Development and Validation of Mortality Prediction Models among Frail Participants in the UK Biobank Study

Identifying effective risk assessment strategies and prediction models for frail populations is crucial for precise mortality risk identification and improved patient management. This study aimed to evaluate whether prediction models incorporating survey data combined with biomarkers, physical measurements, or both could enhance mortality risk prediction in frail individuals than survey-only models.

15 754 frail participants aged 40-72 from the UK Biobank were included. We used Cox models to assess all-cause mortality risk and Light Gradient Boosting Machines for variable selection by sex. Performance was evaluated through discrimination, calibration, and reclassification.

In the survey-only models, we selected 24 predictors for males and 19 for females; age and number of treatments were the top predictors for both sexes. In the biomarker models, we selected 15 predictors for males and 24 for females. In the physical measurement models, we retained 24 predictors for males and 23 for females. The base models showed good discrimination: C-statistic was 0.73 (95% CI, 0.72-0.75) for males and 0.74 (95% CI, 0.72-0.76) for females in development, and 0.70 (95% CI, 0.65-0.75) for males and 0.78 (95% CI, 0.73-0.83) for females in validation. Although incorporating additional predictors led to some improvement in model performance, the overall enhancement was not substantial.

Survey-based models predicted mortality in frail individuals effectively, with only minor improvements from adding biomarkers or physical measurements. These findings highlighted the value of surveys in forecasting outcomes and informed personalized management strategies to improve health for the frail.