Predicting Hospital Admission After a Fall in Adults ≥ 80 Years: An Interpretable Machine Learning Model and Bedside Risk Score

Introduction

Falls are the most common reason for attending the emergency department (ED) in adults ≥80 years. However, existing admission prediction decision support for this age group is lacking. This study aimed to develop an interpretable machine learning (ML) model and a bedside scoring system to predict hospital admission in this age group presenting with a fall to the ED.

Methods

A total of 7,464 retrospective attendances by 5,921 patients ≥80 years presenting with a fall to the ED were extracted from MIMIC-IV v3.1 and MIMIC-IV ED v2.2 (BIDMC, 2008-2019). A total of 110 predictors covering demographics, arrival mode, triage acuity, vital signs, blood results, injury diagnoses, comorbidities, medications and prior healthcare use were included. An XBboost ML classifier model tuned by random search with 5-fold cross-validation on 80% training-set and evaluated on 20% testing-set was developed. SHAP values were calculated. A 19-factor point risk scoring-system was derived by multivariable logistic regression.

Results

Median age was 87, females represented 68.3%, and 55.3% were admitted. 30-day and 1-year mortality were 6.1% and 22.7%. Testing AUROC was 0.88 with 80.6% specificity, 82.8% PPV, 76.8% NPV and calibration slope 1.02. Injury diagnoses represented 36% of the model explanation, blood result 23%, triage acuity 11%, arrival model 8% and comorbidities 8%. Hip fracture, other fractures, intracranial haemorrhage, leucocytosis and raised urea predicted admission. Walk-in arrival, and soft-tissue injury decreased admission risk. The point risk score (AUROC 0.86) stratified attendances into 5-bands with admission rates of 7%, 26%, 58%, 79%, and 94%.

Conclusion

An interpretable ML model using information available during ED assessment was able to predict admission after a fall in adults ≥80 years with good discrimination and calibration. A simple bedside risk point score was able to reproduce its performance. External validation is needed to evaluate its performance, clinical utility, and generalisation.