(133) Can Machine Learning Models Help Predict Potentially Inappropriate Medication Use in Older Adults? A Population-Based Cohort Study

Publication Number: 464

Background: Potentially inappropriate medications (PIMs) are associated with negative health outcomes in older adults, such as increased risk for hospitalizations and adverse drug reactions. Logistic regression is commonly used to predict outcomes related to PIM use and to estimate associated factors. However, other models have been less investigated, especially those exploiting the potential of large databases.

Objectives: To compare the performances of machine learning (ML) models in predicting use of PIMs and quantifying the associated factors, in community-dwelling older adults.

Methods: This is a retrospective population-based cohort study using medico-administrative databases (Quebec Integrated Chronic Disease Surveillance System) in the province of Quebec, Canada. The study population is older adults (at least 66 years old) alive and covered by the public drug insurance plan from April 1st, 2015 to March 31st, 2016. We used logistic regression, random forests, neural networks, gradient boosting machines, and naïve Bayes to predict PIM use. The latter is a binary outcome defined as having at least one prescribed PIM, based on the 2015 Beers criteria. Potential predictors included age, sex, residential status (urban/rural), material and social deprivation indices, number of hospitalizations in 2015, and chronic diseases. Performance metrics were sensitivity, specificity, positive (PPV) and negative predicted values (NPV), and area under the curve (AUC); relative variable importance was measured for each model. Data were separated in training (70%) and testing (30%) sets.

Results: Out of 1,105,295 older adults, 533,719 were PIM users (48.3%). No imbalanced issue, which can negatively affect prediction, was thus observed. Performances according to AUC, PPV, and NPV were similar across the five investigated models (mean values: 0.62, 0.52, and 0.70 respectively), while sensitivity and specificity ranged from 0.38 to 0.52, and from 0.70 to 0.80 respectively. Overall, the two most (least) important variables for prediction were being a woman and anxiety-depressive disorders (heart failure and deprivation indices).

Conclusions: No model clearly outperformed the others. Furthermore, the investigated models yielded similar results in the ranking of prediction variables. ML models may help slightly refine predictions, but other variables may be more helpful in the case of PIM use prediction. Other settings could also be more favorable to ML models, such as unbalanced outcomes or use of only continuous predictors.