Abstract 14932: Prediction of Pulmonary Embolism Using Random Forest Algorithm: A Study on 917 Patients Over a Period of 1 Year in a New York City Public Hospital

Abstract

Introduction: Pulmonary Embolism (PE) can be challenging to diagnose. Artificial intelligence (AI)-generated models can help improve the diagnostic accuracy of detecting PE. Aim: Establish the diagnostic accuracy of AI-generated models in predicting PE as compared to routinely used diagnostic predictive tools (Well’s score, YEARS score, and revised Geneva score). Methods: We conducted a single-center retrospective review of patients who underwent computed tomography pulmonary angiography (CTPA) for diagnosing PE over a one-year period. Twenty-five parameters including age, sex, race, BMI, history of deep vein thrombosis or PE, malignancy, hypertension, COPD, CKD, presenting symptoms like chest pain, dyspnea, cough, palpitations, dizziness, lightheadedness/syncope, hemoptysis, fever, pain on palpation of the lower extremity, vitals including heart rate, respiratory rate, oxygen saturation, findings of lower extremity venous duplex, presence of right ventricular dysfunction on echocardiogram, elevated troponin and pro-brain natriuretic peptide were utilized to train the AI model. Random forest algorithms were trained for predicting PE in 80% of the dataset, while the remaining 20% was used to test the AI model. Subsequently, the accuracy of the model was matched with the accuracy of four alternative modalities: clinical judgment of PE being equally or more likely, Well’s score >4, YEARS score ≥1, and revised Geneva score ≥11. CTPA results were used as the gold standard. Analysis was performed using STATA BE/17 and Python 3.8. Results: 917 patients were included in the analysis (median age: 57 years (IQR: 41-68); female: 59%). The AI model correctly classified patients with 84.6% diagnostic accuracy, outperforming every other modality (Figure 1). Conclusion: AI-generated prediction models can be useful in predicting PE with considerable accuracy. Further, large-scale studies are required to validate these results externally.