Deep Learning Model for Continuous ICU Acuity Scoring

Real-time Dynamic Risk Assessments to Improve ICU patient Outcomes and Support Earlier Clinical Interventions

This deep learning model provides a continuous acuity score for critically ill patients in the intensive care unit (ICU), enabling earlier clinical interventions and improving patient outcomes. Patients experiencing a medical event may have a life-threatening condition or propensity to develop one at any moment. Accurate, timely assessment of patient acuity is essential in the ICU, where rapid changes in clinical status can have life-threatening consequences. Early recognition of evolving illness severity in critically ill patients is invaluable, helping identify patients in need of life-saving interventions and informing shared decision-making process among patients, providers, and families about goals of care and optimal resource utilization. Traditional scoring systems like SOFA rely on static, intermittent calculations and manual data entry, often missing subtle or sudden deterioration and delaying medical response. These limitations can lead to missed opportunities for early intervention and increased burden on clinical staff, who must synthesize large volumes of physiologic data while managing complex care. The demand for advanced patient monitoring solutions is rising, with the global market for patient monitoring devices projected to reach $71.1 billion by 2029.

Researchers at the University of Florida have developed a deep learning–based acuity scoring model that autonomously and continuously analyzes physiologic and patient background data in real-time for ICU patients. By leveraging advanced temporal deep learning models, the system generates dynamic risk assessments and mortality predictions. Unlike static scores calculated every 12–24 hours, this system updates continuously and autonomously without manual data entry. This approach delivers more accurate, interpretable insights to support earlier, data-driven interventions, ultimately improving patient outcomes and reducing the workload on ICU staff.

Application

Continuous ICU acuity scoring system that autonomously analyzes real-time physiologic and patient background data to deliver dynamic risk assessments and mortality predictions, supporting timely, data-driven clinical interventions

Advantages

• Continuously updates patient risk scores, enabling rapid and proactive medical interventions
• Leverages temporal trends and physiologic data, providing more accurate predictions than traditional static scoring systems
• Reduces alarm fatigue and clinician cognitive load while enabling earlier escalation of care
• Highlights clinically relevant time points, improving interpretability and decision-making
• Supports preventative and reactionary care, allowing clinicians to intervene earlier

Technology

This predictive model applies deep learning to generate continuous acuity scores for ICU patients by combining recurrent neural networks (RNNs), gated recurrent units (GRUs), and self-attention mechanisms. The system ingests a comprehensive array of physiologic measurements (blood pressure, oxygenation, urine output, lab values, vasopressor use, mechanical ventilation status) along with patient background information (demographics, comorbidities). RNNs learn patterns in patient data over time, GRUs optimize information flow for efficiency and stability, and self-attention mechanisms identify the most clinically significant time points. This approach enables accurate, interpretable risk predictions tailored for complex ICU environments.