Decision Support Algorithm Using Machine Learning for the Diagnosis of Pulmonary Embolism on Chest X-ray

Introduction: A patient with suspected pulmonary embolism (PE), unable to undergo pulmonary computed tomographic angiography (PCTA) or being in primary care/ emergency room without PCTA is a major clinical issue. Therewithal, despite using risk scores, the number of PCTAs performed is relatively high, increasing healthcare costs and workload, exposing patients to radiation, and putting them into the risk of nephrotoxicity due to contrast agents. The present study aims to reveal whether the evaluation of chest X-rays with the support of machine learning will be helpful in excluding or confirming the diagnosis of PE. Methods: The data set, which included chest X-rays from 200 patients diagnosed with PE by PCTA, 200 chest X-rays from healthy subjects, and 100 chest X-rays from patients with pleural effusion due to various reasons other than PE, was used for training and testing. All X-rays were anonymized. After training and validation, the machine learning algorithm was tested with not only internal but also external test data sets. Results: Internal test results were as the follow; recall 95%, precision 76%, and F1 score 84.4. The disease prediction probabilities of the 5 patients known to have PE were 97%, 87%, 78%, 77%, and 67%, while the disease prediction probabilities of the 5 patients known not to have PE were 53%, 57%, 58%, 71%, and 81%. Conclusion: Even though the findings come from a study with a limited number of patients, they are encouraging. For the patients with suspected PE, unable to undergo PCTA; a more accurate risk assessment from chest X-rays using machine learning can be possible with further studies using larger data.