Computer scientist wins $2.3 million to develop AI that patients can use to manage health

Reviewed by Kate Anderton, B.Sc. (Editor)Sep 17 2019

Samantha Kleinberg, a computer scientist at Stevens Institute of Technology and pioneer in wearable technologies, has secured roughly $2.3 million in three new grants by the National Science Foundation and the National Institutes of Health to develop artificial intelligence that provides personalized information to patients so they can take an active part in managing their health.

It's usable AI. The amount of available data in health and medicine is overwhelming and not all of it is relevant or important for each individual. We combine computational work and the human elements of decision-making to give patients the information they need to actively decide, with their clinician, what their treatment plan should be."

Samantha Kleinberg, computer scientist, Stevens Institute of Technology

The new grants, which bring Kleinberg's total research funding to $5.4 million, underscore a growing market where artificial intelligence and big data are providing added value to healthcare with a focus on diagnosis, treatment, patient monitoring and prevention.

A brief summary of each grant is below:

Uniting Causal and Mental Models for Shared Decision-Making in Diabetes; $917,879: National Science Foundation

Kleinberg will develop computational tools to personalize information patients with diabetes see in order to better decide on a treatment plan. This research is based on a longstanding model in healthcare known as shared decision-making approach, where a patient and clinician work together to understand the patient's preferences and formulate a treatment plan. However, the efficacy of the model can be strained as each participant in the treatment plan--patient, doctor, other caregivers--holds different set of beliefs about disease and treatment, leading to different approaches that can derail results. Additionally, the amount of information given to patients can be too complex and include unnecessary details--which could lead to patients becoming overwhelmed or not understanding. Kleinberg, in collaboration with Jessecae Marsh, a cognitive scientist at Lehigh University, will also create training modules to educate clinicians about how patient beliefs influence trust and decision-making, as a way to facilitate optimal ways to share decision about treatment. Onur Asan, who heads the Humans-Systems Interaction Lab at Stevens, will serve as co-principal investigator.

Harnessing Patient Generated Data to Identify Causes and Effects of Nutrition during Pregnancy; $864,220: National Institutes of Health

During pregnancy, about 9 percent of women develop gestational diabetes. In collaboration with Andrea Deierlein, a nutritional and reproductive epidemiologist at New York University, Kleinberg will use patient-generated health data collected through wearing activity monitors, logging meals in real time using photos that parses out type of food and calories, or recording symptoms to capture what happens before illness develops. While type 2 diabetes typically develops over a long period of time, gestational diabetes occurs within a short, bracketed time period, making this the ideal population to study. Ultimately, the data aim to identify what factors cause this disease and will be used to identify targets for early interventions and to guide decisions during pregnancy.

Moving Beyond Knowledge to Action: Evaluating and Improving the Utility of Causal Inference; $499,454: National Science Foundation

In an effort to build artificial intelligence, researchers are developing algorithms that determine cause and effect and as such, have focused on developing methods to mine through data to find these causal structures. However, causes have been treated as the end goal, without addressing information that is most useful for patients i.e. what they can and/or are willing to act on to improve their health. Kleinberg wants to build artificial intelligence that's useable for the benefit of patients.

For example, imagine monitoring a person with diabetes on their physical activity, food intake and stress levels. Perhaps all three variables coincide with high glucose or other adverse health outcomes. But which of these variable - activity, food intake, or stress - is actually causing the adverse health outcome? Current artificial intelligence is good at finding these correlations, but not as good as determining a causal structure. Moreover, artificial intelligence cannot determine which of these variables people are willing to - or able to - change in order to improve their heath. "Humans don't just use data like a machine," said Kleinberg. "They make decisions based on many things, including prior experiences and assumptions. So not all new information has the same impact on influencing a person's decisions."

The research will focus on understanding what makes the output of an algorithm useful to human decision-makers so that algorithms can be evaluated based on this ability, rather than the percent of causes they find or how fast algorithms accomplish a task. This will introduce new methods that make causal models more usable and personalized, leveraging these results to improve everyday decisions around treatment, diet and exercise.