Machine learning-based mapping of multimodal treatment allocations and prognostic implications of clinical concordance in non-small cell lung cancer

Abstract

Machine learning models of non-small cell lung cancer (NSCLC) treatment usually examine one modality. We developed interpretable models of historical surgery, chemotherapy, and radiotherapy receipt in SEER. We assessed regional transportability and associations between observed treatment-model concordance and survival. We analyzed 186,712 patients with NSCLC diagnosed in 2010–2022 in SEER. The development and nonoverlapping geographic validation cohorts included 72,995 and 113,717 patients, respectively. Seven algorithms modeled historical treatment receipt, interpreted using SHAP. Survival analyses used PSM, 60-month RMST, Fine–Gray and time-varying Cox models, and a 3-month landmark analysis. XGBoost achieved internal/geographic AUROCs of 0.893/0.885 for surgery and 0.853/0.835 for chemotherapy. LightGBM yielded 0.766/0.749 for radiotherapy. Tumor stage, N stage, and age led predictions. Global tests rejected proportional hazards in all nine Cox models; RMST was the primary survival summary. At 60 months, overall-survival RMST differences were 15.41 months for surgery, 9.71 for chemotherapy, and 1.40 for radiotherapy (all P < 0.001). Exploratory radiotherapy analyses showed higher all-cause mortality in stages I (HR 1.564) and II (HR 1.321; both P < 0.001). Radiotherapy survival curves crossed near 35 months; the 3-month landmark overall-survival RMST difference was − 2.94 months. Interpretable ML models captured historical multimodal treatment receipt in SEER. Concordance was associated with survival, but associations varied by modality, stage, and follow-up time. Residual confounding precludes treatment-effect or clinical-utility interpretations. These models should not guide treatment before independent prospective validation with richer clinical, molecular, and treatment-timing data.