Why Mount Sinai Needed to Create an EHR-Mining AI Tool

The new tool is designed to comb through individual patient data in the EHR, finding clinical care connections that doctors might not see.

Healthcare leaders are hoping AI can do the heavy lifting in analyzing data to improve clinical care. They may have an AI tool that can do just that in the EHR.

A research team led by the Icahn School of Medicine at Mount Sinai has developed AI technology that’s designed to mine the EHR for hidden patterns that can be turned into patient-specific diagnostic insights. The tool, called infEHR, finds connections in data stored in the medical record that might otherwise go unnoticed.

"We were intrigued by how often the system rediscovered patterns that clinicians suspected but couldn't act on because the evidence wasn't fully established," Girish N. Nadkarni, MD, MPH, Mount Sinai’s Chief AI Officer, said in a press release issued by the health system. "By quantifying those intuitions, InfEHR gives us a way to validate what was previously just a hunch and opens the door to entirely new discoveries."  

The tool’s value to clinicians lies in the ability to analyze a patient’s medical record as if it were a puzzle, looking through clinical visits, labs and tests, medications and biometric data to find connections. It was tested on small groups of deidentified data from Mount Sinai and UCI Health in California.

“Traditional AI asks, ‘Does this patient resemble others with the disease?’ “ Justin Kauffman, MS, Senior Data Scientist at the Windreich Department of Artificial Intelligence and Human Health at the Icahn School of Medicine, said in the press release. “InfEHR takes a different approach: ‘Could this patient’s unique medical trajectory result from an underlying disease process?’ It’s the difference between simply matching patterns and uncovering causation.”

Kaufman is the lead author and Nadkami, who is also chair of the Windreich Department of Artificial Intelligence and Human Health, director of the Hasso Plattner Institute for Digital Health, the Irene and Dr. Arthur M. Fishberg Professor of Medicine at the Icahn School of Medicine at Mount Sinai, is the senior corresponding author for the research team, which recently published its work in Nature Communications.

The study notes that infEHR “applies deep geometric learning through a procedure that converts whole electronic health records to temporal graphs that naturally capture phenotypic dynamics, leading to unbiased representations.”

In testing the AI tool on datasets from Mount Sinai and UCI Health, researchers checked to see if it could correctly identify newborns who develop sepsis despite negative blood cultures or patients who develop a kidney injury after surgery. The results were compared against current clinical standards.

According to the study, infEHR was 12-16 times more likely to identify newborns with neonatal sepsis without positive blood cultures and 4-7 times more effective in flagging patients at risk of postoperative kidney injury.

The research team also noted that infEHR states how confident it is in those results, and will identify when it can’t come to a supported conclusion.

“Our framework automatically learns phenotypic temporal dynamics from EHRs through a graph neural net-based (GNN) approach,” the study notes. “We learn high-dimensional representations of such graphs to compute the likelihood of an underlying latent disease given the representation (and the EHR it was derived from). The resulting likelihoods have properties important to any clinical test: the probability clearly identifies when the model is uncertain, confidence scales with accuracy, and the pretest probabilities are substantially revised. We express these characteristics quantitatively through high rule-in and high rule-out potentials benchmarked against real-world clinical heuristics.”

Nadkami and his team are making the infEHR coding available to other researchers as they continue to work out the details of the technology. They’re particularly interested in how it might be used for patients in clinical trials.

“Clinical trials often focus on specific populations, while doctors care for every patient,” Kauffman said in the press release. “Our probabilistic approach helps bridge that gap, making it easier for clinicians to see which research findings truly apply to the patient in front of them.”