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

Conventional sPESI discrimination attenuated progressively over time across three mortality horizons.

• AUC for conventional sPESI was 0.727 at 30 days, 0.689 at 180 days, and 0.665 at one year.
• The study used a retrospective cohort of 2547 adults with CT pulmonary angiography confirmed PE.
• This attenuation motivated the development of ML extensions to improve long-horizon risk stratification.

The 10-item ML extension significantly enhanced prognostic performance compared with conventional 6-item sPESI across all time horizons.

• All comparisons between ML models and conventional sPESI were statistically significant (all p < 0.05).
• The 10-item hybrid extension retained continuous age, heart rate, systolic blood pressure, and oxygen saturation alongside the six binary sPESI indicators.
• Three model types were evaluated: logistic regression (LR), XGBoost, and multi-layer perceptron (MLP).
• Models were trained and optimized using 12-month all-cause mortality as the target outcome.

At the primary 12-month horizon, all three ML models outperformed conventional sPESI in discrimination.

• LR, XGBoost, and MLP achieved AUCs of 0.720, 0.726, and 0.712, respectively, at 12 months.
• Conventional sPESI AUC at 12 months was 0.665.
• XGBoost achieved the highest AUC among ML models at the 12-month horizon (0.726).

XGBoost achieved peak discrimination at 30 days with an AUC of 0.796.

• The 30-day AUC of 0.796 for XGBoost compared favorably to 0.727 for conventional sPESI at the same horizon.
• Peak discrimination across all models and all horizons was achieved by XGBoost at the 30-day timepoint.

ML models substantially improved 30-day specificity compared to conventional sPESI while maintaining high sensitivity.

• ML models increased 30-day specificity to 0.326–0.395 compared to 0.131 for conventional sPESI.
• High sensitivity was maintained across ML models at 0.930–0.982 at the 30-day horizon.
• This represents a two- to three-fold improvement in specificity over conventional sPESI.

SHAP analyses revealed horizon-linked attribution shifts in the relative importance of predictors across mortality timepoints.

• Acute hemodynamic markers dominated early (30-day) mortality predictions.
• Age and malignancy emerged as the primary drivers at 12 months under horizon-specific evaluation.
• SHapley Additive exPlanations (SHAP) analyses were used to interpret feature contributions across models.
• This finding suggests that different physiological and clinical variables carry different prognostic weight depending on the time horizon.

The study used a retrospective cohort design with a held-out test set and 10-fold cross-validation for model development and evaluation.

• The cohort comprised 2547 adults with CT pulmonary angiography confirmed acute pulmonary embolism.
• Models were evaluated on a held-out test set at 30, 180, and 365-day mortality horizons.
• Model optimization used 10-item hybrid sPESI features with continuous physiological variables replacing dichotomized versions.