Comprehensive Summary
This study, presented by Kauffman and colleagues, proposes InfEHR: a framework for the automatic computation of disease likelihoods from electronic health records (EHRs). InfEHR, a deep geometric learning framework which converts EHRs into graphs that generate unbiased probabilities, was developed using unstructured data from 11 million EHRs stored in the Mount Sinai Data Warehouse. The medications prescribed in the EHRs were evaluated, along with clinical techniques used in care. Additionally, neonatal patients without significant categorical information missing in EHRs were identified, leading to a final data set of 8067 individuals. The model aggregates signals and refines them into calibrated disease likelihoods. In testing across two populations, consisting of neonatal culture-negative sepsis and postoperative acute kidney injury, InfEHR is shown to outperform physician predictions due to its maintenance of sensitivity and specificity. Further, InfEHR notes when data is insufficient for certain judgements, and provides guidelines for exhaustive diagnosis in said cases.
Outcomes and Implications
Clinicians frequently must make high-stakes decisions in environments where clinical information is incomplete, diagnostic criteria are unclear, and disease presentations vary widely, especially for conditions that lack definitive tests or occur infrequently. In such settings, traditional heuristics or single biomarkers often fail to provide the precision needed to guide care, resulting in delayed interventions for patients who truly need treatment or unnecessary therapies for those who do not. InfEHR addresses this gap by integrating diverse, time-dependent EHR data to generate individualized disease likelihoods that are both accurate and reflective of uncertainty. This capability reduces risks associated with both under-treatment, such as progression of sepsis, and over-treatment, including avoidable antibiotic exposure in neonates. Moreover, because InfEHR requires only limited labeled data, automatically adjusts to differences in local patient populations, and clearly communicates prediction confidence, it is highly promising for real-world clinical support. This is especially relevant in fast-moving acute care settings where rapid, well-informed decisions are critical and clinician capacity is strained.