Fig. 2

The schema of proposed VTE risk assessment ML approach. Firstly, training and test cohorts were constructed and patients in training data were split into different groups according to values of VTE-related clinical variables. For different group \(\:{C}^{\left(i\right)}\) and \(\:{C}^{\left(j\right)}\), their corresponding feature vectors \(\:{v}^{\left(i\right)}\) and \(\:{v}^{\left(j\right)}\) satisfied \(\:{v}^{\left(i\right)}\ne\:{v}^{\left(j\right)}\). Then VTE risk ratio was calculated in every group and groups were sorted accordingly. Next probability of distribution of patients in each group was estimated using VTE incidence rate and only groups with probability < 0.05 were saved. Based on sorted result, accumulated sensitivities and specificities were calculated for every group and groups were classified into high and low risks by thresholding, which formed a new training set based on groups. Using this training set, the proposed model consists of two modules, group-memory module for patients in known groups and group-prediction module for the unknown. Decision tree was used in group-memory module. For group-prediction module, VTE ratios for groups were used instead of high or low risk label, and artificial neural network was fitting