Medication-based mortality prediction in COPD using machine learning and conventional statistical methods.

Abstract

Predicting mortality in chronic obstructive pulmonary disease (COPD) patients supports clinical decision-making and resource allocation. While most existing prediction models rely on clinical, physiological, imaging, or biological measures which are not frequently collected in clinical practice, drug claims data may be electronically accessible during routine visits. We conducted a retrospective cohort study of COPD patients aged ≥40 years in Quebec with ≥2 years of public drug coverage and ≥1 dispensation of maintenance COPD medication. Predictors included sociodemographic characteristics, medication use and adherence for COPD and, in some models, for other chronic conditions. We compared logistic regression to six machine learning (ML) methods and assessed their performance in predicting 5-year all-cause mortality on a separate test dataset. Among 179,168 COPD patients (mean age 70.2 years; 47.4 % male), five-year mortality rate was 24.3 %. Logistic regression achieved an area under the receiver-operator characteristics curve (AUC-ROC) of 0.749 using only COPD medications, rising to 0.778 when adding medications for other chronic conditions. Most ML methods slightly outperformed logistic regression, with deep artificial neural networks (D-ANN) yielding the best performance (AUC-ROC = 0.787; p < 0.001). SHapley Additive exPlanations (SHAP) analysis highlighted non-inhaled anticholinergics, diuretics, antidepressants, and lipid-lowering agents as top predictors. Models using medication-based predictors can predict a substantial part of five-year all-cause mortality in COPD patients and may serve as a useful proxy for clinical, physiological, laboratory, and imaging predictors, which are often unavailable in medico-administrative databases and/or inaccessible to physicians in routine practice; they may also yield additional gains in predictive discrimination when combined with these other sources. While ML approaches, especially D-ANN, showed improved performance, gains over logistic regression were modest.