Classifying Future Healthcare Utilization in COPD Using Quantitative CT Lung Imaging and Two-Step Feature Selection...

Area of Science:

• Pulmonary Medicine
• Data Science
• Medical Informatics

Background:

• Predicting healthcare utilization in Chronic Obstructive Pulmonary Disease (COPD) is crucial for patient management.
• Existing methods for selecting predictive features lack clarity and robustness.

Purpose of the Study:

• To develop and validate a novel hybrid feature selection method for predicting healthcare utilization in COPD patients.
• To compare the performance of the proposed method against common feature selection techniques.

Main Methods:

• A retrospective study involving 454 COPD participants from the Canadian Cohort Obstructive Lung Disease study.
• Utilized demographic data, lung function measurements, and CT imaging features.
• Proposed a two-step hybrid feature selection combining sparse subspace learning (nonnegative matrix factorization) and a genetic algorithm.

Main Results:

• The hybrid method achieved a significantly higher accuracy (82% ± 3%) in predicting healthcare utilization compared to common methods (72%-80%).
• Identified key discriminative features including DLCO, FEV1, RV, FVC, TAC, LAA950, Pi-10, LAA856, and various CT-derived texture and morphologic features.
• 161 out of 454 participants (35%) utilized healthcare services at follow-up.

Conclusions:

• The developed hybrid feature selection method effectively identifies the most discriminative features for predicting future healthcare utilization in COPD.
• This approach offers improved accuracy over existing methods, aiding in better risk stratification for COPD patients.