External validation and adaptation of the Computational Barthel Index Tool in a large nursing home cohort

The Computational Barthel Index Tool (CBIT), originally developed using U.S. Department of Veterans Affairs (VA) data, predicts future dependency in activities of daily living up to one year in advance. This study externally validated CBIT in a large non-VA nursing home population and evaluated whether population-specific adaptation could recover performance following external transport. We analyzed SEER-Medicare linked data, including 843,836 patients and over 8 million Minimum Data Set (MDS) assessments. The predictor and outcome structures of the original CBIT models were reconstructed in the SEER-Medicare cohort. Pretrained models were evaluated without modification for discrimination and calibration at current, 3-, 6-, and 12-month prediction timeframes. Population-specific adaptation was evaluated using progressively larger subsets of SEER-Medicare training data and compared with model reconstruction in the SEER-Medicare. Under strict external validation, discrimination declined relative to the original VA models, particularly for longer-term predictions. For evaluation models, mean AUC decreased from 0.795 in the original models to 0.693 at current prediction, 0.595 at 3 months, 0.595 at 6 months, and 0.592 at 12 months. Re-evaluation models showed stronger but similarly declining external performance. Calibration error increased, particularly for evaluation models at longer timeframes. Subgroup analyses showed modest variation in discrimination across demographic and clinical groups. Population-specific adaptation substantially recovered discrimination, with performance improving as more SEER-Medicare training data were incorporated. Using the adaptation training set, adapted models achieved performance comparable to SEER-Medicare model reconstruction. External validation revealed substantial performance degradation when CBIT was transported to a non-VA population. Population-specific adaptation recovered much of this performance loss, supporting a structured approach to clinical ML transportability in which pretrained models are externally validated first, followed by adaptation when performance degradation is observed. Full model reconstruction remains an alternative when sufficient local data are available.