Mean arterial pressure is all you need in a machine learning model for mean arterial pressure prediction.

Background: Anaesthesiology and intensive care use monitoring to identify patients in danger of deterioration. Traditionally, trends and early warning scores allow clinicians to predict deterioration with moderate reliability. Reduced mean arterial blood pressure has been associated with complications, and models have been sought to predict its value. Machine learning methods with complex inputs have been used for predictive monitoring in hospital care.

Objectives: This study evaluates whether machine learning can predict mean arterial pressure (MAP) from previous values.

Design: This is a monocentre, retrospective, exploratory, observational cohort study using the MIMIC-III-WDB, VitalDB and an internal study centre dataset, training machine learning models on adult patients with invasively measured blood pressure (IBP) as input during an observation window up to 20 min before the prediction horizon (5 to 20 min).

Setting: Kepler University Hospital, Linz, Austria.

Participants: Two thousand three hundred and forty-six patients from the internal dataset, 4741 patients from MIMIC-III-WDB and 3357 patients from VitalDB were analysed.

Main outcome measures: The primary endpoint was model performance in predicting whether MAP would fall below 65 mmHg in a given time frame. In a secondary analysis, we restricted the input set to stable patients with current MAP above 65 mmHg.

Results: Models using the complete training data achieved receiver operating characteristic area under the curves (ROC AUCs) of 0.963, 0.946, 0.934 and 0.923 on the internal dataset for 5, 10, 15 and 20 min of prediction horizon, respectively, and 0.856, 0.837, 0.821 and 0.804 in the secondary analysis. The maximum difference of ROC AUC to baseline measurement (ROC AUC of last measured MAP as trivial estimator) was 0.006 for the complete training data and 0.051 for stable patients. The prediction of MAP may allow clinicians to intervene in time before MAP deterioration becomes clinically relevant.

Conclusion: Predicting MAP below 65 mmHg within 5, 10, 15 and 20 min for patients with and without a MAP above 65 mmHg is possible and requires only MAP as input for machine learning models.

Trial registration: ClinicalTrials.gov (NCT05471193).