Dynamic EHR Fall Risk Outperforms Hester-Davis for Pharmacist Triage

Predictive Model Development

Authors:Hojjat Salehinejad; Ricky Rojas; Kingsley Iheasirim, et al.

Journal:Scientific Reports

Static scores miss evolving risks. By tracking changing vitals and meds, this XGBoost model from the Mayo Clinic achieved an AUC of 0.87 compared to Hester-Davis’s 0.60. Leverage these dynamic signals to prioritize medication safety reviews.

November 25, 2025

Quick Take

A dynamic EHR‑driven Extreme Gradient Boosting (XGBoost) model separated fallers from non‑fallers much better than the bedside Hester–Davis (HD) score.

In this cohort XGBoost AUC was 0.87 (95% CI 0.86–0.88) with sensitivity ≈0.79 and specificity ≈0.77, versus HD AUC ≈0.57–0.60 at thresholds 7 and 20.

Top contributors included HD score, oxygen saturation (SpO₂), heart rate, neurologic diagnoses, behavioral abnormalities, and IV furosemide.

Use these dynamic risk signals as a triage aid to focus pharmacist review time on flagged patients. Pay particular attention to patients receiving IV loop diuretics, sedative or antipsychotic drugs, and opioids, and interpret risk in the context of changing vitals, labs, and mental status.

Why it Matters

Hospital falls are common, often injurious, publicly reportable events that increase mortality, length of stay, and cost. Accurate triage matters for patient safety and for how we deploy pharmacist and nursing resources.

Many units rely on point‑in‑time scores recorded at admission or per shift (for example, the HD score). Static tools can over label low‑risk patients and miss evolving risk as vitals, labs, mental status, and medication exposures change, which wastes effort and creates missed prevention opportunities.

Medication exposures (diuretics, CNS agents, opioids, antipsychotics) are modifiable contributors to falls and map directly to pharmacist stewardship actions.

Inpatient pharmacists need timely, patient‑specific risk signals to prioritize reviews.

What They Did

Retrospective multicenter cohort of 46,695 adult inpatients (4,245 fallers) across 17 hospitals in one integrated health system (Mayo Clinic) from 2018–2023; intensive care, hospice, and psychiatric units were excluded.

Built four dynamic machine learning models (KNN, SVM, Random Forest, XGB) using 92 admission‑day variables plus 8‑ and 24‑hour updates of vitals, daily labs, administered medications, and nursing HD scores.

Addressed class imbalance by down‑sampling non‑fallers for training, tuned hyperparameters, used 10‑fold cross‑validation, and evaluated models on both balanced and real‑world test sets, including calibration and SHAP‑based interpretability.

All predictors came from the EHR, including incident‑report falls, medication administration records, labs, and vitals. Missing values were forward filled, and multiple measures per time interval were aggregated to a median.

What They Found

The multivariable XGB model achieved AUC 0.87 (95% CI 0.86–0.88), accuracy and F1 ≈0.78, sensitivity ≈0.79, and specificity ≈0.77. HD‑based thresholds alone produced AUCs ≈0.57–0.60.

In this cohort the dynamic XGB model misclassified roughly one in five patients, which was markedly fewer errors than the HD thresholds, where about two in five patients were misclassified depending on the cut‑point.

SHAP analysis ranked top predictors as HD score, SpO₂, heart rate, pre‑existing neurologic disease or neurologic admission, behavioral abnormalities, and IV furosemide. Medication exposure, especially IV furosemide, contributed meaningfully to risk but also likely reflected underlying acuity and toileting burden.

The dynamic, time‑updated model produced clearer separation of high‑ and low‑risk patients than the static HD score and yielded patient‑level risk probabilities that track clinical course, including medication‑linked signals as therapies, vitals, and labs change.

Takeaways

Treat dynamic, EHR‑updated fall‑risk models as triage tools to focus pharmacist effort rather than as stand‑alone prescribing engines. Use predictions to prioritize reviews, medication reconciliation, and deprescribing or simplification work.

Interpret medication signals in context. IV furosemide often reflects heart failure, volume shifts, electrolyte risk, and toileting needs as much as direct drug effect, so pharmacists should review dose, route, timing, volume status, electrolytes, and concurrent sedating or blood‑pressure‑lowering agents before changing therapy.

Before adopting a similar tool locally, clarify operational details: how medications are encoded (agent, dose, route, timing, cumulative burden), expected positive predictive value at your fall prevalence, how often scores update, and who will monitor calibration and model drift.

Local validation, defined pharmacist and nursing roles, alert‑threshold planning, and ongoing monitoring are required. Single‑system retrospective training and down‑sampling mean bedside performance and alert burden may differ in your environment.

Strengths and Limitations

Strengths:

Large multicenter EHR cohort (46,695 patients across 17 hospitals) with dynamic 8‑ and 24‑hour time‑series inputs, SHAP‑based interpretability, and explicit calibration assessment using Brier scores.

Robust model development with hyperparameter tuning, 10‑fold cross‑validation, comparison across multiple algorithms, testing on both balanced and natural data distributions, and clear reporting of discrimination and calibration metrics.

Limitations:

Retrospective, single integrated health system whose population was older, less racially diverse, and relatively affluent, so generalizability to other health systems and safety‑net settings is uncertain.

Outcome ascertainment relied on incident‑report falls and excluded ICU and psychiatric units, which may miss unreported events and limits applicability to those settings.

Medication data were primarily modeled as exposure to specific agents rather than detailed dose or administration‑time patterns, which limits direct translation into drug‑specific prescribing rules.

Training used down‑sampling to balance classes; in real‑world low‑prevalence settings, positive predictive value and alert burden at any chosen threshold will differ and need prospective evaluation.

Bottom Line

Dynamic EHR‑updated fall‑risk models, with XGB as the best performer in this study, provide credible signals that can help pharmacists target high‑risk patients and high‑risk medication regimens. Use these models as triage layers alongside HD and clinical judgement, and if you pilot locally, include richer medication encoding, clear workflows for pharmacist review, calibration checks, and continuous monitoring before full operational use.