A Comparative Analysis of Electronic Health Record and Electrocardiogram Waveform Data for Pulmonary Embolism Identification in Critically Ill Patients

Pulmonary embolism (PE) is one of the leading causes of preventable death amongst hospitalized patients, yet current risk assessment tests based on clinical variables have shown inconsistent validity or poor predictiveness. More recent predictive models for PE using electronic health record (EHR) data are promising, but their reliance on comprehensive and integrated EHR data can limit real-time utility, creating a need for more accessible and rapid diagnostic tools. This study compares the performance of an EHR-based model, an electrocardiogram waveform (WF) model, and a fusion model combining both modalities for the identification of PE in critically ill patients. We leverage routinely acquired clinical and ECG waveform data from the 48 hours preceding PE suspicion from a retrospective dataset of 7,132 ICU admissions between 2008 and 2019 (4.60% PE prevalence). PE diagnoses were determined through ICD-9 or ICD-10 diagnostic coding. We find that our WF model, which employs a single, 10-second 12-lead ECG sample, demonstrated comparable predictive performance (AUROC 0.67 (95% CI, 0.64-0.70)) to our EHR-based model (AUROC 0.71 (0.68-0.74)). However, a fusion model combining both modalities did not improve predictive performance (AUROC 0.67 (0.64-0.70)). All our models outperform widely used existing risk stratification scores such as the Revised Geneva score (AUROC 0.54 (0.51-0.57)), the original Wells score (AUROC 0.61 (0.58-0.64)), and the PE Rule Out Criteria (AUROC 0.56 (0.53-0.59)). Our findings underscore the value that ECG waveform data can bring to the detection of PE in critically ill patients by demonstrating its predictive capability compared to existing benchmarks. After additional validation, these models may serve as valuable tools in PE diagnostic clinical workflows.