Background: To develop an accurate model for venous thromboembolism recurrence prediction is important to determine the time of patient’s treatment.
Aims: To obtain Neuro-Fuzzy Networks-based models for the prediction of recurrent venous thrombosis using simple clinical variables.
Methods: From thirty-nine clinical and laboratory factors initially selected, two models were developed, as shown at Table 1.
The set (i), with 7 factors, was determined via Design of Experiments [Res Pract Thromb Haemost 4 (2020) 1149] in conjunction with factors pointed out as significant by Principal Component Analysis in Martins [Unicamp (Thesis) 2018)].
In set (ii) were considered 16 variables, which were the factors obtained via Principal Component Analysis, by Martins. Data from 236 patients were included: 70% were used for training the networks and 30% for simulation. The Neuro-Fuzzy models were developed using MatLab software, varying the parameters of the Takagi-Sugeno model, and consequently the number of fuzzy rules. A 5-fold cross-validation was also performed.
Results: Two best models were selected, one for each set. The model for set (i) showed 83% (AUC: 0.726) of accuracy for training data, and 88% (AUC: 0.909) for simulation data. For set (ii), 85% (AUC: 0.905) of accuracy for training data and 81% (AUC: 0.818) for simulation data. The two models developed showed that Neuro-Fuzzy Nets have a good ability to predict rethrombosis. Both models presented accuracy above 80% and the AUC values indicate that the model developed with ensemble (ii) has better generalization capacity. Finally, cross-validation demonstrated that the results were consistent. These results are important, since it showed that a larger set of independent variables is required to obtain a reliable model.
Set (i) | Set (ii) |
Leukocytes, anticoagulation time, age, location of thrombosis, red blood cells, hemoglobin, and whether the first thrombotic episode was spontaneous or provoked | Same as Set(i), plus D-dimer, hematocrit, erythrocyte size distribution, total cholesterol, HDL and LDL, triglycerides, glycemia, and creatinine |
Conclusions: The results of this study showed that Neuro-Fuzzy models can predict the recurrence of venous thromboembolism from different sets of independent factors. Once validated, this model can be used as a clinical decision support.
To cite this abstract in AMA style:
Ottaiano GY, Annichino-Bizzacchi JM, Filho RM, Martins TD. Development of Neuro-Fuzzy Networks for Venous Thromboembolism Recurrence Prediction [abstract]. Res Pract Thromb Haemost. 2021; 5 (Suppl 2). https://abstracts.isth.org/abstract/development-of-neuro-fuzzy-networks-for-venous-thromboembolism-recurrence-prediction/. Accessed March 22, 2024.« Back to ISTH 2021 Congress
ISTH Congress Abstracts - https://abstracts.isth.org/abstract/development-of-neuro-fuzzy-networks-for-venous-thromboembolism-recurrence-prediction/