SPİNAL ANESTEZİ UYGULANAN SEZARYEN OPERASYONLARINDA İNDÜKSİYON SONRASI HİPOTANSİYON RİSKİNİ ÖNGÖREN YAPAY ZEKÂ ALGORİTMASI GELİŞTİRİLMESİ
Özet
Yavuzel S. Development of Artificial Intelligence Algorithm that Predicts the Risk of Post-Induction Hypotension in Cesarean Section Operations Under Spinal Anesthesia. Hacettepe University, Thesis in Anesthesiology and Reanimation, Ankara, 2024.The aim of this study is to develop deep learning algorithms to predict hypotension after induction in cesarean sections with spinal anesthesia. Preoperative data collected from patients was used for this purpose. A total of 370 patients who applied to the Hacettepe University Hospital Department 81 Delivery Room between February 2023 and December 2023 were included in the study, and a dataset consisting of a total of 45,140 variables, including 122 unique features and the outcome variable, was created. The created dataset was labeled as hypotensive and non-hypotensive and divided into 80%-10%-10% training-validation-test data to train the deep learning models. During model training, important features were selected using the select-K-best method, and ridge regression was applied to prevent overfitting. The developed models included eight different deep learning models from three different deep learning methods (FCNN, 1D-CNN, and LSTM) that could predict hypotension after spinal anesthesia. Approximately 4000 experiments were conducted with different feature combinations for the eight different models. The experiment with the best AUROC for each model was accepted as the best result for that model. These eight models were then compared, and the model with the best AUROC was found to be the 50-Feature experiment of FCNN-2, one of the Fully Connected Neural Network Models (AUROC=0.6883). Additionally, the dataset was classified as hypotensive and non-hypotensive, and risk factors effective in hypotension were determined using classical statistical tests. All features that statistically created a risk of hypotension in classical statistical tests were included in the 50 features. Being over 30 years old (p=0.036), having experienced hypotension in previous cesareans (p=<0.001), a decrease in the amount of IV fluid given from hospital admission (p=0.042), a decrease in ESR value (p=0.033), an increase in CRP value (p=0.049), and the presence of DM (p=0.043) were risk factors for hypotension, while the use of iron preparations (p=0.004) and multivitamin use (p=0.003) emerged as protective factors against hypotension. In terms of model importance, in addition to these features, an increase in MCV, ferritin, and D-dimer values, the presence of PCOS and hypertension, and an increase in fasting duration were identified as important risk factors for the development of hypotension. Apart from the dose of bupivacaine used, no feature related to the application of spinal anesthesia was found to be significant for hypotension. Additionally, no correlation was found between patients' preoperative syndecan-1 levels and the frequency of hypotension after spinal anesthesia (p=0.675). In conclusion, the most successful deep learning algorithm that predicts hypotension after spinal anesthesia, based only on preoperative features, achieved a performance of 68% AUROC. Achieving this performance with only preoperative features makes its clinical practice use both easy and allows anesthesia doctors to early identify patients who will develop hypotension after spinal anesthesia.
