#3989 INTERACTIVE-HD 2.0 PROJECT – PRELIMINARY RESULTS OF LUGANO DATASET

Abstract

Abstract Background and Aims Computational models can be used in nephrology and dialysis divisions in order to predict the intradialytic trends of the main electrolytes, breakdown products, and body fluids volumes. However, predictive models need to be trained with consistent clinical data for a certain amount of HD sessions: if successfully trained, these models can be a useful support tool for a therapy customization, mainly needed for ESRD patients characterized by multiple comorbidities. Strength and limits of similar approaches have to be evaluated before their daily clinical use. Methods The multi-compartmental, multi-solute model optimized during the InterACTIVE-HD 2.0 study needs to be trained with one week's HD sessions, during which data machine and hourly blood compositions are given in input to the model itself. After the training, this model allows an accurate prediction of the patient's response using only the clinical data recorded at the beginning of the therapy, not only in the same sessions used for the model training (ID mode) but also in independent sessions (PRED mode). The preliminary results refer only to the center of Lugano, in which 25 patients were enrolled and whose sessions were monitored in consecutive phases over a 7-month span (for a total of n = 290 sessions); model performance (stratified by age, presence of diabetes and heart disease or hypertension) has been evaluated in terms of normalized Root Mean Squared Error defined as follow: where the subscript i stands for the ith of N intradialytic acquisition time-points, while CLIN and MOD refer to the clinically measured data of interest and to the model simulated ones, respectively; y is the variable of interest, i.e., electrolytes’ hematic concentrations. Inter-stratification variability in terms of model performance has been analysed via a Kruskal-Wallis test. Results From the results’ stratification, it can be observed a good ability of the model to describe all variables, without bias with respect to pathologies in the short term. From 2 months onwards, the presence of hypertension, alone (n = 23), or the presence of cardiovascular diseases, alone (n = 38), leads to an increase in the error. Error that is instead very limited in the main group of patients (n = 229). Regarding age groups’ stratification, all variables are well simulated in the short term; however, already at 2 months, it is highlighted an increase in the error only for patients under 70 years of age (n = 87). On the other hand, for patients older than 70 years (n = 203), the stratification shows more stability. Finally, regarding the stratification based on the presence of diabetes, it can be observed that glycaemia is no longer simulated well only in insulin-dependent patients (n = 75) compared to non-insulin-dependent patients (n = 89) around 5 months; in the case of insulin-dependent patients, for K+ and urea the model should take into account some other phenomenon to be modelled as they are tracked with lower precision already in the short term. Conclusions The stratification highlighted that hypertensive patients should be reevaluated more often, preferably two months after the first training of the model. Moreover, it was observed that the most sensitive group is the one with patients under 70 years of age, thus requiring re-evaluation, already at 2 months; on the other hand, for patients older than 70 years, in the absence of other pathologies, this re-evaluation can be postponed at least 7 months. Finally, the ability to simulate blood glucose is always acceptable in the 7 months, except for insulin-dependent patients in whom a re-evaluation between 2 and 5 months is necessary. Either way, given the general results and what the stratifications have evidenced, we claim the clinical effort is worth it and the model can represent an important support for clinical decisions.