Deriving the anticholinergic medication index for delirium and falls in older people: development and external validation of a risk prediction model

Objective: To develop the anticholinergic medication index (ACMI), a new anticholinergic burden measure derived from a clinical prediction model for falls and delirium, scoring 88 anticholinergic medicines prescribed in primary care.

Design Development and external validation of a risk prediction model.

Setting UK primary and secondary care data from Connected Bradford and the Secure Anonymised Information Linkage (SAIL) databank, 1 January 2019 to 31 December 2019.

Participants: For development cohort, 71 757 Bradford and Craven residents aged >65 years on 1 January 2019 who contributed data to Connected Bradford between 1 January 2019 and 31 December 2019. For external validation cohort, 678 134 Welsh residents aged >65 years on 1 January 2019 who contributed data to SAIL between 1 January 2019 and 31 December 2019.

Main outcome measures: Composite outcome of hospital admission for delirium or hospital admission for a fall or fracture (as an indicator of an injurious fall) during the study period.

Results: 71 757 patients (4147 (5.8%) with the composite outcome) comprised the development cohort and 678 134 patients (43 675 (6.4%) with the composite outcome) comprised the external validation cohort. Based on anticholinergic medications recorded in electronic health records as predictors, a prediction model for falls and delirium was developed and validated. Model coefficients for anticholinergic medications were transformed into ACMI scores. Of 88 anticholinergic medications, codeine, prednisolone, and furosemide were the most commonly prescribed (8.3%, 6.5%, and 6.0% of patients prescribed these anticholinergic drugs at least once in the development cohort, respectively). Model discrimination and calibration in the development cohort was good (C index=0.794, R2=0.06, calibration slope=0.97). The model performed well in external validation with good discrimination (C index=0.721, 95% confidence interval (CI) 0.718 to 0.734, R2=0.04) but with some evidence of miscalibration (calibration slope=0.69). The ACMI medications were scored 0.1 (desloratadine) to 2.3 (amantadine).

Conclusions: The ACMI was designed to be implemented in primary care to automatically calculate anticholinergic burden. Combined with the model predicted risk of delirium or falls, the ACMI allows clinicians to prioritise patients for structured medication reviews to reduce overprescribing.