Researchers at the Icahn School of Medicine at Mount Sinai have developed NutriSighT, an interpretable transformer-based AI model that dynamically predicts which mechanically ventilated ICU patients are at risk of underfeeding (receiving <70% of daily caloric needs) during days 3–7 of their stay. Updating every four hours with routine clinical data, the tool enables early, personalized nutrition interventions to potentially improve outcomes in critically ill patients.
Glimpse:
Published December 17, 2025, in Nature Communications, the study trained NutriSighT on data from 3,284 Dutch ICU patients and validated it externally on 6,456 U.S. patients. It highlights underfeeding prevalence (41-53% by day 3, 25-35% by day 7) and uses factors like vital signs, labs, medications, and sedation to forecast risks hours ahead. As an early warning system not a replacement for clinical judgment the model supports timely adjustments, with future prospective trials planned for real-world impact.
In a major advancement for critical care nutrition, researchers from the Icahn School of Medicine at Mount Sinai have introduced NutriSighT an innovative, interpretable AI model designed to forecast underfeeding risks in mechanically ventilated ICU patients. The tool addresses a persistent challenge: many critically ill patients on ventilators receive inadequate nutrition during the crucial first week, when needs fluctuate rapidly due to evolving clinical conditions.
Underfeeding remains common, with 41% to 53% of patients underfed by day three and 25% to 35% still affected by day seven, potentially delaying recovery, increasing complications, and worsening outcomes. NutriSighT tackles this by analyzing routine ICU data vital signs, lab results (e.g., sodium levels), medications, sedation status, and feeding patterns to predict risks hours in advance, with predictions dynamically updated every four hours as patient status changes.
Developed as a transformer model with learnable positional encodings, NutriSighT outperforms traditional methods like XGBoost in retrospective testing. Trained on de-identified data from AmsterdamUMCdb (3,284 patients) and externally validated on MIMIC-IV (6,456 patients), it focuses on adults ventilated for at least 72 hours. The model’s interpretability is a key strength: it highlights specific contributing factors, empowering clinicians to make targeted adjustments without overriding human expertise.
This early-warning capability could enable nutrition teams to personalize enteral feeding strategies sooner, potentially reducing gaps in care and supporting better recovery for some of the most vulnerable ICU patients.
The research, led by co-senior authors Girish N. Nadkarni, MD, MPH (Chief AI Officer at Mount Sinai Health System and Chair of the Windreich Department of Artificial Intelligence and Human Health) and Ankit Sakhuja, MBBS, MS, builds on Mount Sinai’s leadership in AI-driven healthcare. It follows prior successes like NutriScan (a malnutrition screening tool that won the Hearst Health Prize in 2024).
Next steps include prospective multi-site trials to evaluate real-time impact on patient outcomes, integration into electronic health records, and careful design to prevent alert fatigue. The study underscores AI’s role in shifting ICU nutrition from reactive to proactive and individualized.
“The significance of our study’s findings is that, for the first time, it may be possible to identify which patients are at risk of underfeeding early in their ICU stay and tailor care to their individual needs.”
By
HB Team
