Development of a machine learning model using electrocardiogram signals to improve acute pulmonary embolism screening

Area of Science:

• Artificial Intelligence in Medicine
• Cardiovascular Diagnostics
• Pulmonary Medicine

Background:

• Clinical scoring systems for pulmonary embolism (PE) exhibit low specificity, leading to overuse of computed tomography pulmonary angiography (CTPA).
• There is a need for improved diagnostic tools to enhance PE detection accuracy and reduce unnecessary imaging.

Purpose of the Study:

• To assess the efficacy of deep learning models utilizing electrocardiogram (ECG) waveforms for increasing PE detection specificity.
• To compare the performance of an integrated deep learning model against traditional clinical scoring systems.

Main Methods:

• A retrospective cohort of 21,183 patients with moderate- to high suspicion of PE was analyzed.
• Three machine learning models were developed: an ECG-only model, an Electronic Health Record (EHR) model, and a Fusion model combining EHR data with ECG waveform embeddings.
• Model performance was evaluated using area under the receiver-operating characteristic curve (AUROC) and specificity.

Main Results:

• The Fusion model achieved a superior AUROC of 0.81 (±0.01), outperforming the ECG model (0.59 ±0.01) and EHR model (0.65 ±0.01).
• The Fusion model demonstrated higher specificity (0.18) and performance (AUROC 0.84 ±0.01) compared to four established clinical scores (AUROC 0.50-0.58).
• The model maintained comparable performance across different sex and racial/ethnic subgroups.

Conclusions:

• Synergistic deep learning of ECG waveforms with clinical variables enhances PE detection specificity in patients with at least moderate suspicion.
• This AI-driven approach holds promise for improving diagnostic accuracy and potentially reducing CTPA utilization in PE screening.