Vital Signs-Only Machine Learning Model for Acute Inpatient Deterioration: A Retrospective Multicenter Study

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Clinical Decision Support

Background:

• Clinical deterioration requires timely intervention to prevent adverse outcomes.
• Existing methods for predicting deterioration often suffer from low positive predictive values (PPVs), leading to alert fatigue.
• Developing accurate predictive models compatible with vital signs monitoring devices is essential for early risk identification.

Purpose of the Study:

• To develop and validate a machine learning model for identifying patients at risk of clinical deterioration.
• To ensure the model is compatible with vital signs monitoring devices for seamless integration into clinical workflows.
• To minimize false positives while maintaining clinical relevance in detecting low-prevalence deterioration events.

Main Methods:

• Utilized vital signs data from over 30,000 inpatients across a multihospital system.
• Trained a Light Gradient Boosting Machine model on a large dataset of high-quality vital signs.
• Employed a novel 2-window feature extraction method to capture both short-term and long-term patient status changes.

Main Results:

• The model achieved a sensitivity of 73.4% and a positive predictive value (PPV) of 30.4%.
• Demonstrated a high C-statistic of 0.874, indicating strong discriminative ability.
• The PPV was significantly higher at specific sites (e.g., 54.9% at Rochester), outperforming existing indices.

Conclusions:

• The developed machine learning model effectively predicts clinical deterioration with improved accuracy.
• The novel 2-window feature extraction method enhances predictive performance by analyzing temporal trends.
• High PPV is critical for reducing false alarms and improving the reliability of clinical deterioration alerts.