Artificial neural networks for prediction of recurrent venous thromboembolism

Recurrent venous thromboembolism (RVTE) is a multifactorial disease with occurrence rates which vary from 13 % to 25 % in 5 years after the initial event. Once a patient the first thrombotic event, the probability of recurrence should be determined, as well as the most adequate anticoagulant therapy. To our knowledge based on the published literature, three statistical models have been proposed to calculate RVTE probability. However, these models present several limitations, such as: limited input variables, lack of external validation and applicability only for patients with a first unprovoked thrombosis. Additionally, some of the models have been recognized to fail in determining RVTE when patients have a low risk of recurrence.

An alternative procedure in which three Artificial Neural Network (ANN) models were developed to classify which patients will present RVTE based solely on clinical data.

Data of 39 clinical factors from 235 patients were used to train several ANN structures. The difference among the three models was its inputs. In ANN 1, the inputs were all 39 factors. In ANN 2, 18 factors determined previously as the main predictors of RTVE using Principal Component Analysis (PCA). Finally, in ANN 3, 15 factors combining PCA results with practical aspects. Different number of hidden layers and neurons, and three optimization algorithms were considered. 5-fold cross validation was also performed.

The results showed that all models were capable of performing this task. Different optimization algorithms lead to different results. The best models presented high accuracy. The best structures were 39-10-10-1, 18-10-5-1, and 15-15-10-1 for ANN 1, ANN 2, and ANN 3 models, respectively. The cross-validation showed that the results are consistent.

This work showed that the association of multivariate techniques and ANNs is a powerful tool that can be used to model a complex phenomenon such as RVTE without the restrictions of existing approaches.

After proper validation, these ANN models can be used to help clinicians with decisions regarding VTE treatment.