Abstract 347P
Background
All advanced NSCLC patients (pts) without driver alterations are candidates to IO therapy. Pts with very poor (VP) outcomes could instead be directed toward clinical trials or supportive care. Non-response to IO is often linked to host immune fitness, such as circulating immune profiling (CIP). Despite being informative, noninvasive and cheap, these “ghost biomarkers” face challenges in transitioning to the clinic. Within the APOLLO 11 study, we developed an AI multimodal tool combining real-world data (RWD), CT-radiomics (CT-RAD), and CIP to identify the IO VP population.
Methods
The dataset included fluorescence-activated cell sorting (FACS), CT-RAD and RWD. CT-RAD features were extracted using a CT-scan Foundation Model (FM). FACS analysis focused on identifying circulating low-density neutrophils and myeloid cells. Machine learning (ML)-based multimodal models were evaluated using cross-validation, assessing OS
Results
Among 1031 screen NSCLC pts treated with IO, in 330 patients we performed CT-rad and 154 FACS at baseline IO. 74 pts had all the modalities. 4000 CT-RAD features were extracted and reduced to 15. CT-RAD alone achieved an F1 score of 0.66 ± 0.08 and an AUC of 0.54 ± 0.07. FACS unimodal model achieved an F1 score of 0.80 ± 0.06 and an AUC of 0.77 ± 0.12. The multimodal ML model, including all data modalities, achieved an F1 score of 0.83 ± 0.03 and an AUC of 0.78 ± 0.13. The SHAP analysis reported CD11b, CD10- CD16- immature neutrophils, high neutrophil counts, low BMI, and high smoking pack-years as the main correlated features with VP.
Conclusions
The high predictive accuracy of FACS analysis, both alone and in combination with RWD emphasizes the value of integrating “ghost biomarkers” into clinical practice. The low unimodal CT-RAD performance, despite using FM, highlights the need for other biomarkers. Finally, the potential of RWD combined with complementary biomarkers remains underestimated. Thus, integrating human and cost-effective biomarkers within a simple multimodal tool is essential in the era of AI and IO.
Clinical trial identification
APOLLO11.
Legal entity responsible for the study
The authors.
Funding
Has not received any funding.
Disclosure
V. Miskovic: Financial Interests, Institutional, Invited Speaker: Novartis. L. Mazzeo: Financial Interests, Personal, Other, Honoraria: Novartis; Financial Interests, Personal, Invited Speaker: MSD; Other, Personal, Other, Conference Grants: Daiichi Sankyo, Leo Pharma, Sanofi. R. Ferrara: Financial Interests, Personal, Advisory Board, In April 2022: MSD; Financial Interests, Personal, Advisory Board: BeiGene, BMS, Sanofi, AstraZeneca; Financial Interests, Institutional, Invited Speaker: MSD, Pfizer, AstraZeneca, Roche, Sanofi, Novartis, BMS, Ipsen, Daiichi Sankyo Company. A. Prelaj: Financial Interests, Personal, Other, Training of personnel: AstraZeneca, Italfarma; Financial Interests, Personal, Invited Speaker, The Hive Project: Discussant: Roche; Financial Interests, Personal, Advisory Board, Advisory board in Lung Cancer project: BMS; Financial Interests, Personal, Other, Travel Grant: Janssen; Financial Interests, Personal, Advisory Board: Janssen, AstraZeneca; Financial Interests, Personal, Invited Speaker: Medsir, Novartis, Lilly; Financial Interests, Institutional, Invited Speaker: Bayer, BMS, AstraZeneca, Lilly, MSD, Spectrum, Roche. All other authors have declared no conflicts of interest.