A multimodal AI biomarker PATH-ORACLE improves prediction of recurrence in stage I lung adenocarcinoma

Background Surgical resection is the standard treatment for stage I lung adenocarcinoma, in most cases without additional systemic adjuvant treatment. A substantial proportion of stage I cases recur, with a 5-year survival rate <50%. Clinical data suggest that adjuvant treatment, including immune checkpoint inhibitor therapy, may improve survival in such recurrent cases. Previously evaluated predictors, including the International Association for the Study of Lung Cancer (IASLC) grading system applied to histological sections and transcriptomic profiles, have not been sufficiently accurate or consistent for risk stratification or to guide therapeutic intervention. We hypothesized that these diverse diagnostic measurements carry complementary information that may provide higher prognostic power when combined. Materials and methods We developed PATH-ORACLE, a multimodal deep learning biomarker built on top of the prospectively validated transcriptomic-based Outcome Risk Associated Clonal Lung Expression (ORACLE) score, with the addition of routine histological sections processed by pretrained foundation models. Predictive performance was assessed in two independent cohorts. Results The histology-only predictor outperformed automated IASLC grading and remained prognostic after adjustment for tumor size, grade and T-stage. PATH-ORACLE exceeded both constituent modalities, predicting 1-year recurrence with an area under the curve of 0.86 and 0.84 in the two independent validation cohorts, and separated high- and low-risk groups using a single prespecified cut-off. Conclusions Extensive validation will be needed to convert PATH-ORACLE from a prognostic biomarker to a predictive biomarker that could be used to prioritize stage IB patients for adjuvant targeted therapy, chemotherapy, or immune checkpoint inhibitor therapy with or without liquid biopsy-based monitoring.