1 INTRODUCTION Frequent or occasional symptomatic intradialytic hy- potension (IDH) may influence patient well-being. In a recent study mortality in patients with frequent IDH was significantly higher than in those without such events (1). However, these authors failed to show an independent effect of frequent or occasional episodes of IDH on mortality after adjustments for covariates. Symptomatic IDH episodes requiring reverse Trende- lenburg positioning, a pause in the ultrafiltration and/or infusion of isotonic saline and/or hypertonic glucose are difficult to predict and therefore to pre- vent. To avoid intra-dialytic hypotension sodium profiling, blood temperature and on-line blood volume moni- toring and, in patients showing a significant peri- and post-dialysis alkalemia, a reduction in bicarbonate concentration of the dialysate have been suggested (2- 4). However, even applying these classical prevention tools a substantial number of dialysis patients will con- tinue to show abrupt and sometimes severe IDH episodes (2-4). The ability to predict IDH is probably related to the nephrologist’s experience and on the characteristics of the individual patient, but this has not been systematically analysed. There is a large amount of clinical and biochemical data collected during each dialysis session that could be helpful in predicting IDH, but their interpretation, even using statistical models as multivariate logistic re- gressions is difficult. The limitations presented by the experience/intuition and statistical procedures are mainly due to the non-linearity and multidimension- ality of the problems analysed. In a previous study we JNEPHROL 2005; 18: O O RIGINAL RIGINAL INVESTIGA INVESTIGA TION TION www.sin-italy.org/jnonline/vol18n4/ Predicting intradialytic hypotension from experience, statistical models and artificial neural networks Luca Gabutti 1 , Martine Machacek 1 , Claudio Marone 1 , Paolo Ferrari 2 1 Renal Services, Ente Ospedaliero Cantonale, Bellinzona, Switzerland 2 Department of Renal Medicine, Fremantle Hospital, Perth, Australia ABSTRACT: Background: Symptomatic intradialytic hypotension (IDH) associated with increased mortality in he- modialysis patients is difficult to predict and hence prevent. Artificial Neural Networks (ANNs) are promising tools to solve multidimensional non-linear problems. The aim of the study was to verify in which way mathematical models, statistics or knowledge of patients influence the ability of the nephrologists to predict IDH. Methods: The performance of ANNs was compared with that of independent nephrologists supported by a logistic re- gression giving odds ratio for each studied variable (NEPHiS) or of nephrologists in charge of the patients without (NEPHc) or with statistical support as for NEPHiS (NEPHcS). Data from 98 hemodialysis patients were analysed in order to select patients with frequent IDH (>10% of the dialysis sessions). Complete data on 1979 dialysis sessions from 7 patients were retrieved. The ability to predict the occurrence of hypotension episodes was compared (ROC curves) between ANNs, NEPHc/S (N=7) in Switzerland and NEPHiS from independent dialysis centers in Western Australia (N=10). Results: ANN gave the most accurate correlation between estimated and observed IHD episodes compared to NEPHc (p<0.001), but a similar performance was attained by NEPHcS (p<0.001). NEPHiS were superior to NEPHc (P<0.05), but inferior to ANN (P<0.01). For a sensitivity of 80%, specificity was 44% for ANNs, 33% for NEPHcS and 20% for NEPHc. Conclusions: ANNs are superior to nephrologists in predicting IDH episodes; however when supported by a statistical analysis, nephrologists reach ANNs in their prediction ability. IDH still remains difficult to predict even with mathe- matical models. Key words: Artificial neural networks, Hemodialysis, Prediction, Dialysis-related hypotension