AI System Finds Crucial Clues For Diagnoses In Electronic Health Records todayheadline

In hospitals where seconds matter, physicians often face a data paradox: vast electronic records but little time to extract meaning. Researchers at the Icahn School of Medicine at Mount Sinai have now developed an artificial intelligence system that transforms this flood of information into structured insight. The tool, called InfEHR, interprets how clinical events relate over time, connecting the dots that traditional analytics often miss. Their findings, published in Nature Communications, show that InfEHR can detect rare or ambiguous diseases with remarkable accuracy while knowing when it is uncertain.

InfEHR, short for Inference on Electronic Health Records, converts each patient’s chart into a network of interconnected medical events such as lab tests, vitals, medications, and symptoms. Using deep geometric learning, the AI infers how one event influences another and estimates the probability of hidden conditions. Rather than forcing every patient into the same diagnostic mold, the model adapts its reasoning to individual medical trajectories.

Decoding Uncertainty In Real Time

In tests at Mount Sinai in New York and UC Irvine in California, the researchers trained InfEHR using a small set of doctor-confirmed examples. The AI then analyzed thousands of deidentified patient records to see whether it could flag elusive cases that defy standard rules, such as newborns developing sepsis despite negative blood cultures or adults developing kidney injury after surgery. Both are conditions where time and uncertainty can be deadly.

In neonatal sepsis without positive cultures, InfEHR identified affected infants up to 16 times more effectively than traditional clinical heuristics. For postoperative kidney injury, it flagged at-risk patients up to seven times more effectively. The system was also able to admit when the evidence was insufficient and return “not sure,” a safeguard rarely built into AI models.

“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,” said Girish N. Nadkarni, MD, MPH, senior author and Chair of the Windreich Department of Artificial Intelligence and Human Health at Mount Sinai.

Unlike most diagnostic models that apply the same logic to all patients, InfEHR constructs a personalized network from each patient’s history. It identifies relational patterns, such as combinations of vital sign shifts and medication timing, that may suggest hidden disease processes. The framework automatically learns which temporal patterns matter and how confident it should be in its own conclusions.

From Hunch To Probability

The algorithm’s core strength lies in its probabilistic reasoning. Instead of simply classifying patients as sick or healthy, InfEHR estimates how likely a given disease explains the observed clinical events. This generative approach, the authors note, lets the AI weigh both evidence and uncertainty in the same way physicians do when forming differential diagnoses.

In technical terms, InfEHR learns to represent each patient’s record as a temporal graph, a web where edges encode how events influence one another over time. Deep geometric learning then allows the system to update its understanding as new data arrive, reflecting the evolving reality of clinical care. The result is a model that not only predicts disease but also signals when its confidence is low, guiding doctors to focus their judgment where the machine hesitates.

“Traditional AI asks, ‘Does this patient resemble others with the disease?’ InfEHR takes a different approach: ‘Could this patient’s unique medical trajectory result from an underlying disease process?’” said lead author Justin Kauffman, MS.

The research team envisions broader applications. Because the framework requires little labeled data, it could extend to rare diseases and adapt across hospitals without extensive retraining. The authors are making the code publicly available so other teams can test and refine it. Future work aims to integrate clinical trial data, allowing InfEHR to suggest personalized treatments for patients whose profiles fall outside typical study populations.

For now, InfEHR represents a shift from reactive to inferential medicine, a system that can see what human eyes might miss in a sea of fragmented data. The model’s ability to quantify uncertainty may prove as valuable as its accuracy, turning hesitation itself into a clinical signal.

Nature Communications: 10.1038/s41467-025-63366-6