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

Area of Science:

• Pulmonary Medicine
• Health Informatics
• Machine Learning in Healthcare

Background:

• Predicting mortality in chronic obstructive pulmonary disease (COPD) is crucial for patient care and resource management.
• Traditional prediction models often require data not routinely collected in clinical practice.
• Drug claims data offers an accessible alternative for mortality prediction in COPD patients.

Purpose of the Study:

• To evaluate the efficacy of medication-based predictors for forecasting 5-year all-cause mortality in COPD patients.
• To compare the performance of logistic regression against various machine learning (ML) methods for mortality prediction.
• To identify key medication classes that serve as significant predictors of mortality in COPD.

Main Methods:

• A retrospective cohort study involving 179,168 COPD patients aged 40 years and above.
• Analysis included sociodemographic data, COPD medication use, and medications for other chronic conditions.
• Compared logistic regression with six ML methods, including deep artificial neural networks (D-ANN), for 5-year mortality prediction.

Main Results:

• The 5-year mortality rate among the study cohort was 24.3%.
• Logistic regression models achieved an AUC-ROC of 0.749 (COPD meds only) to 0.778 (adding other meds).
• Deep artificial neural networks (D-ANN) demonstrated the highest predictive performance with an AUC-ROC of 0.787, outperforming logistic regression modestly.

Conclusions:

• Medication-based prediction models can effectively estimate 5-year all-cause mortality in COPD patients.
• These models serve as a viable proxy when clinical, physiological, or imaging data are unavailable.
• Machine learning, particularly D-ANN, offers slight improvements in predictive accuracy over traditional methods for COPD mortality.