Machine learning algorithm for the diagnosis of pulmonary embolism: a proof of concept study

Introduction: Pulmonary embolism (PE) is a common life-threatening condition. Underdiagnosis and delays in the diagnosis are common, mainly due to unspecific clinical presentations. Aim: We used clinical parameters to train a Machine Learning algorithm for the diagnosis of PE in Emergency Department settings. Methods: We analyzed data from 190 patients that underwent Computed Tomography Pulmonary Angiogram (CTPA) due to clinical suspicion of PE in two tertiary hospitals. Gold standard was the diagnosis of pulmonary embolism in the CTPA scan. Using this data, we trained an algorithm in order to discern between patients with and without PE. Results: The best performance was achieved by the AdaBoost classification algorithm coupled with the Wrapper feature selection technique. In Table 1 we report accuracy, sensitivity and specificity of the proposed algorithm compared to some of the validated prediction rules for the assessment of pretest probability of PE (i.e. Wells score, Wells simplified, revised Geneva, revised Geneva simplified), on the same patient set. Conclusion: The proposed algorithm outperforms current scoring systems, especially in terms of accuracy and sensitivity; however, further training and validation with richer datasets is needed in order to assess its generalization capability. The incorporation of laboratory parameters as input to the algorithm is expected to enhance the predictive accuracy.