Grant to Fund Artificial Intelligence StudyPosted: Updated:
Two Indiana University scientists have received a nearly $700,000 grant from the National Science Foundation to develop artificial intelligence tools for use in medical decision-making. Lead investigator Kris Hauser says the technology can detect patterns in large amounts of electronic health records.
January 22, 2014
Bloomington, Ind. -- Two computer scientists from Indiana University Bloomington's School of Informatics and Computing and a medical doctor at the Regenstrief Institute on the IU School of Medicine campus have been awarded over $686,000 to further develop the use of artificial intelligence as a tool in medical decision-making.
The National Science Foundation grant will allow lead investigator Kris Hauser, fellow assistant professor of computer science Sriraam Natarajan and Regenstrief investigator Dr. Shaun Grannis to develop and test prototype decision support systems in three clinical settings: cardiology, clinical depression and emergency room re-admission.
"Electronic health records contain a treasure trove of data, but they are simply too large for humans to process," Hauser said. "But AI can detect patterns matching a patient's disease progression and recommend up-to-date, cost-effective treatment plans to a human doctor. The goal of this project is to build a disease-agnostic prototype decision support system that integrates evidence-driven decision-making and the advice of human domain experts to provide well-informed recommendations."
Clinical partners in the prototype testing will include Centerstone Research Institute, the research arm of Centerstone, a behavioral health service provider in Indiana and Tennessee, Wisconsin-based Marshfield Clinic, South Bend (Ind.) Memorial Hospital and Wake Forest School of Medicine.
The team is seeking to use statistical relational learning techniques to mine for patterns in large electronic health care databases, and then input those patterns into a mathematical framework. That framework, called a partially observable Markov decision process, uses documented observations along with observation probabilities to maintain a probability distribution over a set of possible treatment sequences. Through exhaustive searches, the decision support system then identifies the optimal treatment sequences.
Past research has shown statistical relational learning techniques are the best machine learning tools for predicting cardiac arrest based on demographic and lifestyle data. In other work conducted on a clinical depression dataset, treatment plans generated by partially observable Markov decision processes have been shown to outperform existing fee-for-service methods by reducing costs 50 percent and improving outcomes by 40 percent.
The team intends to combine the two machine learning tools by using statistical relational learning techniques to develop a disease progression model used by a partially observable Markov decision process. The goal is to achieve further improvements in recommendation quality and computational scalability for complex treatments.
"Recent studies have demonstrated the immense value leveraging health care data can have on providers' ability to offer more efficient and effective care," said Grannis, who is also an associate professor of family medicine. "This grant will allow us to combine two proven AI-based clinical decision support approaches to create a more powerful, more accurate tool."
Hauser added that the project aims to develop intelligent clinical decision support techniques for recommending optimal action plans -- including both diagnostic tests and medical interventions -- for treating chronic disease; performing multi-step and adaptive treatments; and modifying long-term health habits.
Evidence continues to build showing that using artificial intelligence for clinical decision support can both reduce health care costs and improve patient outcomes. A successful prototype integrating evidence-driven decision-making could also shed light into policy questions surrounding U.S. health care costs and practices.
"Patients, doctors, hospitals, insurers and policymakers coexist in a complex and often dysfunctional web of hidden information, costs and objectives, but they also face the problem of information overload," Hauser added. "AI systems could digest relevant information and put it all on the table, ultimately making health care more transparent and cost-effective."
Additional members of the research team include Regenstrief's Dr. Jon Duke, Centerstone Research Institute's Casey Bennett, Dr. Raman Mitra at South Bend Memorial Hospital, Dr. Michael Caldwell at Marshfield Clinic and Dr. Jeffrey Carr at Wake Forest.
Source: Indiana University