86P Artificial intelligence tools to predict early progression of non-small cell lung cancer patients treated with immunotherapy: Preliminary evidence from a single-centre experience of black-box and white-box models pairing

Immunotherapy (IO) alone or in combination with chemotherapy (CT-IO) has become the standard of care for most patients with metastatic Non-Small Cell Lung Cancer (NSCLC). Nevertheless, 20-30% of patients do not benefit from these treatments facing a negative risk-benefit balance and the identification of reliable predictive factors for a better patients' selection is an important unmet need. In this retrospective single-centre analysis, patients with metastatic NSCLC treated with first line (CT)-IO between 2015 and 2021 were eligible. Clinical, pathological and laboratory data were retrieved. Primary aim was to identify the features predicting early progression (radiological progression or death within 3 months from treatment start) by artificial intelligence (AI) models. 201 patients were included: 121 IO (all PD-L1 ≥50%), 80 CT-IO (all PD-L1 <50%). 46% in IO group and 36% in CT-IO group were early progressors. Early progression correlated with poor performance status, elevated neutrophil percentage (>80%) and neutrophil/lymphocyte ratio (≥8), lower-range PD-L1 status (50%-75%, in IO only), use of steroids (in IO only), bone and liver metastases (in CT-IO) by a statistical exploratory data analysis. In IO only population, black-box models (Automated Machine Learning, AutoML) encompassing clinical, laboratory and pathological data reached an AUC of 0.82 in predicting early progression (best model: Voting Ensemble). White box tools (SHapley Additive exPlanations) paired with AutoML generated models increased the interpretability of the results despite slightly lower precision scores. Dimensionality reduction algorithms such as t-SNE (t-distributed stochastic neighbor embedding) and SVM (Support Vector Machine) were applied on our dataset to rebalance sample size with feature list, obtaining a reliable pattern of cases distribution. Artificial intelligence models can predict trustworthy patterns of early progression in IO-treated patients. Pairing black box and white box methods is feasible and could increase robustness and interpretability of the results.