Comparative predictive performance of anticholinergic burden scales for anticholinergic-related hospitalisation and emergency department visits in older adults: a population-based study of model development and temporal validation.

introductionAnticholinergic burden scales are widely used to guide medication review in older adults, yet their ability to predict clinically important adverse outcomes remains uncertain, as most evidence is based on associations rather than validated risk prediction. In the absence of a gold standard, the clinical value of these scales depends on whether they can discriminate between individuals at risk and provide accurate absolute risk estimates.

aimTo compare the predictive performance of six anticholinergic burden scales for hospitalisation/emergency department (ED) visits related to anticholinergic adverse effects, assess internal and temporal validation of cognitive impairment models across scales, and examine the risk gradient across increasing anticholinergic burden levels for cognitive impairment across scales.

methodA retrospective cohort study using linked population-level administrative health data of adults aged ≥ 65 years in Western Australia (2015-2019). Development cohorts were from 2015-16 (N = 323,682) and 2016-17 (N = 334,304), and the temporal validation cohort was from 2017-18 (N = 330,684). Six anticholinergic burden scales were calculated annually. Logistic regression models with common predictor structure including demographic and clinical predictors were used to predict any hospitalisation/ED visit related to anticholinergic adverse events (falls/fractures/dizziness, cognitive impairment, or constipation/urinary retention). Model performance was assessed using the c-statistic, calibration slope, and Brier score. Cognitive impairment models were further evaluated using bootstrap internal validation (200 replications), temporal validation, and predicted risk estimation across exposure deciles.

resultsAcross initial model comparisons, Korean Anticholinergic Burden Scale (KABS) showed the highest c-statistics for each outcome, although between-scale differences were small. Cognitive impairment showed the highest discrimination across scales and was selected for further validation. In temporal validation, c-statistics for cognitive impairment ranged from 0.795 to 0.806, with the highest value observed for KABS, and calibration slopes ranging from 1.024 to 1.032 across scales. Predicted risk of cognitive impairment increased across exposure deciles for all scales, with the highest-decile risk ranging from 5.27% for TSDD-SAMS to 7.58% for KABS.

conclusionKABS showed slightly higher and more consistent predictive performance than the other scales, particularly for cognitive impairment, although between-scale differences were modest. Presenting validated absolute risk estimates across exposure groups may improve clinical interpretability and support risk stratification, but wider validation and clinical judgement remain essential before routine use.