Researchers at the Icahn School of Medicine at Mount Sinai have developed an AI-based tool that could help clinicians identify critically ill ICU patients at risk of underfeeding while on mechanical ventilation. The study points to a data-driven approach to improving nutritional care during one of the most vulnerable phases of intensive care.
Adequate nutrition in the first week of ventilation is crucial, as patients’ metabolic needs can change rapidly. Yet underfeeding remains common. “Too many ventilated ICU patients fail to receive sufficient nutrition during this critical period,” says Ankit Sakhuja, Associate Professor of Artificial Intelligence and Human Health at Mount Sinai. “We wanted a timely, practical way to identify who is most at risk, so care teams can intervene earlier.”
Early risk prediction
The research team developed NutriSighT, an interpretable AI model that analyzes routinely collected ICU data, including vital signs, laboratory values, medications and feeding records. The system generates updated risk predictions every four hours, forecasting which patients are likely to be underfed on days three to seven of ventilation.
NutriSighT was trained and validated using large, deidentified ICU datasets from both Europe and the United States, demonstrating robustness across different care settings. The model does not operate as a black box: it highlights which factors, such as blood pressure trends, sodium levels or sedation, contribute most to underfeeding risk.
Widespread and persistent problem
The study, published in Nature Communications, confirms that underfeeding is common in early ICU care. Between 41% and 53% of ventilated patients were underfed by day three, with 25% to 35% still underfed by day seven. These findings underline the clinical relevance of timely nutritional risk assessment.
By providing early warnings, the AI tool could support more personalized feeding strategies, help nutrition teams prioritize patients, and inform future clinical trials aimed at optimizing nutrition in critical care.
The researchers emphasize that NutriSighT is designed as a decision-support tool, not a replacement for clinical judgment. “The goal is to give clinicians better information at the right moment,” says Girish N. Nadkarni, Chair of the Windreich Department of Artificial Intelligence and Human Health at Mount Sinai and Chief AI Officer of the Mount Sinai Health System. “Ultimately, it’s about delivering the right nutrition to the right patient at the right time.”
Toward implementation
Future work will focus on prospective, multi-center trials to determine whether acting on NutriSighT’s predictions improves patient outcomes. The team also plans careful integration into electronic health record systems and expansion toward broader, individualized nutrition targets.
If successful, the approach could mark an important step toward more personalized, data-driven ICU care, using AI to close a long-standing gap in nutritional support for critically ill patients.
