Harvard Researchers Develop AI to Predict Colon Cancer Survival

Researchers at Harvard Medical School and National Cheng Kung University in Taiwan have developed an artificial intelligence (AI) model that predicts how aggressive a colon cancer is, the likelihood of survival, and the optimal therapy for each patient.

The MOMA Model

The Multi-omics Multi-cohort Assessment (MOMA) model precisely forecasts the severity of colorectal malignancies, the probability of a patient's survival, and the most effective treatment plan by analyzing images of tumor samples.

By aiding clinicians and patients in managing this elusive disease, which frequently displays varying behavior even among those with comparable disease characteristics who receive identical treatment, this tool can prove to be highly beneficial.

Kun-Hsing Yu, the study co-senior author and assistant professor of biomedical informatics at the Blavatnik Institute at HMS, explained that the model can perform tasks that pathologists cannot do based on image viewing alone.

Yu led an international team of pathologists, oncologists, biomedical informaticians, and computer scientists in developing the tool. He also stated that "We fully expect that this approach will augment the current clinical practice of cancer management."

The researchers acknowledge that multiple factors affect the prognosis of any individual patient, and no model can perfectly predict survival.

However, the new model can help guide clinicians to follow up more closely, consider more aggressive treatments, or recommend clinical trials for experimental therapies if their patients have worse predicted prognoses based on the tool's assessment.

Creating the AI Tool

To create this tool, the researchers utilized data from nearly 2,000 patients with colorectal cancer from diverse national patient cohorts, including over 450,000 participants.

During the training phase, the model was provided with patient information such as age, sex, cancer stage, and outcomes, as well as tumor-specific data like genomic, epigenetic, protein, and metabolic profiles.

Subsequently, the researchers presented the model with pathology images of tumor samples and instructed it to identify visual markers related to tumor types, genetic mutations, epigenetic alterations, disease progression, and patient survival.

The researchers tested its performance on a new set of tumor sample images from different patients to simulate "the real world" conditions.

They compared the model's predictions with the actual patient outcomes and other clinical information. The results were promising as the model accurately predicted the patients' overall survival after diagnosis and the duration of cancer-free years.

Additionally, the model accurately predicted an individual patient's response to different therapies based on the presence of specific genetic mutations in their tumor. This outperformed human pathologists and current AI models in both areas.

The research team plans to use the model in clinics and hospitals after a randomized trial to refine it. With 1 million annual colorectal cancer deaths worldwide, this tool could provide clarity to doctors making prognoses and treatment decisions for patients.