New AI Tool May Help Predict Best Treatments for CRC

Researchers have developed an artificial intelligence machine-learning platform that can predict the prognosis and likely treatment response of patients with colorectal cancer (CRC) using histopathology images, according to a new study published in Nature Communications.

Specifically, the tool can aid doctors in identifying a “molecular diagnosis” based on a patient’s tumor and cancer characteristics, Kun-Hsing Yu, MD, PhD, the study’s senior author and an assistant professor of biomedical informatics at Harvard Medical School in Boston, Massachusetts, told Medscape Medical News.

The Multi-omics Multi-cohort Assessment (MOMA) “successfully identified indicators of how aggressive a tumor was and how likely it was to behave in response to a particular treatment,” as well as patients’ overall and disease-free survival, noted Harvard Medical School in a press release. “Based on an image alone, the model also pinpointed characteristics associated with the presence or absence of specific genetic mutations — something that typically requires genomic sequencing of the tumor.”

The researchers designed the tool to offer “transparent reasoning,” so that if a clinician asks it why it made a certain prediction, it would be able to explain its reasoning and the variables it used, the press release noted.

“We first allow AI to explore any correlation, and then we try to explain those correlations using existing pathology terms that experts will be able to understand,” Yu told Medscape Medical News.

Although the tool is freely available to clinicians and researchers, it’s not yet ready for clinical use. When it is, the tool has the potential to provide timely, accurate decision support based on tumor imaging.

Colorectal cancer is the second most common cause of death from cancer in the United States, with more than 53,000 deaths each year, and the patient population has been gradually skewing younger over the past two decades.

Although clinicians already use histopathology and genetic analysis to guide treatment, the process can take several days or weeks in some areas, and these services may not be available in all parts of the world.

“Currently, a clinician has to send a [tissue] sample from the tumor specimen to genomic sequencing labs and wait for a week, sometimes up to three or more weeks, to get genomic sequencing results,” Yu said. That means a patient’s anxiety grows as they wait to find out which treatments might benefit them or how they might respond to a particular treatment, he said.

Additionally, current knowledge for predicting patient survival, beyond considering the patient’s cancer stage, age, and general health status, is limited, Yu said.

Predictive Ability

The MOMA platform was trained on information from 1888 patients with colorectal cancer from three national cohorts: 628 patients from The Cancer Genome Atlas (TCGA) program, 927 patients from the Nurses’ Health Study with Health Professionals Follow-Up Study (NHS-HPFS), and 333 patients from the Prostate, Lung, Colorectal and Ovarian (PLCO) Cancer Screening Trial.

During the training, they fed the model information about the patients’ age, sex, cancer stage, and outcomes, as well as their tumors’ “multi-omic” information: the cancers’ genomic, epigenetic, protein, and metabolic profiles. Researchers showed the AI model digital, whole-slide histopathology images of tumor samples and asked it to look for visual markers related to tumor types, genetic mutations, epigenetic alterations, disease progression, and patient survival with the goal of enabling the platform to detect patterns that are undiscernible to the human eye.

They then tested the MOMA platform’s ability to interpret images by feeding it new tumor sample images from different patients and asking it to predict their survival and progression-free survival.

Researchers found that the tool successfully identified overall survival outcomes in patients with stage I or II cancer in the TCGA cohort, which they further validated with the NHS-HPFS and PLCO cohorts. The platform revealed that “dense clusters of adenocarcinoma cells are highly indicative of worse overall survival outcomes” and that the interaction of cancer cells with smooth muscle cells in cancerous areas predicted poorer overall survival.

MOMA was slightly more effective in predicting progression-free survival for stage I and stage II colorectal cancer across all three cohorts.

“Compared with the overall survival prediction, our progression-free survival model puts more emphasis on infiltrating lymphocytes and regions associated with extracellular mucin in its prediction,” the authors note.

Prediction of overall survival and progression-free survival for stage III colorectal cancer showed similar levels of accuracy, they note.

The tool also successfully assessed patients’ likely response to immunotherapy using predictions of microsatellite instability (MSI), since high MSI indicates a better response to immune checkpoint inhibitors.

MOMA outperformed a different machine-learning algorithm in predicting the copy number alterations (CNAs) and other features related to cancer development, and it predicted the likelihood of a BRAF mutation, which is linked to poorer prognosis.

Pushing the Envelope?

MOMA presents an “intriguing new avenue of adding to how we think about and assess someone who has cancer,” Stacey Cohen, MD, an associate professor in the clinical research division of Fred Hutchinson Cancer Center at the University of Washington Medicine in Seattle, told Medscape Medical News.

However, the tool as it’s currently described appears primarily to duplicate what clinicians already are doing, which is considering a wide range of factors — including pathologic features, patient features and demographics, and the patient’s other medical illnesses — to develop a treatment plan within the context of current guidelines, noted Cohen, who was not involved in the project.

“I’m looking for these types of models to not just prognosticate an outcome but to really predict how someone should be treated, and to do that better than [using] standard clinical features,” Cohen said. “To some degree, they’re taking this AI model and trying to catch up to what we’re currently doing. Clearly, if they could do that, they can then push the envelope.”

Cohen acknowledged that a strength of using an AI platform is the speed at which it can provide its predictions in areas with few medical resources and few healthcare professionals — as long as the necessary imaging is available and physicians have a way to use the platform.

“On the one hand, I do see this as an opportunity to share the wealth of knowledge in a more rapid fashion, but I don’t think anybody is going to let a computer program dictate their treatment without a human medical oncologist being able to interpret that information,” Cohen said. “It still will require a lot of education by the users and not just by the people who are designing the study.”

Although the MOMA platform looked at multiple pathologic features in multiple cohorts, the results remain limited by the fact that the patients in those cohorts were treated decades ago, before many current treatments may have been available, Cohen said.

She also added that the cohorts did not have much ethnic diversity. In the NHS-HPFS, the largest cohort, 57% of the patients were White, and researchers lacked data on race for 42% of patients, so only about 1% of participants were of a known non-White race. Similarly, 47% of the TCGA patients were White and 41% had no data on race, leaving only 12% of patients from known, non-White racial backgrounds, including 10% Black or African American.

Additional studies that focus on specific patient populations are needed to evaluate the model’s applicability in clinical settings, the investigators note. More research is required to “identify the optimal prognostic prediction methods and enable personalized treatments and advance care planning,” they add.

These are the early days for this type of technology, Cohen noted.

“I’m very excited to see how this technology develops and how it could be potentially additive or improve upon our current treatment planning for patients,” she said.

Yu developed the invention “Quantitative Pathology Analysis and Diagnosis using Neural Networks,” whose patent is held by Harvard University, and has consulted for Curatio. Co-author Kana Wu, PhD, is a stakeholder and employee of Vertex Pharmaceuticals. The study’s funding sources included the National Institute of General Medical Sciences, the Google Research Scholar Award, the Blavatnik Center for Computational Biomedicine Award, the National Science and Technology Council Taiwan, and the National Center for High-performance Computing Taiwan. Cohen has advised or consulted for Natera Inc.

Nature Communications. Published online April 13, 2023. Full text

Tara Haelle is a health/science journalist based in Dallas. Follow her at @tarahaelle

For more news, follow Medscape on Facebook, Twitter, Instagram, YouTube, and LinkedIn

Source: Read Full Article