FC095: Development and Validation of an Ai-Based Triage Algorithm for Intensified Prevention of Intradialytic Hypotension

Abstract

Abstract BACKGROUND AND AIMS Complications during hemodialysis (HD) sessions remain one of the main challenges in the dialytic patients’ management. Intradialytic hypotension (IDH) is among the most frequent adverse events in HD treatments. However, the IDH incidence varies widely across studies from 7.5% up to 68% of HD sessions, because of the lack of a univocal IDH definition. Besides, inadequate IDH management can lead to adverse outcome including increased mortality risk, hospitalization and vascular access thrombosis. IDH can be partially prevented by improved hydration management, as well as fine tuning of dialysis and pharmacological prescriptions. Nevertheless, a personalized approach to IDH prevention is still challenging. Therefore, new, cutting-edge approaches to predict these events is of critical importance to advance dialysis practice. We sought to develop and validate an AI-based classification model to predict the risk of IDH in the next HD treatment. METHOD The IDH was defined as any treatments with IDH symptoms reported in the clinical chart with a fall of ≥20 mmHg in systolic blood pressure (SBP) from pre-dialysis to nadir intradialytic levels plus ≥2 responsive measures. The aim of IDH model is to predict the risk of IDH in the next HD session exploiting a high-performance machine learning classifier called XGBoost. A feature selection strategy based on a tradeoff between prediction accuracy and model complexity was performed to determine the set of input variables. All the model's variables are routinely collected in clinical practice or they are automatically recorded from dialysis machine: no additional effort is required by the clinical staff to use the risk model. The accuracy of the prediction was evaluated as the average of the Area Under the ROC Curve (AUC-ROC) on 10 different random sampling of the dataset. In each randomization 70% of the dataset has been used for training and 30% for testing. The appropriateness of the model was also assessed computing the Risk Rate Ratio of the incidence of the IDH in 4 risk classes. RESULTS The final dataset comprised 8217 adult patients for a total of 1.7 million of HD sessions collected in EuCliD database by Fresenius Medical Care (FMC) clinics in Portugal and Spain between January 2018 and November 2021. The IDH incidence was of 10%. Baseline patients’ characteristics are the following: male: n = 5361 (65%); age (year): 67.93 ± 14.39; dialysis vintage (days): 752.58 ± 1742.7; BMI: 27.08 ± 5.76; and 3532 (43%) patients affected by diabetes. The average of AUC-ROCs was 0.80 [95% confidence interval (95% CI) 0.79–0.81]. Variables associated with the greatest impact on risk estimates were different metrics on intra-dialytic blood pressures, fluid overload, treatment time, clinic IDH rate and use of diuretics (Figure 1). We defined three thresholds determining four distinct risk groups, ranging from low-risk patients to high-risk patients. The relative Risk Rate Ratio for each class is reported in Table 1. CONCLUSION The developed AI model showed a good discrimination and excellent calibration. Moreover, the definition of four risk groups together with new continuous and interactive monitoring protocols, will promote the application of new personalized digital therapeutic program to prevent IDH. The great advantage of this strategy is that the triad ‘patient-nurse-physician’ will tightly cooperate to prevent IDH achieving a concrete improvement of the patient's quality of life.