Abstract 13169: Validation of an Artificial Intelligence Algorithm for Diagnostic Prediction of Coronary Disease in Patients With Acute Thoracic Pain: Comparison With an Human Intelligence Model

IntroductionThe evolution of computers ability to process information has enabled algorithms capable of translation an infinity of patterns into probabilities. The accuracy of this artificial intelligence (AI), compared with traditional statistical models, is not established in the medical field.HypothesisTo test the concept of AI in the development of an accurate probabilistic algorithm for predicting CAD in the acute chest pain scenario, comparing its performance with the traditional model of logistic regression.MethodsThe Registry of Acute Thoracic Pain consists of a consecutive sample of 963 patients admitted to the Coronary Unit due to chest pain, whose clinical and laboratory data are prospectively collected (24 candidates for predictors) and the presence of CAD (defined by angiographic obstruction > 70%) systematically investigated. From the first 2/3 of the patients (training sample) two probabilistic models of CAD were builta machine learning algorithm and a traditional logistic model. The performance of these two predictive strategies were evaluated in the remaining third of the patients (test sample).ResultsThe training sample consisted of 642 patients, with a 52% prevalence of CAD. The logistic model was composed of the following independent predictorsage, male, gender, ECG ischemia, positive troponin, signs of ventricular failure, non-pleuritic pain, smoking, previous CAD and physical / emotional stress at the moment of pain. The machine learning algorithm was composed of all candidates for predictors. In the test sample, the area below the ROC curve in predicting CAD was 0.81 (95% CI = 0.77 - 0.86) for the AI model, similar to that obtained in the logistic model (0.82; 95% CI = 0.77 - 0.87) - P = 0.68. Predicted-observed linear regression analysis and Hosmer-Lemeshow test showed adequate calibration for the AI model (r = 0.95; α = - 0.11 and β = 1.23; χ2 = 13; P = 0.11) and for the logistic model (r = 0.98; α = -0.02; β = 0.99; χ2 = 5.5; P = 0.70).ConclusionsThe AI model was accurate and calibrated in the probabilistic prediction. Considering the continuous improvement of accuracy as artificial intelligence is exposed to new cases, our result suggests that this approach should be explored in the construction of medical probabilistic models.