Factor-specific decomposition of mortality predictions produced individualized profiles that varied by population and prediction horizon.

Objective

To examine whether aggregate all-cause mortality predictions from additive Cox models can be represented as individualized profiles of four prespecified factor contributions, and to assess variation by population, follow-up interval, and acute-severity adjustment. STUDY

Design

AND

Setting

Retrospective analyses used NHANES 1999-2018 linked mortality data and MIMIC-IV v2.2. Separate Cox models included diabetes, hypertension, current smoking, obesity, age, and sex. Fixed-reference Shapley values decomposed the four-factor portion of the fitted log relative hazard; dominant meant the largest strictly positive contribution. MIMIC chronic factors incorporated index-hospitalization discharge diagnoses, and OASIS used data through 24 hours after ICU admission; these analyses retrospectively explained fitted models rather than admission-time prediction. Only C underwent bootstrap optimism correction; calibration and Brier metrics were apparent.

Results

NHANES included 48,390 adults and 7,265 deaths; optimism-corrected C was 0.8525. Smoking had the largest mean contribution (0.1562), whereas hypertension was the most frequent positive dominant factor (25.12%); 29.50% were none-positive. MIMIC-IV included 50,917 first adult ICU stays and 9,010 deaths by 90 days; corrected C was 0.6556. Diabetes was dominant for 27.67%, and 72.33% were none-positive. OASIS increased corrected C to 0.7384; no chronic factor contributed positively in the pooled model. In an interval-specific sensitivity among patients still at risk after day 30, diabetes was positively associated during days 31-90. Fixed-reference Shapley values equalled conventional decomposition of the additive Cox linear predictor.

Conclusion

Individual factor profiles described the composition of existing mortality predictions and were conditional on cohort, follow-up interval, and adjustment. They did not identify causal effects or treatment priorities.