Development and Validation of a Machine Learning Model to Explore Tyrosine Kinase Inhibitor Response in Patients With Stage IV EGFR Variant–Positive Non–Small Cell Lung Cancer

Key Points Question Can an end-to-end deep learning network model be used to identify patients with stage IV epidermal growth factor receptor (EGFR) variant–positive non–small cell lung cancer who will not benefit from EGFR–tyrosine kinase inhibitor (TKI) therapy? Findings In this diagnostic/prognostic study of 342 patients receiving EGFR-TKI therapy, a bidirectional generative adversarial network model demonstrated a 36% reduction in the progression-free survival of patients at high risk for rapid progression but no significant difference in the progression-free survival between these patients and those receiving first-line chemotherapy. The proposed deep learning semantic signature eliminated all manual interventions required while using previous radiomics methods and had a better prognostic performance. Meaning An end-to-end clinically applicable approach is promising for quantitatively identifying the benefit of EGFR-TKI therapy.

External validation cohort 1: Both chest non-enhanced and contrast-enhanced CT were performed on every patient using one of the two multi-detector row CT (MDCT) systems (GE Lightspeed Ultra 8, GE Healthcare, Hino, Japan or 64-slice LightSpeed VCT, GE Medical systems, Milwaukee, Wisconsin). All CT images were reconstructed with the standard kernel.
External validation cohort 2: 1-3mm slice thickness with or without contrast CT scans were obtained with SIEMENS SOMATOM Definition Flash scanners (Munich, Germany). The following parameters were used to obtain HRCT images: collimator with 64 × 0.6 mm, section thickness of 1 mm, reorganization interval of 0.66 mm, and tube voltage of 120 kV.
In order to reduce the impact of the variation in images from different sources on the learning efficiency of BigBiGAN, the following image standardization was performed to all the input images.

Treatment details and follow-up
Erlotinib (n=67), gefitinib (n=194), icotinib (n=51), afatinib (n=12), and osimertinib (n=18) were administered to patients in the three EGFR-TKI cohorts. Three patients in the training cohort were diagnosed with advanced NSCLC when they were first admitted to our hospital and then treated with EGFR-TKI therapy. The clinical treatments provided in the chemotherapy cohorts were as follows: of 72 patients with pathologically confirmed SCC, 39 received gemcitabine plus cisplatin, 10 received docetaxel plus carboplatin, 11 received paclitaxel plus cisplatin, and 12 received docetaxel plus cisplatin. Fifty-one patients with pathologically confirmed adenocarcinoma received bevacizumab/pemetrexed plus carboplatin. All drug doses were administered in accordance with the current clinical guidelines and the patient's condition.
The averaged follow-up interval was 4-6 weeks in patients treated with EGFR-TKI therapy.
Asymptomatic patients were followed up every 6 weeks, imaging was performed every 8-12 weeks, and for symptomatic patients more flexible and frequent follow up plan were developed. Patients treated with chemotherapy were reviewed every 3 weeks on average.

The radiomics signature we used for comparison
A prognostic radiomic signature which previously reported for EGFR-TKI efficacy prediction was used for comparison (please see the reference [16] in the manuscript). In order to construct this signature, 1032 phenotypic features were designed to be automatically extracted from the manually segmented tumor region for each patient. All the features were grouped by: 3D, texture, Gabor, and wavelet features that covered one-, two-and three-dimensional features. Then, 12 differently expressed radiomic phenotypic descriptors and their corresponding weights were obtained from the feature set in the training cohort for prognostic prediction by using the LASSO Cox proportional hazards regression. The established signature was applied to stratify the training cohort into slowand rapid-progression subgroups of EGFR inhibitor, which was achieved by using the X-tile. The signature is presented as following.  eFigure 4. Results of the comparison without the training cohort. Kaplan-Meier curves of the lowprogression-risk (blue) and high-progression-risk (red) patients who received EGFR-TKI therapy, and patients with EGFR mutation-positive (green) and EGFR wild-type (yellow) who received firstline chemotherapy. The dotted lines represent the median PFS of patients in each cohort. EGFR: epidermal growth factor receptor; TKIs: tyrosine kinase inhibitor; PFS: progression-free survival