Combination of Deep-Learning Models Provides Significant Risk-Assessment of Survival in Pancreatic Cancer Patients, Holding Promise as a Prognostic Technology

On October 13, 2020 NantHealth, Inc. (NASDAQ: NH), a next-generation, evidence-based, personalized healthcare company, presented virtually with NantOmics and ImmunityBio on September 29-30 2020, a session entitled "Deep-learning image-based tumor, stroma and lymphocytes spatial relationships and clinical features that affect survival in pancreatic cancer patients," at the American Association for Cancer Research (AACR) (Free AACR Whitepaper) virtual special conference on pancreatic cancer (Press release, NantHealth, OCT 13, 2020, View Source [SID1234568386]). This Nant technology is an example of how digital pathology solutions may support the treatment of cancer.

Schedule your 30 min Free 1stOncology Demo!
Discover why more than 1,500 members use 1stOncology™ to excel in:

Early/Late Stage Pipeline Development - Target Scouting - Clinical Biomarkers - Indication Selection & Expansion - BD&L Contacts - Conference Reports - Combinatorial Drug Settings - Companion Diagnostics - Drug Repositioning - First-in-class Analysis - Competitive Analysis - Deals & Licensing

                  Schedule Your 30 min Free Demo!

Prepared in collaboration with NantOmics, ImmunityBio, and the University of Colorado School of Medicine, the presentation examined differential survival in pancreatic cancer patients via stromal and lymphocyte density. In this study, the contributing researchers developed an automated deep-learning system to provide risk assessment upon spatial relationships between tumor, stroma, and lymphocyte regions in pancreatic pathology images from 82 pancreatic adenocarcinoma patients who underwent chemotherapy. Using Gaussian mixture models, researchers identified optimal thresholds in the image-based features and organized patients into unsupervised clusters, then linked those to differences in survival. Risk models were generated on standard clinicopathological features and used to compare against the proposed image-based risk models.

The study’s key findings include:

Cox PH models trained on image features more significantly stratified risk on unseen test patients than an optimal set of clinicopathological features
Image-based models suggest low-risk patients have:
Higher tumor-infiltrating lymphocytes despite having fewer overall lymphocytes
Higher tumor-adjacent stroma
The combination of both risk-models proved to be superior in training and testing sets and achieved better separation in survival curves than either model by itself
Image-based risk-associated features are independently prognostic of clinicopathological features
Even with a limited sample-size, the results showed significance to warrant a larger study
"Our combined analysis with NantOmics, ImmunityBio and the University of Colorado School of Medicine, provided results with a level of significance that show great promise for future studies using deep-learning models as a prognostic tool for pancreatic cancer patients," said Christopher Szeto, Director of Machine Learning, NantHealth. "These results provide a strong clinical platform not only to better understand pancreatic cancer for evidence-based prognosis but also reinforce the essential nature of advanced technology in the continued evolution of medicine."

The AACR (Free AACR Whitepaper) Virtual Special Conference: Pancreatic Cancer is a conference focused on bringing together experts across industries to better understand and further advance pancreatic cancer research and treatment. The conference attracts a range of attendees, including those in government, clinical roles, patient advocacy, and companies who make groundbreaking technology and discoveries in oncology.