Personal tools

You are here: Home / Research / Biomedical Signal Processing & Instrumentation / Dr. G Clifford / ICU Decision Support

ICU Decision Support

ICU Event Prediction

This project aims to develop novel prediction algorithms in the Intensive Care Unit. Together with Cerner Corporation, we have developed prediction models for a wide range of scenarios on the largest available databases of both public and private data.

In 2012 our group won the Physionet/Computing in Cardiology Challenge for mortality prediction. Our algorithms are being licensed and ported to commercial devices.


Alistair Johnson (DPhil)
Nic Dunkley (DPhil)
Andrew Kramer (Supervisor)
Gari Clifford (Supervisor)

False Alarm Reduction

Over the past two decades, high false alarm (FA) rates have remained an important yet unresolved concern in the Intensive Care Unit (ICU). High FA rates lead to desensitization of the attending staff to such warnings, with associated slowing in response times and detrimental decreases in the quality of care for the patient.

Atzema et al (Journal of Emergency Medicine 2006, No 24, p62) found that less than 2% of all alarms require the clinician to take action to reverse the patient’s condition. Many published articles mention that ECG and impedance respiration are the primary sources of most alarms. False arrhythmia and heart rate alarms can usually be attributed to the use of a single ECG lead for analysis and also because information from other independent signals is not used to confirm or suppress calls based solely on information gleaned from the ECG.

A study published by Dr Clifford's team while at MIT (J Biomed Inform. 2008 Jun;41(3):442-51), based on data from the MIMIC II database, indicates that, on average, 43% (min 23%, max 91%) of life threatening alarms are false calls. Fusing information available in the invasive arterial blood pressure (ABP) signal resulted in a 60% reduction in false extreme tachycardia, extreme bradycardia, ventricular tachycardia (VT), ventricular fibrillation (VF), and asystole alarms. However the report also noted that the number of alarms studied (around 300 to 1000 of each type) was insufficient to build a prospective device.

Moreover, extensions to using the system in the absence of arterial lines is needed (since at least 1/3 of patients in the ICU do are not monitored for blood pressure continuously). This study seeks to address these shortcomings and, in conjunction with an industrial partner, allow for the development of a real-time robust monitoring system with greatly improved false alarm rates. Our research is funded by Mindray North America who license our technology.


Qia Li (Postdoc)
Gari Clifford (Supervisor)

Bayesian ECG

This project extends work originally performed in Oxford in 2001 to develop a Bayesian model of the electrocardiogram. Since then, our group has applied the model to a wide array of problems, ranging from TWA analysis, to QT interval measurement and separation of the magnetohydrodynamic effect.

We are also developing novel Bayesian voting methods for improving labeling of medical data (in particular ECGs) and for classifying data quality and cardiac rhythms.


Tingting Zhu (DPhil)
Joachim Behar (DPhil)
Roberta Colloca (DPhil)
Johannes Krug (PhD)
Julien Oster (Postodoc)
Gari Clifford (Supervisor)