Proof-of-concept study ‘highlights that using AI to integrate different types of clinically informed data to predict disease outcomes is feasible’ researchers say
Artificial intelligence (AI) and machine learning are—in stepwise fashion—making progress in demonstrating value in the world of pathology diagnostics. But human anatomic pathologists are generally required for a prognosis. Now, in a proof-of-concept study, researchers at Brigham and Women’s Hospital in Boston have developed a method that uses AI models to integrate multiple types of data from disparate sources to accurately predict patient outcomes for 14 different types of cancer.
The process also uncovered “the predictive bases of features used to predict patient risk—a property that could be used to uncover new biomarkers,” according to Genetic Engineering and Biotechnology News (GEN).
Should these research findings become clinically viable, anatomic pathologists may gain powerful new AI tools specifically designed to help them predict what type of outcome a cancer patient can expect.
“Experts analyze many pieces of evidence to predict how well a patient may do. These early examinations become the basis of making decisions about enrolling in a clinical trial or specific treatment regimens,” said Faisal Mahmood, PhD (above) in a Brigham press release. “But that means that this multimodal prediction happens at the level of the expert. We’re trying to address the problem computationally,” he added. Should they be proven clinically-viable through additional studies, these findings could lead to useful tools that help anatomic pathologists and clinical laboratory scientists more accurately predict what type of outcomes cancer patient may experience. (Photo copyright: Harvard.)
AI-based Prognostics in Pathology and Clinical Laboratory Medicine
The team at Brigham constructed their AI model using The Cancer Genome Atlas (TCGA), a publicly available resource which contains data on many types of cancer. They then created a deep learning-based algorithm that examines information from different data sources.
Pathologists traditionally depend on several distinct sources of data, such as pathology images, genomic sequencing, and patient history to diagnose various cancers and help develop prognoses.
For their research, Mahmood and his colleagues trained and validated their AI algorithm on 6,592 H/E (hematoxylin and eosin) whole slide images (WSIs) from 5,720 cancer patients. Molecular profile features, which included mutation status, copy-number variation, and RNA sequencing expression, were also inputted into the model to measure and explain relative risk of cancer death.
The scientists “evaluated the model’s efficacy by feeding it data sets from 14 cancer types as well as patient histology and genomic data. Results demonstrated that the models yielded more accurate patient outcome predictions than those incorporating only single sources of information,” states a Brigham press release.
“This work sets the stage for larger healthcare AI studies that combine data from multiple sources,” said Faisal Mahmood, PhD, Associate Professor, Division of Computational Pathology, Brigham and Women’s Hospital; and Associate Member, Cancer Program, Broad Institute of MIT and Harvard, in the press release. “In a broader sense, our findings emphasize a need for building computational pathology prognostic models with much larger datasets and downstream clinical trials to establish utility.”
Future Prognostics Based on Multiple Data Sources
The Brigham researchers also generated a research tool they dubbed the Pathology-omics Research Platform for Integrative Survival Estimation (PORPOISE). This tool serves as an interactive platform that can yield prognostic markers detected by the algorithm for thousands of patients across various cancer types.
The researchers believe their algorithm reveals another role for AI technology in medical care, but that more research is needed before their model can be implemented clinically. Larger data sets will have to be examined and the researchers plan to use more types of patient information, such as radiology scans, family histories, and electronic medical records in future tests of their AI technology.
“Future work will focus on developing more focused prognostic models by curating larger multimodal datasets for individual disease models, adapting models to large independent multimodal test cohorts, and using multimodal deep learning for predicting response and resistance to treatment,” the Cancer Cell paper states.
“As research advances in sequencing technologies, such as single-cell RNA-seq, mass cytometry, and spatial transcriptomics, these technologies continue to mature and gain clinical penetrance, in combination with whole-slide imaging, and our approach to understanding molecular biology will become increasingly spatially resolved and multimodal,” the researchers concluded.
Anatomic pathologists may find the Brigham and Women’s Hospital research team’s findings intriguing. An AI tool that integrates data from disparate sources, analyzes that information, and provides useful insights, could one day help them provide more accurate cancer prognoses and improve the care of their patients.
New biomarkers for cancer therapies derived from the research could usher in superior clinical laboratory diagnostics that identify a patient’s suitability for personalized drug therapies and treatments
Once approved for clinical use, not only would these biomarkers become targets for specific cancer therapies, they also would require development of new diagnostic tests that anatomic pathologists could use to determine whether a biomarker was present in a patient.
If yes, the drug can be administered. If no, the patient is not a candidate for that drug. Thus, this research may produce both diagnostic biomarkers and therapeutic targets.
Relevance of In-Depth Tumor Profiling to Support Clinical Decision-Making
In the Swiss “Tumor Profiler” (TuPro) project, the research team is examining the cellular composition and biology of tumors of 240 patients with melanoma, ovarian cancer, and acute myeloid leukemia. Recruitment for the study began in 2018. Today, the melanoma cohort is fully enrolled, and the ovarian cancer and acute myeloid leukemia cohorts are nearing complete enrollment.
“The Tumor Profiler Study is an observational clinical study combining a prospective diagnostic approach to assess the relevance of in-depth tumor profiling to support clinical decision-making (“fast diagnostic loop”) with an exploratory approach to improve the biological understanding of disease (“exploratory science loop”),” the TuPro website states.
The graphic above taken from the Tumor Profiler project paper illustrates how the TuPro study’s workflow entails patient enrollment, sample collection, analysis by different technology platforms, and data integration, creation and discussion of molecular research and summary reports, discussion of treatment options in pre-tumor boards and the final treatment decision in tumor boards. (Photo copyright: Cancer Cell.)
In their published paper, the Swiss researchers say these three cancers were selected for the study “based on the potential clinical benefit and availability of sufficient tumor material for simultaneous analysis across all technologies.”
According to a University Hospital Basel blog post, the TuPro project examination of each cancer tumor goes “much further than the limited use of molecular biological methods” used by leading hospitals. “This results in huge amounts of data per patient, which we process and analyze using data science methods,” stated data scientist Gunnar Rätsch, PhD (above), Professor for BiomedicalInformatics at ETH Zurich and one of the study’s corresponding authors, in the blog post. This research could lead to new precision medicine biomarkers for clinical laboratory cancer diagnostics and therapies. (Photo copyright: ETH Zurich.)
The TuPro Project’s findings are available to doctors who analyze them at interdisciplinary tumor board meetings and generate treatment options, creating a “fast diagnostic loop” with an estimated four-week turnaround time from surgery to tumor board. “This approach has the potential to alter current diagnostics and paves the way for the translation of comprehensive molecular profiling into clinical decision-making,” the study’s authors wrote in Cancer Cell.
Could Oncologists Be Making Better Precision Medicine Decisions?
In its writeup on the TuPro Project’s research, Precision Oncology News concluded that the Swiss study “is rooted in the researchers’ notion that oncologists are not making the best personalized treatment decisions for patients by relying just on targeted DNA profiling using next-generation sequencing and digital pathology-based tests.
“The researchers within the TuPro consortium hypothesized that integrating a more comprehensive suite of omics tests could lead to a more complete understanding of patients’ tumors, including providing insights into the tumor microenvironment, heterogeneity, and ex vivo responses to certain drugs. This, in turn, could help inform the best course of treatment,” Precision Oncology News added.
“With the Tumor Profiler study, we want to show that the widespread use of molecular biological methods in cancer medicine is not only feasible, but also has specific clinical benefits,” said TuPro consortium member Viola Heinzelmann-Schwarz, MD, Head of Gynecological Oncology at University Hospital Basel, in an ET Zurich news release.
New Precision Medicine Biomarkers from TuPro’s Molecular Analysis
Researchers in the study also are investigating whether and what influence the molecular analysis had on doctors’ therapy decisions.
The University Hospital Basal blog notes the long-term benefits of the Tumor Profiler approach is to expand the personalized-medicine therapy options for patients, including determining whether patients would benefit in certain cases “if they were not treated with drugs from standard therapy, but with drugs that have been approved for other types of cancer.”
Anatomic pathologists and clinical laboratory scientists will want to take note of the TuPro project’s ultimate success or failure, since it could usher in changes in cancer treatments and bring about the need for new diagnostic tests for cancer biomarkers.
