Machine learning-extended sPESI for 1-year mortality prediction in pulmonary embolism

Area of Science:

• Cardiology
• Medical Informatics
• Machine Learning in Medicine

Background:

• The simplified pulmonary embolism severity index (sPESI) is a widely used bedside tool for assessing acute pulmonary embolism (PE) risk.
• Current sPESI limitations include dichotomized physiological variables and validation primarily for 30-day outcomes.
• There is a need for enhanced prognostic precision and extended risk stratification beyond 30 days for PE patients.

Purpose of the Study:

• To develop and validate a machine learning (ML) extension of the sPESI using bedside data for improved prognostic accuracy.
• To extend risk stratification for pulmonary embolism patients to a 12-month horizon.
• To compare the performance of ML models against the conventional sPESI for predicting mortality at various time points.

Main Methods:

• A retrospective cohort of 2547 adult PE patients confirmed by CT pulmonary angiography was analyzed.
• Logistic Regression, XGBoost, and Multi-layer Perceptron models were developed using a 10-item hybrid sPESI extension.
• Models retained continuous physiological variables (age, heart rate, systolic blood pressure, oxygen saturation) and were trained to predict 12-month all-cause mortality.

Main Results:

• The ML extension significantly outperformed the conventional sPESI in prognostic discrimination across 30-day, 180-day, and 12-month horizons (p<0.05).
• At 12 months, ML models achieved Area Under the Curve (AUC) values of 0.712-0.726, compared to 0.665 for sPESI.
• ML models demonstrated improved 30-day specificity (0.326-0.395 vs. 0.131) while maintaining high sensitivity (0.930-0.982).

Conclusions:

• Machine learning models incorporating continuous physiological data enhance prognostic precision for pulmonary embolism.
• This ML extension provides improved long-horizon risk stratification, capturing both acute and latent patient vulnerabilities.
• The bedside-feasible ML approach facilitates individualized management strategies for PE patients.