Multimodal artificial intelligence models from baseline histopathology to predict prognosis in HR+ HER2- early breast cancer: Subgroup analysis.

101 Background: Prognostic assessment in HR+ HER2- early breast cancer (EBC) remains challenging given relatively low rates of disease progression. Modern artificial intelligence (AI)-based techniques have already provided substantial medical progress, particularly in prostate cancer. We have leveraged ArteraAI’s multimodal artificial intelligence (MMAI) platform to develop a research-level prognostic model in HR+ HER2- EBC, based on the WSG PlanB and ADAPT trials. Here, we quantify the value added by MMAI within clinically relevant subgroups. Methods: Histopathology image data was generated from pre-treatment breast biopsy and surgical hematoxylin and eosin (H&E) slides from the WSG PlanB and ADAPT trials. Patients with available images and complete data (n=5259) were allocated (stratified by trial, randomization arm and distant recurrence (DR)) to development (60%) and validation (40%) cohorts. An MMAI-based model using image data combined with clinical prognostic variables (age, T and N stage, tumor size) was developed to predict risk of DR. Univariable and multivariable Fine-Gray models were used to assess performance in the validation cohorts; subdistribution hazard ratios (sHR) refer to validation cohorts and are reported per standard deviation increase of the model scores (image-alone or combined). Pre-specified prognostic subgroups for analysis were defined by nodal status, menopausal status, and central tumor grade. All statistical tests were 2-sided at .05 significance. Results: The trained MMAI score was significantly associated with risk of DR in the validation cohort (sHR [95% CI] = 2.3 [2.0-2.8]) as a whole and in all considered subgroups. The score remained significant (sHR [95%CI] = 2.2 [1.7-2.8]) after adjusting for clinical prognostic factors. Moreover, the MMAI image component alone had significant prognostic value (sHR [95%CI] = 1.6 [1.3 - 1.9]) in the validation cohort. Remarkably, the MMAI image component alone had significant prognostic value separately within the G2 and G3 sub-groups, with sHR of about 1.5 per standard deviation increase, and also in most of the other predefined clinical subgroups. Conclusions: Preliminary results from the current MMAI breast model provide evidence that ArteraAI MMAI technology can be leveraged for outcome prediction in HR+ HER2- EBC using H&E-stained images to further personalize breast cancer management. The ability of image-only AI models to provide significant prognostic value within grade subgroups suggests that self-supervised AI has identified some novel image features with prognostic value beyond grade. To put the results into the clinical context, comprehensive validation analyses will be presented at the meeting.