Using Machine Learning Model to Predict NSCLC Progression After Upfront Pembrolizumab

By: Julia Cipriano, MS
Posted: Thursday, October 24, 2024

A pilot study demonstrated the predictive capability of a machine learning model integrating baseline CT radiomic features for identifying patients with advanced non–small cell lung cancer (NSCLC) who may experience disease progression after first-line monotherapy with the PD-1 inhibitor pembrolizumab, according to Calum MacAulay, PhD, of BC Cancer Research Institute, Vancouver, and colleagues. Their findings, which were presented during the International Association for the Study of Lung Cancer (IASLC) 2024 World Conference on Lung Cancer (WCLC; Abstract OA03.04), may facilitate decision-making in the treatment of those with high PD-L1 expression.

“Health Canada approved pembrolizumab in the first-line setting for advanced NSCLC with a PD-L1 expression level of at least 50% and no EGFR/ALK aberrations,” the investigators explained. “The KEYNOTE-024 trial showed 55% of such patients experience disease progression on pembrolizumab monotherapy.”

The investigators leveraged data from two retrospective cohorts: a discovery training cohort (n = 97) and an external validation cohort (n = 17). All patients underwent CT imaging before the initiation of treatment (ie, baseline) and 9 to 12 weeks after receiving first-line pembrolizumab monotherapy. Using a radiomic feature extraction pipeline, the investigators extracted 2D radiomic shape, texture, and intensity features from axial CT images of lung tumor lesions. They trained a fivefold cross-validated machine learning model on a combination of baseline demographic, clinical, and CT radiomic features to determine whether the patients had disease control or progressive disease.

A receiver operating curve analysis indicated a good performance for the best-fit logistic regression model, which integrated a combination of baseline variables in predicting progressive disease (training: AUC = 0.91; validation: AUC = 0.88). Based on Kaplan-Meier plots, patients from both the training (log-rank P < .0001) and validation (log-rank P < .01) cohorts who were predicted to have a high vs low risk of experiencing progressive disease had statistically worse overall survival.

Disclosure: For full disclosure information, visit wclc2024.iaslc.org.

2024 World Conference on Lung Cancer