Predicting Patient Outcomes Via Machine Learning Models

Collaborative research to test method in opioid use disorder 

STONY BROOK, NY, October 3, 2024 –Two Stony Brook University researchers are developing a way to use machine learning models to predict patient outcomes. The collaboration involves Richard N. Rosenthal, MD, a Professor in the Department of Psychiatry and Behavioral Health in the Renaissance School of Medicine (RSOM), and Fusheng Wang, PhD, a Professor in the Departments of Biomedical Informatics and Computer Science in the RSOM and College of Engineering and Applied Sciences (CEAS). Their work specially centers on optimizing prediction of risk related to opioid use disorder and opioid overdose.

The research is supported by a $1.05 million grant from the Patient-Centered Outcomes Research Institute (PCORI), an independent funding organization that supports patient-centered comparative clinical effectiveness research in the U.S. Wang is the leading Principal Investigator of the PCORI grant. His research on using machine learning to predict patient risk was the basis for the award, developed around work he spearheaded since 2017.

Wang focuses on creating models to predict how likely patients are to develop opioid use disorder and opioid overdose. Every individual patient varies in opioid risk. Using a prediction model, Wang and Rosenthal seek to develop a more useful machine learning tool for clinicians to foresee patient risk and then plot their course of treatment for each patient. Key to the process is pulling data from patient’s medical records to make the prediction.

“I think probably the most important contribution in this type of model is our stakeholder-in-the-loop approach,” explains PI Wang. “Stakeholders, including clinicians and patients, will participate in the full cycle of model design, development and evaluation. I think for the health care domain, that’s really something that is missing. if we can provide a framework with this particular model, the lessons learned can be very useful for others to adopt a similar methodology.”

He adds that with this revolutionary “stakeholder in the loop approach,” clinicians can provide feedback to the prediction model to make the output both more accurate and more clinically usable, a method that makes the machine learning model more patient-centric.

“Most AI model development in health care is done by the developers so that there is little if any feedback into the process by the end users, such as clinicians,” says Rosenthal. “As a result, because of how uncurated machine learning works, the doctors are frequently left with non-intuitive models that they can’t explain to patients for making treatment decisions, so most models are underutilized.”

One of the challenges that Wang, Rosenthal and their colleagues have to overcome is that patient data is very complex, with lots of clinical variables. What variables contribute to the prediction of the risk are unclear. This is another reason why the stakeholder-in-the-loop approach is important because it gives clinicians the chance to add in clinical knowledge.

“A doctor wants to know all the information as quickly as possible, as comprehensive as possible,” emphasizes Wang. “If the machine learning model generates a prediction, then we need to really have a good precise summary about the patient, why the patient is predicted with such a risk.”

The project brings in patient partners, clinicians, computer scientists, researchers and community representatives from the New York State Office of Mental Health and Suffolk County Department of Health to collaborate together.

The team hopes to expand this research method for other diseases, such as cardiovascular diseases. They also hope to implement the model in a clinical setting, such as the emergency department, in order to test its accuracy and effectiveness.