Cross-Generational Transportability of Cognitive Risk Models: A SHAP-Based Analysis from Aging Adults to University Students

This study evaluates the portability of cognitive risk models across generations by training a LightGBM classifier on 65,914 aging adults from the LASI dataset (age 45+) and testing the trained algorithm on 297 university students (age 18-30). While the initial direct transfer only achieved an AUC of 0.586 due to distribution shifts, applying Platt scaling calibration and threshold optimization improved performance, achieving an AUC of 0.672 and an F1-score of 0.529, which is better than the baseline model. Moreover, the analysis revealed that the calibration of probability bridges the performance gap between different generations, while the domain shift from objective cognitive testing in the elderly to subjective cognitive complaints in students accounts for most of the performance drop.