By Maxine Bookbinder

July 21, 2016 The urgency to find treatment for his ailing father inspired an MIT professor to develop a method to successfully predict cancer therapy outcomes, replacing intuition with data, potentially saving millions of dollars, and extending patients’ lives up to six months.

Professor Dimitris Bertsimas, Sloan School and Operations Research Center at MIT, and his team have developed models to accurately predict the outcomes of clinical trials testing combination chemotherapies for gastric and esophageal cancers before they are run. The findings indicate that the MIT models might improve the efficacy of the regimens selected for testing in Phase III clinical trials without increasing toxicity. Gastric cancer is the third leading cause of cancer deaths globally but has no standard chemotherapy regimen.

Generally clinical trials evaluate novel combinations of previously-used drugs without evaluating new drugs. But Bertsimas wanted to design a computational method to evaluate chemotherapy regimens for Phase II and Phase III trials. He says this is the first study to use statistical models to predict clinical trial outcomes of arbitrary drug combinations and to perform out-of-sample evaluations of the predictions.

In 2007, the professor’s father was diagnosed with inoperable non-metastatic gastric cancer. Bertsimas researched therapies used at five major U.S. hospitals; each used different therapies and combinations. “I did a simple calculation by reading papers reviewing clinical trials from the hospitals. Then I created a graph with a horizontal axis showing drug toxicity and a vertical axis for survival rates.” His father lived two years, twice the expected outcome for his cancer.

This personal study inspired Bertsimas to research further. “If I had more than a month, what could I do?” He researched data from clinical trials for gastric cancer from 1980 to today and developed a program to study outcomes of different drugs, drug combinations, and resulting toxicity and survival rates. Then, he used this data to predict outcomes from other clinical trials.

He and his team extracted data from 414 published journal articles describing the treatment methods and patient outcomes for a total of 495 treatment arms of gastric cancer clinical trials.

Analytics Formula

The study has three parts; Bertsimas calls part one Descriptive Analytics, which electronically maps the outcomes, based on survivability, toxicity, and cost, of a multitude of drug combinations and subsequent predictions. “We took data from a universe of patients – how sick they are, what drugs used, schedules, outcomes, toxicities and so forth and how long they survived. This data provides information to doctors and patients about which therapies work best for specific cancers and demographics, as well as trade-offs when using certain medications.”

Part two is Predictive Analytics, in which the team builds statistical models based on trials. “By observing the outcomes of various combinations of drugs, we see footprints of the individual drugs,” says Bertsimas. “Using machine learning methods, we build statistical models of outcomes as a function of the drugs used and their dosages. This way, each trial is a formula for us; we can make predictions of things we have not seen.” He adds that his team accurately predicts what specific drugs will do in certain combinations, allowing the researchers to inform doctors and patients about which therapies are stronger than others. Bertsimas describes a setting in which, for example, drugs A and B are used together, and drugs C and D are used together, but A and C are not. He then runs statistical models on all the combinations. While the team cannot make individual recommendations, it can map predictions for specific demographics.

“We know this works,” says Bertsimas. “We took data from 1980 – 2005 and predicted what their methods would do in 2006 without seeing those results. We then continued from 1980-2006 to predict the outcomes for trials in 2007 and so on. The predictions were very accurate.”

Part Three is Prescriptive Analytics, data that Bertsimas and his team use to find promising – as well as unpromising -- trials. “The 10% predicted to be lowest in terms of survival result in only 1% that result in above median survival. Say we make predictions among 100 combinations for average length of survival. Take the 10 that resulted in the 10 lowest predictions. Only one of these 10 would result in actual survival times among the top 50. In other words, by avoiding these trials, we gain significant cost savings without sacrificing significant benefits. We know which trials won’t work and will work. We don’t need to rely on intuition.”

The team also designs trials centered on maximizing survival while limiting toxicity levels. The resulting combination of drugs adds about 4-6 months of life for gastric cancer, which has an average survival rate of about 10 months. “We are not going to cure cancer but we can extend survival rates, lower toxicity levels, and offer a better quality of life,” says Bertsimas.

The MIT researchers are attacking other cancers and illnesses; they recently completed statistical data for breast cancer and are working with American Society of Clinical Oncology to publish the results for gastric cancer and breast cancer to allow public access to its findings.

The process is labor-intensive. Undergraduate and graduate students scrutinize “the most-cited papers in the world for these diseases” to extract data, develop algorithms, predict outcomes, and create a database of these findings. Bertsimas’ team read 1,500 papers for breast cancer and 400 for gastric cancer.

The team is now examining thousands of clinical papers to create a database for lung and colon cancer. Bertsimas is also collaborating with doctors at New York City’s Presbyterian Hospital for breast cancer.

In addition, the professor is expanding his research; he is working on personalized diabetes management based on data from 1.5 million people and could expand this to cancer, stroke, and high blood pressure, among other diseases.