Yapay sinir ağları kullanılarak serebral iskemi ve hemorajiolaylarının önceden tahmin edilmesi

Stroke, which has two main types as cerebral ischemia and hemorrhage, is a serious health condition that is among the leading causes of death and permanent disability worldwide. Despite this fact, a generally valid, effective treatment method is yet to be found in the struggle against stroke. This situation has led to the concentration of studies on the prediction of possible attacks in order to minimize the impact of stroke events. However, the acute nature of stroke events and the high number of factors affecting the risk of stroke, some of which are still unknown, make it difficult to establish an accurate risk estimation method. A novel, 3000-patient dataset is formed and used in training and testing of the proposed model within the context of study, using the data obtained from Gazi University Hospital retrospectively. The model proposed in the study consists of 2 main stages. In the first stage, the dataset input to the model is clustered using fuzzy c-means algorithm and the clustering results obtained are used to modify the dataset. In the second stage, this modified dataset is taken as input and the risks of the patients in the set for 4 different stroke types are estimated with ensemble learning based classifiers that consist of 4 different base models for each stroke type. With a novel approach, the weights of the base models used in this stage are determined dynamically according to their performances in validation dataset, separately for each cluster. By this way, not only a high model performance is obtained, but also the operating time and computational load of the model is reduced considerably. Furthermore, the clustering process and the usage of different weights in the resulting clusters provided the model with resistance against possible biases in the datasets. The accuracy of the model in predicting general stroke, ischemic stroke, hemorrhagic stroke and transient ischemic attack events are obtained as 99,89%; 99,90%; 99,81% and 99,87%; respectively. When these accuracy values are considered, it is observed that the model proposed in the study shows almost perfect performance. In this context, this model, with its ability to perform risk estimation accurately for 4 different types of stroke by considering a large number of factors can aid the medical experts by functioning as a decision support system.