Session Item

Chemo-immunotherapy improves survival only in a subset of advanced non-small cell lung cancer (NSCLC) patients, and established response biomarkers, such as PD-L1 expression, have limited predictive value. In this study, we aim to develop a robust lung cancer imaging biomarker for chemo-immunotherapy response.

A cohort of 209 patients receiving first-line chemo-immunotherapy for advanced NSCLC at Dana-Farber Cancer Institute between 2015 and 2021 was utilized. Baseline thoracic computed tomography (CT) scans obtained prior to initiation of combined therapy with pembrolizumab, carboplatin, and paclitaxel were retrospectively collected and analyzed. Patients without baseline imaging or who have received prior treatment were excluded. A contrastive self-supervised learning (SSL) algorithm pre-trained for visual recognition on more than 11,000 CT data samples (with and without pulmonary lesions) was used for feature extraction. LASSO-Cox regression was used to select features strongly associated with the primary outcome of interest, progression-free survival (PFS).  A k-nearest neighbors classifier was then applied to the resultant feature vector as well as a holistic model including additional combinations of clinical, genomic, and immunophenotypic parameters. These classical variables, including PD-L1 expression pattern, tumor mutation burden (TMB), driver mutation status, histology, age, ECOG performance status, BMI, sex, and race, were also independently evaluated. All models were trained on 157 patients and cross-validated on an independent test set of 52 patients. The models’ performance in predicting PFS was quantified using the area under the receiver operating characteristic curve (AUC).   

The median (range) age at cycle one day one of therapy was 64 (56-72) years. Median ECOG performance status was 1 (IQR = 0). 179 (85.6%) patients had adenocarcinoma histology; 113 (54.1%) were female; 186 (89%) were white; median (range) BMI was 25.8 (22-29); 49 (23.4%) had high baseline PD-L1 expression (tumor proportion score≥50%), and the top decile for TMB was ≥32 mutations/Mb. Following dimension reduction of the radiographic feature vector (4096 to 117), SSL representations alone predicted PFS with AUC = 0.66. This predictive performance increased when combined with age and performance status (AUC = 0.73). Age and performance status alone had AUC = 0.52. PD-L1 expression and TMB had a combined AUC = 0.53. Other combinations of clinical variables did not significantly change model performance. Driver mutation status was not predictive.  

SSL representations of CT are predictive of PFS in treatment naïve advanced NSCLC patients receiving chemo-immunotherapy, particularly when these features are combined with clinical and demographic risk factors. They offer a predictive performance benefit over existing methods based on PD-L1 expression and TMB. Automated high-throughput analysis of lung cancer images can aid stratified personalized precision medicine.