Using AI, Startup Mines Health Care Data to Improve Patient Care
Imagine reading six million pages of text or scrolling through 200,000 photos on your smart phone. That is the amount of data, on average, that a doctor should comb through to diagnose and treat a single patient. Most of this information is contained in clinical text, which is difficult to analyze by conventional data analysis methods because it almost always contains errors and is riddled with missing information. As a result, critical information often remains undetected, which can lead to unfortunate consequences. One is adverse drug events, the fourth-leading cause of death in the United States and estimated to cost close to $140 billion per year.
A Columbia startup, Droice, is using artificial intelligence to help doctors make sense of clinical data. "Artificial intelligence is critical to the future of medicine," said Harshit Saxena, a Droice cofounder who just graduated with a master's degree in computer science. He and three other Columbia graduate students—Mayur Saxena (no relation to Harshit), Tasha Nagamine from the Electrical Engineering department, and Aleksandr Makarov—created a platform to assist doctors in choosing an appropriate treatment, whether it is drugs, medical devices, surgeries, lab tests, or other interventions.
Their platform analyzes possible treatment combinations for severalcommon conditions, including diabetes, asthma, and heart disease, which together are responsible for 80 percent of all U.S. prescriptions.
Choosing an appropriate treatment is incredibly complicated. “Every patient is unique,” Harshit Saxena said. “Every treatment should be tailored to the individual sitting in front of the doctor.” It is not easy. More than 2,000 drugs are prescribed in the U.S.; asthma alone has more than 100 FDA-approved drugs. Each drug performs differently depending on a patient’s age, sex, genetic profile, and co-conditions that include diseases, susceptibilities, and allergies. Doctors must also consider a multitude of relevant scientific literature, clinical trials, and FDA guidelines.
Nagamine, who researches machine learning for her PhD in Prof. Nima Mesgarani's lab, explains that traditional methods for choosing a treatment plan can be difficult to use in such a complex setting. “The answer lies in artificial intelligence—more specifically, in advanced natural language processing, which can analyze huge amounts of patient information in the form of text from electronic health records. This gives us critical insights into the best treatment strategy for a patient, which might otherwise be impossible for a doctor to find,” she said.
Droice’s software predicts the performance of a treatment by combining medical research with an analysis of how that treatment has performed on millions of patients in the past. From this collective information, it predicts how the same treatment will perform on a new patient. It is a sophisticated use of artificial intelligence, but for the Droice team, it wasn’t enough to build the software; to truly assist doctors, the predictions had to be delivered in a form that doctors could easily interpret and trust while not disrupting their workflow.
Over several months, the Droice team sought feedback in designing an intuitive interface—interviewing doctors, entering hackathons, and taking advantage of the advice and support provided by alumni and other Columbia connections. Based on this feedback, they embedded Droice intelligence in the software that doctors use to view and update patients’ electronic health records. With a single glance, doctors can see predictions for each drug treatment, backed by trusted scientific papers. “If our software makes predictions for 10 treatment options, we point to the relevant papers in each case so doctors can understand the reason why,” Harshit Saxena said.
In an indication of the interest surrounding the use of artificial intelligence in health care, two hospital-wide rollouts are already in progress, with 576 doctors currently using Droice software. As results come in, the team will use this additional data to refine the software. A paper to make their results public is planned for mid-2017.
“By using artificial intelligence techniques,” Mayur Saxena said, “we are able to help doctors cope with difficult medical decisions and deliver a positive impact on the health and well-being of patients.”