Online Multi-Agent Monitoring in Intensive Care Units
Kaouther Nouira · HAL (Le Centre pour la Communication Scientifique Directe) · 2010
The thesis work focuses on automatic and intelligent identification of patient status in intensive care units (ICUs). The objective was to propose an intelligent system capable of monitoring in real time, patients in critical condition by limiting the number of false alarms.Patients in ICUs, suffering from various pathologies, require assistance and intense follow-up with special care, which justifies continuous and real-time monitoring. Improving patient monitoring in ICUs is at the heart of much research and all the studies on this subject combine to demonstrate the need to improve medical monitoring tools.The current monitoring system is based on threshold algorithms. Currently, the alarms triggered reflect a simple crossing of a threshold by one or more measured variables. This system introduces noise pollution instead of informing the medical staff in a relevant way of real clinical events.Studies have shown that the rate of non-significant alarms triggered varies between 40% and 90% of all alarms. This can lead to a decrease in the vigilance of the nursing staff. This leads the medical staff, in most cases, to deactivate these alarms or to increase the thresholds. This behavior leads to an increase in the mortality rate in ICUs.The major drawback of these current systems is that they do not take into consideration the correlation between physiological variables. If a variable exceeds the threshold, the system triggers an alarm without checking whether the other variables with which it is correlated have changed their behavior or not. We have therefore proposed a multi-agent system where each agent monitors a variable. In the event of an anomaly, the agent sends a message to the supervisor agent who will decide whether to trigger the alarm based on messages from other agents.A second drawback of these systems is that they only consider exceeding the variable threshold and not the change in behavior in general (i.e. change in variance, change in level, slope, seasonality, etc.). This was remedied by the use of time series technology.To this end, we have proposed a new intelligent monitoring system based on time series techniques and multi-agent systems. The idea was to replace the cognitive part of the agent with an anomaly detection algorithm. The model used was the ARIMA model. Being an autoregressive prediction model, we used it to predict the new value before its appearance. Once captured by the system, the actual value will be compared against the predicted value. There are two cases: (1) if the actual value belongs to the confidence interval of the predicted value, then it is normal, (2) if the actual value does not belong to the confidence interval of the predicted value , then it is abnormal.To remedy the problem of outliers, which are essentially due to artifacts, the agent only sends an alert message to the supervisor if at least five successive values have exceeded the predicted values.The major advantage of this architecture is that it does not require any learning time. From the first 60 values captured (one value every 3 ms), the model can be initialized. As values arrive, the model will be updated.