Alarms Early Detection in Dialytic Therapies via Machine Learning Models

Abstract

Hemodialysis (HD) is a clinical treatment for patients affected by Chronic Kidney Disease (CKD). The goal of a treatment is to purify the patient’s blood using dialysis machines, devices that act as artificial kidneys. However, a common problem is the alteration of the patient’s health status due to side effects or to machine malfunctions that may occur during treatment. A dialysis machine is a complex apparatus consisting of a control system of several quantities (e.g., pressure, flow rate, temperature, conductivity, etc.) capable of alerting medical operators when an alarm occurs. In the present work, a Machine Learning (ML) predictive model able to act in advance with respect to the dialysis alarm system was developed. Several machine learning models were tested and a comparison study was carried out. Datasets for training and testing the models came from treatments performed by dialysis machines manufactured by Mozarc Medical®. Among the models tested, the Random Forest (RF) classifier was identified as the more promising one and was then used to perform a parametric sensitivity study. By using a time window of 10 seconds, the RF model provided a Recall of 79% and an F1-Score of up to 85% on test data, demonstrating the good generalization ability that is always required by predictive models such as the one analysed in this paper.