Impact of Altering Data Granularity Levels on Predictive Modelling: A Case Study of Fall Risk Prediction in Older Persons

Classification systems are widely used in medicine for knowledge representation. The hierarchical relationships between concepts in a classification system can be exploited in prediction models by looking for the optimal predictive granularity level. In this study, we used the Anatomical Therapeutic Chemical (ATC) classification system to cluster medications in the context of predicting medication-related falls in older persons. We compared the performance of fall risk prediction by describing medications at varying granularity levels of the ATC classification system. We found that the level of abstraction significantly affects the predictive performance in terms of both discrimination (measured by the receiver operating characteristic curve AUC-ROC) and calibration. An implication of these findings to the researchers is that data representation at different granularity levels can influence the predictive performance. The optimal granularity level can be determined by experimentation.