A data-driven clinical prediction rule for pulmonary embolism

Purpose: Pulmonary embolism (PE), a life-threatening emergency is underdiagnosed because of a non-specific presentation.First-level exams (clinical exhamination, electrocardiography, blood gas analysis and laboratory tests)have low sensitivity and specificity.Clinical prediction rulers (CPRs)such as Wells and Geneva Revised, have been derived from different combinations of these exams. Our aim was to perform a comparison between the two score systems in our population and to derive a new CPR using an Artificial Neural Network (ANN). Methods: We enrolled 755 consecutive outpatients with suspect of PE (351 males, mean age 71±14years) and analyzed 24 clinical,instrumental and laboratoristic variables and Wells and Geneva scores.Logistic regression with ROC curves was used to evaluate the the diagnostic reliability of the scores. To derive the new classifier,the dataset was first split (in supervised classification step)into a train and a test subset containing 2/3 ad 1/3 of the patients' dataset,respectively.To find the optimal configuration of the new classifier we tested two different ANNs:a non-linear feed-forward ANN with back-propagation and a Levenberg-Marquardt network.For both we fixed the topological configurations of the network (one hidden layer,one output neuron)and stressed the system to find the optimal number of neurons in the hidden layer for the best configuration among highest AUC with the highest number of hit in the validation process and the minimum epochs.We repeated this study changing the dimension of the input dataset in two ways:excluding interactively some features or performing the reduction of the dimensionality of the feature space with principal component analysis. The application of the trained ANN to a "map set" gave,for each patient,the probability of belonging to the "pathological" or "healthy" class, obtaining the new CPR.Automatic classifications were compared with the manual ones, calculating the Jaccard coefficient, giving a measure of the quality. The system was implemented in Matlab using Neural Network toolboxes and PRTools. Results: In our population,Wells performed better than Revised Geneva (AUC 0.75%,0.63%,respectively),while our CPR (feed-forward ANN with back-propagation) obtained an average AUC of 0.86% from the train set and Jaccard coefficient 0.86 from the map set.The optimal ANN configuration was with 3 neurons in the hidden layer.The difference among the three ROC curves resulted statistically significant. Conclusions: An ANN-based CPR performs better in the clinical prediction of PE than classical rulers without increasing the number of items required for the analysis.