Mount Sinai: AI That Asks Its Own Questions Could Transform Clinical Diagnostics

Artificial intelligence in health care is often discussed in terms of automation and pattern recognition, but a new system developed at the Icahn School of Medicine at Mount Sinai signals a more profound shift: AI that can tailor its diagnostic reasoning to individual patients and recognize when it lacks enough information to proceed. The system, called InfEHR, challenges traditional models of clinical support by operating not just as a predictor, but as a dynamic inference engine.

Published in Nature Communications, the Mount Sinai study introduces a framework that goes beyond matching patients to known diagnostic clusters. Instead, InfEHR constructs a personalized network from a patient’s full medical history, linking disparate clinical events over time into meaningful patterns. This allows it to not only spot rare conditions but also flag uncertain cases where traditional AI might confidently deliver a wrong answer.

From Pattern Matching to Diagnostic Inference

What distinguishes InfEHR from other clinical AI systems is its foundational approach. Traditional algorithms often ask: “Does this patient resemble others with this disease?” InfEHR instead explores: “Could this patient’s unique trajectory result from a specific disease process?” This orientation toward causality over correlation redefines what AI can contribute to clinical diagnostics.

By transforming longitudinal electronic health records into relational networks, InfEHR identifies meaningful connections among symptoms, tests, treatments, and time points. These networks are then used to train geometric learning models that can detect underlying diseases, even in the absence of typical biomarkers. This approach proved especially potent in two areas: neonatal sepsis without positive cultures and post-surgical kidney injury, both conditions where current diagnostics are notoriously unreliable.

Clinical Performance Without Deep Data Dependence

One of the most significant findings from the Mount Sinai study is that InfEHR achieves high diagnostic accuracy with limited labeled data. Rather than relying on massive training datasets, the system calibrates itself with a small number of clinician-validated examples, then generalizes across populations and hospital settings.

In neonatal sepsis, InfEHR outperformed existing clinical rules by a factor of 12 to 16 in identifying affected newborns. For post-operative kidney injury, it was 4 to 7 times more effective. These gains were achieved without compromising safety: InfEHR was explicitly designed to recognize data insufficiency and respond with uncertainty, a rare and crucial feature for clinical AI tools.

This design choice addresses a key gap in current machine learning deployments, which often lack the ability to express diagnostic humility. In real-world care, the ability to acknowledge uncertainty can prevent harmful decisions, reduce clinician overreliance, and reinforce trust in AI-assisted support.

System-Level Implications for Health Leaders

For CMIOs, CIOs, and enterprise data leaders, InfEHR represents more than a novel algorithm. It challenges the current procurement and implementation frameworks that prioritize scalability and vendor maturity over contextual performance and interpretability.

The system also demonstrates how academic health systems can lead in defining new AI paradigms. By focusing on inference and transparency, Mount Sinai is reframing what “clinical-grade” AI should look like. This model prioritizes safety, patient specificity, and deployment agility over universal applicability or brute-force prediction.

Health systems evaluating AI partners or building in-house capabilities should note three core design principles from InfEHR:

• Contextual inference over cohort resemblance. Diagnostic AI must adapt to individual patient journeys, not just aggregate similarity.
• Embedded uncertainty as a safety feature. Tools must flag when data are insufficient, not just when confidence is high.
• Flexible learning with minimal labeling. Systems should generalize well with sparse examples, enabling faster validation and cross-site relevance.

From Diagnostic Insight to Treatment Relevance

Mount Sinai researchers plan to extend InfEHR’s capabilities into therapeutic decision support. By integrating data from clinical trials, the system could help clinicians assess whether published findings apply to their specific patients, particularly those underrepresented in research cohorts.

This capability could address one of health care’s most enduring challenges: translating controlled trial results into messy, real-world care. For example, a treatment proven effective in middle-aged adults may or may not benefit an elderly patient with multiple comorbidities. InfEHR’s probabilistic reasoning offers a path toward bridging that gap.

Ultimately, InfEHR invites a rethinking of how clinical AI systems are conceptualized, evaluated, and governed. It moves the conversation beyond static performance metrics and into dynamic, patient-specific reasoning. For health care leaders seeking more than just automation, this represents a pivotal evolution, from prediction to understanding.