(ASPOT14) Drug Burden Index is a Strong Modifiable Predictor of 30-Day Hospitalisation in Community-Dwelling Older Adults: Machine Learning Analysis of InterRAI Data.

Background: Older adults (≥65 years) account for an inordinately high proportion of hospitalisation and in-hospital mortality. Machine Learning (ML) models have been built and validated to predict many clinical outcomes, such as hospitalisation and mortality. However, there remains a significant need to further optimise ML models, to improve accuracy, generalisation and interpretability for clinicians.

Objectives: This study aimed to;
i) Develop three ML classification models to predict 30-day-hospitalisation.
ii) Evaluate three ML classification models in terms of discriminatory power.
iii) Identify important modifiable risk factors in predicting 30-day-hospitalisation.

Methods: In this retrospective cohort study, 14,198 community-dwelling older adults (≥ 65 years) assessed between June 2012 and July 2014 from the New Zealand Inter-Resident Assessment Instrument database were used to develop three ML models to predict 30-day-hospitalisation; Random Forest (RF), XGBoost (XGB), and Logistic Regression (LR). The three ML models utilised all 26 variables. 30-day hospitalisation was defined as any hospitalisation related to any clinical event within 30 days. Variable importance plots were used to identify predictors of 30-day hospitalisation, including non-modifiable variables such as comorbidities, and modifiable variables such as the Drug Burden Index (DBI), which quantifies exposure to medications with anticholinergic and sedative activity. The model performance metrics included accuracy, sensitivity, specificity, positive and negative predictive value, and the area under the receiver operating characteristics curve (AUC-ROC).

Results: The AUC-ROC for the RF, XGB and LR models were 0.97, 0.90 and 0.72, respectively, signifying good discriminatory power. The models also demonstrated good positive predictive value (0.88, 0.86, 0.66, respectively) and negative predictive value (0.94, 0.80, 0.66, respectively). Variable importance plots identified the DBI (var. imp. 91%, 53%, 52%, respectively) and alcohol consumption (var. imp. 97%, 68%, 98%, respectively) as important modifiable variables in predicting 30-day-hospitalisation. The ML models also identified several known non-modifiable variables of hospitalisation, including mobility, fall history and activities of daily living.

Conclusions: Identifying modifiable risk factors such as DBI and alcohol consumption may be of high clinical relevance. Clinicians may influence these factors and proactively lower the risk of 30-day hospitalisation. This is the first study to identify the DBI as an important modifiable variable in predicting 30-day hospitalisation for adults ≥65 years. It is crucial, however, that these models undergo extensive validation through clinical studies.