Tree-based model offers best balance for mobility prediction
Across 14 methods, a tree-based model scored 0.673 for ranking risk and 0.147 for prediction error, with a training time of 9.2 minutes.
Limit: Prediction performance was measured only within cohorts, not in independent populations
Researchers conducted a benchmarking analysis of 14 prediction methods using data from 14,821 older adults in four long-term US aging studies. They defined mobility limitation as self-reported walking difficulty or walking speed below 0.8 meters per second. They repeatedly trained models on some participants and tested predictions on others within each cohort.
Several methods performed similarly at ranking risk. An oblique random survival forest, which combines many branching prediction rules, offered the most favorable overall balance. It combined strong risk ranking, good agreement between predictions and outcomes, and moderate training time. When training speed mattered most, regression methods that limit model complexity may be preferable. The findings supported both approaches as strong starting choices for predicting when events occur in aging research.
Why it matters
Mobility limitation presents a challenge for aging studies with widely spaced assessments. This benchmark addresses how researchers can estimate when it begins while balancing prediction quality with practical demands.
Caveats
Testing took place within each cohort, so the benchmark did not establish performance in independent populations. It evaluated predictions, not ways to prevent mobility limitation.
The paper
Wake Forest University School of Medicine; University of Maryland School of Medicine
Experimental Gerontology, Oct 18, 2026
