NEURAL-NETWORK IN THE CLINICAL-DIAGNOSIS OF ACUTE PULMONARY-EMBOLISM

The purpose of this investigation was to test the hypothesis that computer-based pattern recognition can accurately assess the likelihood of acute pulmonary embolism (PE) based on readily obtainable clinical characteristics. Data were obtained from 1,213 patients who participated in the collaborative study of the Prospective Investigation of Pulmonary Embolism Diagnosis (PIOPED). Characteristics of the history, physical examination, electrocardiograph, chest radiograph, and arterial blood gases of patients with suspected acute PE were presented to a backpropagation neural network. The 1,213 patients were divided into training set A (n = 606) and test set B (n = 607). These groups were then reversed into training set B (n = 607) and test set A (n = 606). A receiver operating characteristic (ROC) curve was constructed from PIOPED clinical assessment, and from neural network clinical assessment in groups A and B. Areas under the respective ROC curves were 0.7450, 0,7477, and 0.7324. All differences were not significant. Areas under ROC curves for PIOPED clinical assessment combined with ventilation/perfusion (V/Q) scan results were compared with neural network clinical assessment combined with V/Q scan results in groups A and B. The respective ROC areas were 0.8324, 0.8203, 0.8496 (all differences not signficant). These data show that neural networks were able to predict the clinical likelihood of PE with an accuracy comparable to experienced clinicians.