In intensive care units (ICUs), medical monitors trigger an alarm when physiological parameters exceed or fall below certain thresholds. For example, when the patient’s heart rate goes below 60 bpm, an alarm will be triggered so medical staff is informed and can take action. The problem with this approach is, that there are so many alarms at an ICU, that the staff becomes fatigued and desensitized, hence not responding properly to every alarm.

This project aims to predict alarms from trends in the physiological parameters. Thereby, alarms – which are supposed to indicate acutely critical conditions – can be replaced by scheduled tasks. Instead of triggering an alarm when the heart rate is already below 60 bpm, we can detect that the heart rate will probably fall below 60 bpm during the next hour and notify the medical staff to look after the patient when they have time. Thus, minimizing stress in staff and critical conditions in patients.

To this end, we use large clinical databases such as MIMIC-III and HiRID. These databases contain all events, measurements, and patient records gathered in an ICU in a certain timeframe. This allows to perform time-series forecasting on patient data using machine learning models such as Recurrent Neural Networks (RNNs) and Hidden Markov Models (HMMs).