Early Prediction of ICU Hypotension Using Machine Learning

A Prospective Observational Machine Learning Study for the Early Prediction of Hypotension in Adult Intensive Care Unit Patients

This prospective observational study aims to develop and internally validate a machine learning model for the early prediction of hypotension in adult intensive care unit patients. The model will use routinely collected non-invasive vital signs, heart rate, medication-dose records, and fluid-balance data recorded during standard ICU care. No intervention will be assigned by the study, and patient management will not be changed according to the model output. The primary aim is to predict hypotension 30 minutes before its occurrence; shorter 5- and 15-minute prediction horizons will also be evaluated.

Hypotension is a frequent hemodynamic event in critically ill patients and may occur before clear clinical deterioration is recognized. Earlier identification of patients at risk may support closer clinical attention and more timely evaluation. This study is designed as a prospective, observational machine learning study in adult intensive care unit patients.

Routinely available ICU data will be collected at five-minute intervals, including systolic, mean, and diastolic non-invasive blood pressure, heart rate, medication-dose entries, and fluid-balance records. These data will be used to construct time-dependent features reflecting recent values, short-term changes, and rolling trends. Hypotension will be defined at each five-minute time point as systolic blood pressure below 90 mmHg, mean arterial pressure below 65 mmHg, or diastolic blood pressure below 60 mmHg.

The primary prediction horizon will be 30 minutes. Separate secondary analyses will evaluate 5- and 15-minute prediction horizons. A gradient-boosted decision-tree model will be developed and internally validated using patient-level data partitioning to avoid assigning observations from the same patient to both training and validation sets. Model performance will be assessed using discrimination, classification performance, and calibration measures. Feature-importance analyses will be used to describe the variables contributing to model predictions.

The study is observational. No treatment, medication, device, alarm, or clinical decision will be assigned by the study protocol. The prediction model will be developed and evaluated using collected data and will not be used to guide real-time patient management during the study period.

Inclusion Criteria:

• Age 18 years or older
• Admission to the adult intensive care unit during the study period
• Length of stay in the intensive care unit of at least 24 hours
• Availability of routine intensive care unit monitoring data
• Availability of non-invasive blood pressure and heart rate measurements recorded during ICU monitoring
• Availability of medication-dose and/or fluid-balance records during ICU monitoring

Exclusion Criteria:

• Age younger than 18 years
• Length of stay in the intensive care unit of less than 24 hours
• Absence of usable blood pressure monitoring data
• Records with irrecoverable timestamp inconsistencies
• Insufficient monitoring duration for feature construction and future outcome labeling