Multimodal AI using host, tumor, and ghost biomarker for predicting immunotherapy efficacy in NSCLC.

Abstract

8578 Background: Almost all NSCLC patients (pts) without driver alterations received a first-line based immunotherapy (IO). It is unclear if pts with very poor (VP) overall survival (OS <6 months) will benefit from IO, advancing the hypothesis that this population might benefit more from next-generation drugs or supportive care. Non-response to IO is often linked to host immune fitness, such as circulating immune profiling (CIP). These "ghost biomarkers" can guide clinicians in making critical decisions, such as sparing IO in VP pts. APOLLO 11 study, aimed to develop an AI multimodal tool combining real-world data (RWD), CT-radiomics (CTRAD), and CIP to identify IO survival prediction VP pts. Methods: Data collected at Istituto Nazionale Tumori di Milano included: CTRAD features extracted using both a CT-scan Foundation Model (FM CTRAD) and PyRadiomics (pyRAD) features. The two methods were compared. Fluorescence-Activated Cell Sorting (FACS) analysis focused on identifying circulating low-density neutrophils and myeloid cells. Machine Learning (ML) multimodal pipelines for both classification (using LASSO as feature selector) and survival (using COX-ML) were developed using respectively OS < 6 months as threshold and overall survival as continuous outcome. SHAP explainability was applied to identify the most influential features contributing to model predictions. Results: Among 932 screened NSCLC pts treated with IO a (720 retrospective-R, 212 prospective-P), 495 had available baseline CT scans, with 638 lesions in the lung (397), lymph nodes (208), and pleura (29). Baseline FACS analysis was performed on 236 patients (162 R, 74 P). 117 pts had all three modalities. 4000 FM RAD features were extracted and reduced to 52 using PCA, while 107 features with pyRAD. Bimodal models with RWD and FACS achieved better performance (AUC 0.71±0.11) than bimodal models with RWD and CTRAD achieving an AUC 0.66±0.07 with pyRAD and 0.62±0.08 with FMRAD). The multimodal ML model, including all data modalities, achieved AUC 0.76 (±0.12). SHAP showed that high frequency of total myeloid cells (CD11b) and of immature neutrophils (CD10-CD16-), high LDH, high ECOG, low BMI, and two rad features as the most important for predicting VP. The survival multimodal model achieved a c-index of 0.76 ±0.10. Conclusions: The multimodal tool including all data modalities demonstrated superior performance in predicting OS. SHAP identified all data modalities as relevant, highlighting key "ghost biomarkers" for predicting OS and identifying VP pts. Given that this biomarker is fast (results within one day) and cheap (approximately €/$300), it can be easily integrated with RWD and RAD, which are readily available in clinical practice. This tool can reduce financial toxicity by guiding pts to the appropriate treatment. Finally, pyRAD features compared to FM features seem to perform better in bimodal models. Clinical trial information: NCT05550961 .