Immunotherapy Efficacy Prediction in Cancer: An Artificial Intelligence Approach with Unannotated H&E Whole-Slide Images

Developing a solution to predict clinical outcomes for immunotherapy that is accurate, scalable, affordable, clinically meaningful, and globally accessible is an unmet medical need. Precise prediction of patient response to immunotherapy from pretreatment biopsy images will enable the delivery of immuno-oncology drugs to suitable patients and reduce the risk of administering unnecessary toxicity to patients who do not benefit from it. We propose an AI-based framework to produce stratifying algorithms that only need routinely obtained unannotated hematoxylin and eosin (H&E)-stained whole slide images. This design choice eliminates the need for pathologist annotations, ensuring affordability and scalability. Our solution, developed with data from 418 durvalumab patients, was validated both for head and neck squamous cell carcinoma and non-small cell lung cancer with data from 283 durvalumab patients, demonstrating its versatility and ease of adaptation for different indications. The results obtained using test data from clinical trials, different from training data, exhibit clinically meaningful improvement between those classified as positive and negative. For median overall survival (OS), the enhancement is in the range [55.9%, 198%] and [0.49, 0.70] for the hazard ratio for OS. For median progression-free survival (PFS), the improvement ranges within [39%, 195%], while the hazard ratio is within [0.46, 0.86] for PFS. Our solution complements the current biomarker, programmed death lig- and 1, for immunotherapy therapy, presenting an opportunity to develop more accurate solutions. In addition, as the algorithm was developed in a hypothesis-free approach, the analysis of the converged solution may enhance expert understanding of the pathomechanisms driving the response to immunotherapy. Its scalability and accuracy, combined with the AI-based engineering nature of the solution, bring the opportunity of being globally deployed using the cloud. Our technique has the potential to unlock opportunities not available for patients before by enabling the generation of efficient and affordable diagnoses for immunotherapy.