A machine learning model to identify pulmonary embolism in patients admitted to intensive care

Pulmonary embolism (PE) is a leading cause of preventable death, yet statistical prediction models have shown inconsistent validity. Our primary objective was to determine if a machine learning model trained with data routinely collected in clinical care can successfully identify acute PE in critically ill patients.

Leveraging two multicenter datasets acquired nationally (development cohort) and within the Johns Hopkins Health System (external validation cohort), we trained machine learning models with features extracted from demographics, comorbidities, physiologic and laboratory data available following intensive care unit (ICU) admission. The primary endpoint was the identification of acute PE during ICU admission. Model performance was contrasted with two benchmark PE risk scores.

PE was diagnosed in 2647 of 164,383 (1.61%) and 754 of 64,923 admissions (1.16%) in the development and external validation datasets respectively. Using data from the first 48 h after ICU admission, the mean (95% CI) discrimination measured by area under the receiver characteristic curve (AUROC) was 0.829 (0.808-0.852), 0.704 (0.681-0.727), and 0.667 (0.653-0.681) for our logistic regression machine learning model and for the two benchmark scores, respectively; mean area under the precision recall curve was 0.150 (0.138-0.162), 0.080 (0.071-0.089), and 0.081 (0.071-0.091), respectively. Discrimination was maintained in the external validation dataset with an AUROC of 0.819 (0.802-0.836).

Findings indicate that PE can be detected accurately in ICU patients using routinely collected clinical data. The machine learning model successfully validated and outperformed existing benchmark risk scores. Such a model could become a valuable tool for assessing the likelihood of PE among critically ill patients.