MO913: Exploring Factors Influencing Intradialytic Hypotension Using Deep Learning

Abstract

Abstract BACKGROUND AND AIMS Intradialytic hypotension (IDH) is an independent risk factor affecting the life prognosis in patients undergoing haemodialysis. Various studies using statistical methods have been conducted to identify the factors that cause IDH, but the factors and the interactions among the factors have not yet been clearly elucidated. In this study, we used a neural network model (deep learning) using factors that can be measured in real time as input parameters to identify factors strongly influencing the risk of IDH, with the ultimate goal of predicting IHD using only factors that can be measured in real time. METHOD A total of 25 parameters obtained during dialysis treatment in 208 patients were selected as the input parameters for deep learning; the 25 parameters included 18 items selected from the patient background and vital data (including the dialysis time, treatment mode, systolic blood pressure, diastolic blood pressure and mean blood pressure), and seven items calculated from the above data (including the difference between the actual and planned dialysis times and the difference in body weight before and after dialysis). As the evaluation indices, we used F-measure, which is calculated from precision and recall, and attention weight, which indicates the % influence of an input parameter on deep learning. We compared the F-measure and attention. RESULTS In the learning using all the input parameters, the correct response rate was 69.6%, the recall rate was 64.2%, the fit rate was 41.7% and the F-measure was 41.7%. The highest value of attention weight among the input parameters was 21.1% for the occurrence of IDH in the most recent treatment. Even in the pattern in which the highest F-measure was found, the attention weight of the occurrence of IDH in the most recent treatment was 21.9%. In all patterns, the top three attention-weighted items were the occurrence of IDH in the most recent treatment, the systolic blood pressure and the body weight before dialysis. The items with high values of the F-measure and high attention weight were considered essential factors for the prediction model. CONCLUSION By using deep learning to compare the accuracy of prediction of various input patterns, the occurrence of IDH in the most recent treatment, systolic blood pressure and body weight before dialysis were identified as the factors most strongly influencing the risk of IDH.