Prediction of Tumor PD-L1 Expression in Resectable Non-Small Cell Lung Cancer by Machine Learning Models Based on Clinical and Radiological Features: Performance Comparison With Preoperative Biopsy

Abstract

Objective: We investigated if PD-L1 expression can be predicted by machine learning using clinical and imaging features.

Methods: We included 117 patients with c-stage I/II non-small cell lung cancer who underwent radical resection. A total of 3951 radiomic features were extracted by defining the tumor (within tumor contour), rim (contour ±3 mm) and exterior (contour +10 mm) on preoperative contrast computed tomography. After feature selection by Boruta algorithm, prediction models of tumor PD-L1 expression (22C3: ≥1%, <1%) of resected specimens were constructed using Random Forest: radiomics, clinical, and combined models. Their performance was evaluated by 5-fold cross-validation, and AUCs were compared using Delong test. Next, study groups were categorized as patients without biopsy (training set), and those with biopsy (test set). Predictive ability of biopsy was compared to each prediction model.

Results: Of 117 patients (66 ± 10 years old, 48% male), 33 (28.2%) had PD-L1≥1%. Mean AUC of PD-L1≥1% for the validation set in radiomics, clinical, and combined models were 0.80, 0.80, and 0.83 (P = .32 vs. clinical model), respectively. The diagnosis of malignancy was made in 22 of 38 (58%) patients with attempted biopsies, and PD-L1 was measurable in 19 of 38 (50%) patients. Diagnostic accuracies of PD-L1≥1% from 19 determinable biopsies and 38 all attempted biopsies were 0.68 and 0.34, respectively. These were out performed by machine learning: 0.71, 0.71, and 0.74 for radiomics, clinical, and combined models, respectively.

Conclusions: Our machine learning could be an adjunctive tool in estimating PD-L1 expression prior to neoadjuvant treatment, particularly when PD-L1 is indeterminable with biopsy.

Keywords: Immune-checkpoint inhibitor; Neoadjuvant immunotherapy; Programmed cell death ligand 1.