Prediction of in-hospital adverse clinical outcomes in patients with pulmonary thromboembolism, machine learning based models

Abstract

Background: Pulmonary thromboembolism (PE) is the third leading cause of cardiovascular events. The conventional modeling methods and severity risk scores lack multiple laboratories, paraclinical and imaging data. Data science and machine learning (ML) based prediction models may help better predict outcomes.

Materials and methods: In this retrospective registry-based design, all consecutive hospitalized patients diagnosed with pulmonary thromboembolism (based on pulmonary CT angiography) from 2011 to 2019 were recruited. ML based algorithms [Gradient Boosting (GB) and Deep Learning (DL)] were applied and compared with logistic regression (LR) to predict hemodynamic instability and/or all-cause mortality.

Results: A total number of 1,017 patients were finally enrolled in the study, including 465 women and 552 men. Overall incidence of study main endpoint was 9.6%, (7.2% in men and 12.4% in women; p-value = 0.05). The overall performance of the GB model is better than the other two models (AUC: 0.94 for GB vs. 0.88 and 0.90 for DL and LR models respectively). Based on GB model, lower O₂ saturation and right ventricle dilation and dysfunction were among the strongest adverse event predictors.

Conclusion: ML-based models have notable prediction ability in PE patients. These algorithms may help physicians to detect high-risk patients earlier and take appropriate preventive measures.

Keywords: gradient boosting machine; logistic models; machine learing; outcome analysis; pulmonary embolism; risk factors.

Figure 1

Comparison of the receiver operating characteristic curve (ROC curve) of machine learning models. AUC, area under the curve.

Figure 2

Comparison of the precision-recall curve (PR curve) of machine learning models. AUC, area under the curve.

Figure 3

SHAP summary plot sorts the features used in the gradient boosting model, concerning their importance in the final decision-making process. The y-axis shows the variables, while the x-axis represents the Shapley value of each variable. The colors illustrate the relationship between features and prediction (composite adverse outcomes), red shows the direct relationship, whereas blue means inverse relation. SHAP, SHapley Additive exPlanations; Echo, echocardiography; Lab, laboratory test; CT angio, pulmonary CT angiography.

Figure 4

Graph of classification-error of the gradient boosting model versus the number of trees, showing the decreasing trend of errors in both training and validation datasets.