PUB - Publikationen an der Universität Bielefeld

### Introduction Recently, there have been significant advances in the application of AI-based models to support diagnostic decision making – both in general and in the context of epilepsy. When developing these models, one tries to reduce the number of variables in order to increase the model performance and to reduce the required interaction time with the final system. Statistical feature selection methods can be used to determine features with the strongest impact on the decision. In contrast to this, we propose a dynamic feature selection scheme that determines the best sequences of features in each case individually. Our novel method determines the next most relevant variable or diagnostic test to consider at each step of the decision making process, in order to reduce the uncertainty of the decision making as quickly as possible. ### Methods A Bayesian network with 35 variables is used to predict the probabilities of three possible diagnoses (epileptic seizure, psychogenic non-epileptic seizure or syncope), given the available information about the patient. Our proposed algorithm iteratively selects the best next feature on the basis of mutual information to determine the variable whose observation would result in the highest information gain. At each step, the evidence for the variable in question is updated and the predictions of the network are recomputed. The iterative process is continued, until the uncertainty of the diagnosis is sufficiently lowered. Our approach relies on the entropy of the predicted diagnosis to estimate the uncertainty. For evaluation, we use a dataset with 177 patients (63 syncope, 61 epileptic-seizure, and 53 psychogenic-seizure) as described in the work of Wardrope et al. (2020). ### Results The Bayesian network reaches a baseline accuracy of 80.23% in the case of complete evidence (all 35 features). With our approach, the mean number of queried features can be reduced by over 50 percent with a negligible decrease in accuracy. With an entropy threshold of 0.1, the mean number of queried features is reduced to 12.5 ± 10.73 with an overall accuracy of 79.10%. Lowering the entropy threshold to 0.05, the mean number of queried features slightly increases to 16.3 ± 11.77, while the accuracy is only reduced to 79.66%. Furthermore, it is noteworthy that in 51% of all cases less than 8 features were required. All 35 features were queried only in 16% of all cases. ### Conclusions We have shown that our method is capable of significantly reducing the number of features needed to correctly differentiate between epileptic seizures, psychogenic non-epileptic seizures and syncopes. This is very relevant from a medical (and an economical) point of view, as the correct diagnosis is found faster and without having to perform unnecessary tests. Further work will include evaluating sequences of queried features across cases to check which sequences of features might be most predictive or discriminative for one diagnosis. In addition, this work can be used as a basis to create guidance tools that can suggest optimal next questions or remind of rare features, which might only be relevant in specific cases.