

OBJECTIVES: The EQ-5D is widely used to measure health-related quality of life (HRQoL), especially in economic evaluations. When EQ-5D data are partially or entirely missing, index scores may be estimated using alternative data sources either directly by predicting index values or indirectly by predicting dimension responses and computing the index. This study compared statistical and machine learning (ML) approaches to identify the most accurate and efficient model for predicting EQ-5D-3L scores using a large, representative dataset from England. A secondary aim was to identify key predictors to inform research and policy.
METHODS: Data were drawn from seven waves of the Health Survey for England (2003-2006, 2008, 2011, and 2012) for individuals aged 16+. Six models were evaluated: two statistical methods (Ordinary Least Squares and logistic regression), two tree-based algorithms (XGBoost classification and regression), and two neural networks (NNs). Both direct (predicting index scores) and indirect (predicting dimension responses) approaches were tested. A holdout validation strategy trained models on 2003-2008 data, tuned on 2011, and tested on 2012. Performance was assessed using Mean Absolute Error (MAE) and Health Severity Group (HSG) accuracy. Feature importance was assessed using permutation analysis and SHAP.
RESULTS: NNs outperformed all models, with the best direct NN (one hidden layer) achieving the lowest MAE (0.088) and highest HSG accuracy (73.34%). XGBoost models were also competitive (MAE as low as 0.092; HSG accuracy up to 72.01%) and computationally efficient. Statistical models showed similar MAE but lower HSG accuracy, especially for pain and anxiety. Indirect prediction improved most models, though NNs performed well in both approaches. Key predictors: employment, chronic conditions, self-rated health, and recent illness highlighted the dominance of health-related over demographic factors.
CONCLUSIONS: NNs best captured complex EQ-5D-3L patterns and show promise when direct HRQoL data are unavailable. Findings highlight key health-related predictors and inform targeted interventions and policy.