Home » News » MIT Professor Leverages Machine Learning to Find Promising Cancer Treatments

MIT Professor Leverages Machine Learning to Find Promising Cancer Treatments

When his father was diagnosed with stage IV, non-operable gastric cancer in 2007, Dr. Dimitris Bertsimas knew that combination chemotherapy was the best course of treatment. He visited several of the leading cancer hospitals in the nation—Dana Farber, Massachusetts General, MD Anderson and Memorial Sloan Kettering—to see what specific therapies they would propose for his father.

“They each told me very distinct therapies, almost with no drugs in common,” says Bertsimas. “I didn’t know how to compare them.”

Dr. Dimitris Bertsimas

Dr. Dimitris Bertsimas

So Bertsimas, who is a professor of operational research at MIT did a simple back-of-the-envelope calculation. He created a graph on which the horizontal axis plotted how toxic a particular drug is and the vertical axis plotted how many months patients who took that drug typically survive. “A dot in that two-dimensional plot tells you how much toxicity and how much survivability a drug has,” he explains. “What you want is left and high—many months of survival and low toxicity.”

Bertsimas advised his father to pursue treatment at MGH; he lived for 24 months, 3 times longer than the estimated 8-month survival rate for patients with similarly advanced gastric cancer.

Bertsimas realized that his rough calculations could be systematized to help other patients like his father. He returned to MIT where he and his students developed a database that pulled data about the survivability and toxicity associated with certain cancer drugs from about 550 papers recording clinical trial results for gastric and gastro-esophageal cancers. Armed with information about the past efficacy of certain drugs, the database allowed Bertsimas and his team to advise patients and doctors about the relative benefits of one treatment versus another treatment.

The team also began to develop a machine-learning model to predict the toxicity and survivability of new combinations of drugs that have previously not been tried together. “Today people decide trials by intuition,” explains Bertsimas. “It’s humanly impossible to know the outcomes of hundreds and hundreds of trials, historically.” Machine learning and analysis provided with Bertsimas a systematic method of finding promising treatments and understanding what the tradeoffs are between various treatments.

Clinical trials for combination chemotherapy drugs cost anywhere between $10 million and $30 million—by helping research institutions only spend money on the most promising trials, Bertimas’ model could potentially help them avoid significant costs. Bertsimas is in discussions to complete a clinical trial in breast cancer at New York Presbyterian and gastric cancer at Dana Farber, and has future plans to optimize the algorithm to consider personalized data and target therapies for particular demographics.

Similar posts

Leave a Reply

Your email address will not be published. Required fields are marked *

Follow us!

Send this to a friend