EHR-Based One-Year Fall Risk Tool Enables Targeted Medication Reviews

• An electronic health record (EHR)-based XGBoost model predicted 1‑year fall hospitalisation with high discrimination: AUROC 0.979 (internal) and 0.939 (external), and substantially greater sensitivity than the Morse Fall Scale (0.569 vs 0.139).

• Pharmacists can use the risk score to prioritise targeted medication reviews and deprescribing of fall-risk-increasing drugs (FRIDs), focusing limited resources on high-risk older adults (frequent emergency department use, polypharmacy, diabetes).

• Falls in older adults cause significant morbidity, frequent emergency department visits and hospitalisations and major healthcare costs

• conventional bedside tools require staff time, underperform in community settings and miss many high-risk older adults.

• Pharmacy teams face a triage challenge: high prevalence of polypharmacy and FRIDs combined with limited pharmacist capacity makes universal manual review impractical—imprecise screening wastes time and risks missing patients who would benefit from deprescribing.

• Embedding EHR-derived risk flags into clinical decision support enables pharmacy stewardship programmes to prioritise targeted medication reviews and allocate scarce clinical resources more efficiently.

• Retrospective territory-wide study using Hospital Authority Data Collaboration Laboratory (HADCL) EHRs: ~4.9 million records from 1.14 million adults aged ≥65 years (2013–2017)

• outcome = 1‑year fall hospitalisation defined by ICD‑9 codes.

• Assembled 260 predictors (demographics, utilisation, diagnoses, medication dispensing including FRIDs, laboratory summaries)

• trained six supervised machine-learning algorithms and selected XGBoost after hyperparameter tuning (optimising AUPRC).

• Used a 70/20/10 train/test/validation split, handled missing laboratory data with indicator (dummy) variables, benchmarked against logistic regression, externally validated in the Hong Kong Diabetes Register (HKDR) and compared performance to the Morse Fall Scale in an age‑ and sex‑matched subcohort.

• XGBoost achieved AUROC 0.979, AUPRC 0.764 and PPV 0.614 in internal validation, substantially outperforming logistic regression (AUROC 0.885, AUPRC 0.169, PPV 0.210).

• External validation in the Hong Kong Diabetes Register (n = 13,917

• 522 fall events) yielded AUROC 0.939 and AUPRC 0.312, supporting robustness in a separate cohort.

• In an age‑ and sex‑matched subcohort versus the Morse Fall Scale, XGBoost sensitivity was 0.569 versus 0.139 (MFS)

• XGBoost F1 = 0.626, indicating markedly improved case detection while balancing false positives and negatives.

• Top predictors were emergency department attendances, fasting plasma glucose (FPG), outpatient/inpatient visit counts, prior falls, age and medication use (notably renin–angiotensin system inhibitors, nitrates and antiplatelets)

• these medication-related signals provide concrete targets for prioritised pharmacist review and deprescribing.

• The model’s improved performance was driven by rich EHR signals—utilisation patterns, laboratory results and dispensing histories—rather than any single conventional bedside measure.

• Embed the XGBoost score into the HA EHR as a daily high-risk dashboard/worklist

• use the study’s scaled cut-off (score = 92, which captured ~90% of events in validation) as an initial triage threshold for targeted medication review.

• Route flagged patients into pharmacy medication‑review queues prioritising FRIDs (renin–angiotensin system inhibitors, nitrates, antiplatelets), frequent ED/inpatient users and people with diabetes

• allocate protected pharmacist time and provide brief SHAP-guided training to interpret model drivers in the chart.

• Implement governance requiring documented pharmacist clinical judgment before deprescribing, an audit-and-feedback loop to monitor calibration and performance, and planned prospective validation with periodic retraining in the local EHR environment.

• Operationally treat the model as a territory-wide pre-screening radar that surfaces candidates for medication review

• pharmacists remain the decision-makers who verify signals, adjust therapy and own patient safety.

Strengths:

• Territory-wide EHR dataset (≈1.14M patients, 4.9M records) with 260 multidimensional predictors, hyperparameter tuning, SHAP explainability and external validation.

• Reporting AUPRC alongside AUROC, and benchmarking versus logistic regression and the Morse Fall Scale, strengthens performance assessment and implementation readiness.

Limitations:

• Retrospective 2013–2018 data and use of ICD‑9 hospital codes capture only hospitalised falls, potentially underestimating non‑hospitalised events and limiting temporal generalisability.

• Lifestyle, frailty and environmental variables were unavailable

• external validation was limited to the HK Diabetes Register and prospective, real‑time testing in operational workflows remains pending.

An EHR-derived XGBoost score reliably flags 1‑year fall-hospitalisation risk and is ready for EHR-integrated pilot deployment to triage pharmacist-led medication reviews and deprescribing, with governance, prospective validation and ongoing monitoring.