Computer Aided Detection System for Prediction of the Malaise during Hemodialysis

Abstract

Monitoring of dialysis sessions is crucial as different stress factors can yield suffering or critical situations. Specialized personnel is usually required for the administration of this medical treatment; nevertheless, subjects whose clinical status can be considered stable require different monitoring strategies when compared with subjects with critical clinical conditions. In this case domiciliary treatment or monitoring can substantially improve the quality of life of patients undergoing dialysis. In this work, we present a Computer Aided Detection (CAD) system for the telemonitoring of patients' clinical parameters. The CAD was mainly designed to predict the insurgence of critical events; it consisted of two Random Forest (RF) classifiers: the first one (RF1) predicting the onset of any malaise one hour after the treatment start and the second one (RF2) again two hours later. The developed system shows an accurate classification performance in terms of both sensitivity and specificity. The specificity in the identification of nonsymptomatic sessions and the sensitivity in the identification of symptomatic sessions for RF2 are equal to 86.60% and 71.40%, respectively, thus suggesting the CAD as an effective tool to support expert nephrologists in telemonitoring the patients.

Figure 1

Timeline flowchart of the monitored hemodialysis session. The patients' monitoring yielding the measurement of features after 60 and 180 minutes is emphasized; in particular two Random Forests (RF) are used to detect the session trend and eventually a malaise warning.

Figure 2

The figure shows a comprehensive overview of the training (blue box) and the prediction algorithm (yellow box).

Figure 3

The AUC comparison of both classifiers for different values of the proportionality factor k of the standard deviation in the Gaussian noise.

Figure 4

ROC curve of the first (a) and the second (b) classifier.