Machine Learning to Refine Selection Criteria of Post-Operative Radiotherapy for Resectable Stage III-N2 Non-Small Cell Lung Cancer

Abstract

<h3>Purpose/Objective(s)</h3> The role of post-operative radiotherapy (PORT) for completely resected N2 non-small cell lung cancer (NSCLC) is controversial, given the recently published LungART and PORT-C trials. We sought to utilize machine learning to identify a subset of patients who may still benefit from PORT based on the extent of nodal involvement. <h3>Materials/Methods</h3> Patients with completely resected N2 NSCLC were identified in the National Cancer Database diagnosed between 2004 and 2017. The query was subsequently limited to patients with pN2 disease, known pathologic T stage, and no evidence of distant metastasis. Patients were excluded if they had unknown follow-up, did not undergo surgery (lobectomy, bilobectomy, or pneumonectomy), did not undergo lymph node dissection, had positive margins, had unknown number of positive lymph nodes or total lymph nodes examined, or received neoadjuvant radiation. We trained a machine-learning based model of overall survival (OS). SHapley Additive exPlanation values (SHAP) were used to identify prognostic and predictive thresholds for the number of positive lymph nodes (LNs) involved and lymph node ratio (LNR). Multivariable (MVA) Cox proportional-hazards regression was used for confirmatory analysis. <h3>Results</h3> A total of 16,789 patients with completely resected N2 NSCLC were identified. Of the patients that met inclusion criteria, 6181 (36.8%) received PORT. When radiation treatment technique could be determined, about half received 3D conformal radiation therapy (18.7% of all patients) and about half received IMRT (19.9% of all patients). Median follow-up for the entire cohort was 32.0 months (0-183.1 months). Median OS of the whole population was 41.7 months (95%CI: 40.7-43.1). For patients who received PORT, median OS was 47.6 months (95%CI: 45.2-49.7). For patients who did not receive PORT, median OS was 38.6 months (95%CI: 37.1-39.9). Using SHAP values, we identified thresholds of 3+ positive LNs and a LNR of 0.34+. On MVA, PORT was not significantly associated with OS (HR: 0.96; 95%CI: 0.92-1.01); (p=0.111). However, a subset analysis of patients with 3+ positive LNs revealed that PORT improved OS (HR: 0.91; 95%CI: 0.86-0.97; p=0.002). Patients with 3+ positive LNs had a 5-year OS of 38% (95%CI: 36-40%) with PORT compared to 31% (95%CI: 29-32%) without PORT. Patients with 1-2 positive LNs had a 5-year OS of 49% (95%CI: 47-51%) with PORT compared to 44% (95%CI: 43-46%) without PORT. On a separate subset analysis, in patients with a LNR of 0.34+, PORT improved OS (HR: 0.90; 95%CI: 0.85-0.96; p=0.001). Patient with positive LNR of 0.34+ had a 5-year OS of 38% (95%CI: 36-40%) with PORT compared to 29% (95%CI: 28-31%) without PORT. Patients with a positive LNR of 0.00-0.33 had a 5-year OS of 47% (95%CI: 45-49%) with PORT compared to 43% (95%CI: 42-44%) without PORT. <h3>Conclusion</h3> This study suggests that patients with 3+ positive LNs or a LNR of 0.34+ may present a subpopulation of patients who could benefit from PORT. These findings warrant further investigation in a future prospective study.