Artificial intelligence (AI) applications in hematology offer the possibility of radical innovations in disease diagnostics. At a session of the 66th American Society of Hematology Annual Meeting and Exposition, Damon E. Houghton, MD, MS, presented data further validating the performance of an AI algorithm for predicting the likely presence of pulmonary embolism (PE) based on electrocardiogram (ECG) signals.1
Accurate diagnosis of PE is challenging for emergency room physicians because clinical presentation can vary widely, and other cardiovascular emergencies may be difficult to distinguish. Dr. Houghton pointed out that venous thromboembolism is the fifth leading cause of serious misdiagnosis-related harm according to the Agency for Health Care Research and Quality. Correct and prompt diagnosis of PE is essential, particularly for patients with larger PEs who require treatment with fibrinolytics, anticoagulants, or both.2
Multiple well-validated clinical scoring tools have been developed to help diagnose PE, such as the PERC or Wells’ stratification tools; these incorporate different clinical factors to help estimate risk. However, Dr. Houghton noted, “These are limited by variable uptake in our clinical practice and acceptance by first-line clinicians, and they have perhaps only moderate diagnostic accuracies.”
Computed tomography pulmonary angiogram (CTPA) can be used to rule out PE, particularly in patients whose clinical features and D-dimer scores put them at higher risk, but it is expensive, not always available, unsafe in renal disease, and carries standard risks of radiation exposure. In contrast, ECG is a quick, easy, noninvasive test that almost all emergency department (ED) patients with potential cardiopulmonary symptoms receive.1,2
With his colleagues, Dr. Houghton, a hematologist and vascular medicine specialist at Mayo Clinic in Rochester, Minnesota, previously developed an AI deep neural network model that incorporated machine learning techniques to analyze ECG data to help predict PE. This initial analysis included more than 79,000 patients, of whom 7,400 were found to have acute PE in the ED via CTPA.2
In his presentation, Dr. Houghton shared data validating these findings in a new population, reporting on results from more than 18,000 patients who had ECG performed within six hours of admission, ultimately including 455 patients found positive for PE via CTPA or D-dimer testing (4.3%).1
Using their previously developed algorithm, Dr. Houghton and colleagues found that the AUC (area under the curve) was 0.69, identical to their original cohort. Just using ECG, they found that patients could be stratified into low (0.74%; negative predictive value [NPV] 99.3%), elevated (2.27%; NPV 97.7%), and high-risk (8.02%; NPV 92.0%) categories.1
The team used additional clinical data to explore and train multiple new algorithms, such as adding age, sex, or oxygen saturation as a factor. They found that the best results came from an algorithm that used both AI-ECG and D-dimer results, achieving an AUC of 0.93. By using a negative D-dimer in the elevated-risk group, the NPV could be increased to greater than 99.9%, further reducing the need for CTPA imaging.1
Dr. Houghton pointed out that other groups have also used AI tools to develop ECG recognition algorithms for detecting PE, with impressive results.3,4 “Compared to traditional risk stratification, we think that an AI/ECG-based workflow could lead to earlier PE diagnoses and potentially even help prevent missed or delayed PE diagnoses,” he said.
“A simplified two-component ECG/D-dimer risk stratification tool might overall simplify risk stratification for PE in comparison to existing models,” concluded Dr. Houghton. “We think that this algorithm could fit in very well in real-world practices.”
Any conflicts of interest declared by the authors can be found in the original abstract.
References
1. Houghton DE, Liu K, Lopez-Jimenez F, et al. The AI-Mayo PE (AIM PE) study: validation of an artificial intelligence algorithm using electrocardiograms to predict pulmonary emboli. Abstract 811. Presented at the 66th American Society of Hematology Annual Meeting and Exposition; December 9, 2024; San Diego, California.
2. Wysonkinski WE, Meverden RA, Lopez-Jimenez F, et al. Electrocardiogram signal analysis with a machine learning model predicts the presence of pulmonary embolism with accuracy dependent on embolism burden. Mayo Clin Proc Digital Health. 2024:2(3):453-462.
