Development of an automated machine learning classifier for chronic obstructive pulmonary disease in non-small cell lung cancer patients using electronic health records

Résumé fourni par la source

Chronic obstructive pulmonary disease (COPD) affects 40–60% of non-small cell lung cancer (NSCLC) patients but remains frequently underdiagnosed and undertreated, contributing to adverse outcomes [1]. Our objective was to develop an automated machine learning (ML) classifier to identify potential COPD, and potentially untreated COPD among NSCLC patients. We performed a retrospective cohort study of 4,604 NSCLC patients. Manual chart review (development set; DS; [ n = 600]; test set; TS; [ n = 294]), using Global Initiative for Obstructive Lung Disease (GOLD) criteria, served as the reference standard for COPD diagnosis. Predictor selection used least-absolute-shrinkage-and-selection-operator (LASSO) logistic regression. Two logit thresholds were chosen to create binary classifiers: a ≥ 95% specificity quality-improvement threshold (QIT) for COPD identification, and Youden’s index threshold (YT) for survival analysis and association with radiographic emphysema. The classifier achieved an area-under-the-receiver-operating-characteristic-curve of 0.93 in the DS and TS (95% CI:0.91–0.95; 95% CI:0.91–0.96, respectively). The QIT yielded logit cutoffs ≥ 0.9242 (DS) and ≥ 0.9625 (TS), with sensitivities 69% and 72%, specificities 95% and 96%, positive-predictive-value (PPV) 93%, and negative-predictive-value (NPV) 77% and 80%, respectively. In the DS and TS, potential COPD prevalence was 48% and 46%, with 41% and 46% potentially untreated; the QIT-classifier identified 69% and 72% of potential COPD cases and 40% and 50% of potentially untreated cases, respectively (PPV 85%, 95% CI:73–93% and 86%, 95% CI:71–95%). YT yielded a sensitivity and specificity of 88%. In Cox-proportional hazards models ( n = 4,501, 1,959 deaths), the YT-COPD classifier was associated with increased mortality, with significant effect measure modification by clinical stage ( p = 0.003; stratified HR YT-COPD: stage I: 1.71, 95% CI:1.30–2.44, p < 0.001; Stage II: 1.72, 95% CI:1.15–2.59, p = 0.009; Stage III: 1.67, 95% CI:1.34–2.08, p < 0.001; Stage IV: 1.29, 95% CI:1.15–1.44, p < 0.001). Furthermore, classifier-predicted COPD (YT) was strongly associated with radiographic emphysema ( p < 0.001). This ML classifier effectively emulates structured chart review to identify potential COPD and potentially untreated COPD in NSCLC, with prognostic validity and potential for scalable, low-cost quality improvement and research. External validation in other institutions and EHR systems is warranted before broader implementation.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.