The discovery of immune checkpoint inhibitors (ICIs) revolutionized the treatment of advanced melanoma, however, there is currently no easy way to predict patient response or their risk for immune-related toxicities without conducting RNA sequencing – a method rarely found at community oncology clinics. Researchers report an emerging technique that aims to solve this problem by integrating artificial intelligence deep learning in histology specimens with patient clinical data in the form of an algorithm that may predict ICI response in advanced melanoma patients. Findings from the study were published in Clinical Cancer Research and show great promise for advancements in the future of melanoma care.
AI Model Predictions
Led by Paul Johannet, MD from the NYU Grossman School of Medicine in New York City, a team of researchers evaluated a training cohort from New York University and validation cohort from Vanderbilt University to develop a multivariable classifier that would utilize deep learning to combine histology slides with patient data. The researchers generated a receiver operating
characteristic curve and stratified patients as either high or low risk for cancer progression using an optimal threshold. Progression-free survival was compared between the groups and validation of the classifier was performed on two slide scanners, the Aperio AT2 and Leica SCN400. The clinical outcomes measured were the progression of disease or response, including complete and partial responses.
The artificial intelligence algorithm – based on deep convolutional neural networks (DCCN) – analyzed digital images of hematoxylin and eosin-stained slides of metastatic lymph nodes as well as subcutaneous tissue to identify patterns related to patient response. A response classifier was used to predict whether an untreated tumor would positively respond to ICI treatment or whether it would progress.
To assess the performance of the DCCN classifier, the researchers calculated the area under the curve (AUC) which determines the model’s accuracy in both study cohorts. Finally, to augment prediction accuracy of the model, the study’s authors performed multivariable logistic regression models combining the DCCN predictions with conventional clinical measures.
Immunotherapy Response
Dr. Johannet and his team reported that the multivariable classifier predicted the response to ICI in patients with areas under the curve of 0.800 and 0.805 on imaging from Aperio AT2 and Leica SCN400, respectively. The DCCN prediction model achieved an AUC of approximately 0.7 in both groups, while an AUC of 1 represents perfect prediction.
Additionally, the classifier accurately stratified patients into high versus low risk for disease progression. Patients from the Vanderbilt cohort classified as high risk for progression had significantly worse progression-free survival than those defined as low risk.
The results of the AI model remained consistent regardless of which therapy patients received implicating that some biomarkers may not be necessarily specific to the checkpoint target, the authors explained.
Histology slides and patient clinical data characteristics are readily available; they have great potential to be used in the prediction of immunotherapy responses and disease progression in melanoma patients.
“We believe this computational approach has the potential for integration into clinical practice,” the authors concluded. “This could help oncologists identify patients who are at high versus low risk for progression through immunotherapy.” Although, the results warrant further study, including larger participant cohorts as well as additional validation parameters, before the model can be extended to clinical practice.