When An AI Can Call Code Blue

Artificial intelligence can help spot the early signs of deteriorating patients in hospital wards, but its application is yet to be tested.

By Sherif Gonem, University of Nottingham

NOTTINGHAM, Sept 15 – Monitoring vital signs — heart rates, respiratory rates, blood pressure, and temperature — can give nurses and doctors a pretty good idea of how unwell someone is. But researchers began noticing something was going wrong in United Kingdom hospital wards — crucial signs of patient deterioration were being missed. 

Researchers at the University of Nottingham are developing an Artificial Intelligence (AI) algorithm that can take account of the vital signs history of a patient, including the trend over time and the degree of variability. Tools like these can help medical staff spot early signs of deteriorating patients and treat life-threatening conditions before it’s too late.

Reviewing the charts from patients who have suffered a cardiac arrest often reveals a pattern of worsening vital signs in the hours beforehand. It was soon recognised that combinations of abnormal vital signs were associated with a higher risk than a single abnormal sign.

Which is why Early Warning Scores are largely used to assign a number to each vital sign and added together to give a total score. For example, one widely used score in the UK is the National Early Warning Score-2 (NEWS-2) — if you score five or more you should get an urgent medical review, seven or more and the medical team will come running.

But there is still a problem — even the best scoring systems produce a large number of false alarms. This can lead to doctors being called unnecessarily, potentially diverting them from seeing a truly unwell patient.

Even worse, repeated false alarms can lead to a psychological phenomenon known as alarm fatigue, when clinical staff stop taking the alarms seriously due to the ‘crying wolf’ effect. The scores being used can also miss subtle signs of declining health. For instance, if a vital sign is moving in the wrong direction but still within the normal range. 

The problem with current early warning scores is that they are too simplistic. They needed to be twenty years ago, when figures were calculated by hand on paper charts. But now many hospitals are moving to electronic recording of vital signs.

The large datasets of vital signs recorded within these systems can be used to train and validate AI algorithms to detect the deterioration of a patient’s health more effectively. Studies on how to best utilise AI in these circumstances have also been done by groups in Oxford and Portsmouth and Cambridge.

In both of these cases, the AI algorithms showed improved accuracy compared to previously published scores.

But proving that an algorithm works on paper or within a computer processor is not the end of the story. Real-world clinical trials are needed to see whether more accurate AI algorithms will be beneficial for patients. 

In one recently published study, researchers in the United States introduced a system of alerts, driven by a bespoke machine learning-based scoring system, to a network of 19 hospitals in Northern California.

The investigators found a reduction in deaths after using the system, but it’s not clear whether there would be the same results with simpler pen and paper scores since a direct comparison was not done. It may seem obvious that a more accurate AI-based score will be beneficial, but this remains to be proved. 

We know that nurses do not follow early warning scores blindly — they use their clinical judgement when deciding whether to call a doctor. So perhaps human intuition is already filling in the gaps left by the simple early warning scores, and a more complex AI score is not needed. The bottom line is we need to do the trials to find out. 

Implementing AI solutions in a clinical setting comes with other challenges. The new algorithms need to be integrated with existing electronic health record systems, which can pose technical problems. More importantly, the output of AI algorithms needs to be transparent and interpretable to be trusted by clinicians. 

University of Nottingham researchers are currently developing a module that will provide a verbal or graphical explanation for the output of their AI algorithm. If doctors and nurses can see why a patient has been given a high-risk score, they can make a more informed judgement as to how urgently that patient needs to be seen. One explainable prediction model was recently developed by researchers in Aarhus, Denmark.

There are other exciting developments on the horizon. Wearable sensors have been developed that can monitor vital signs continuously rather than intermittently.

These could provide earlier detection of deteriorating patient health by picking up changes in vital signs immediately, rather than waiting for the next scheduled set of clinical observations. AI algorithms may be needed to handle the continuous streams of data produced by these devices. 

Nottingham University Hospitals NHS Trust is undertaking the first large-scale UK pilot of a wearable patch for continuously monitoring respiratory rates, through funding from the digital arm of the National Health Service. The results are expected in late 2023. 

Researchers anticipate that AI and wearable devices will help keep patients safe in hospital wards within the next five to ten years. These novel technologies will act as extra eyes and ears, but they will complement rather than replace the clinical judgement of health care professionals.

Dr Sherif Gonem is an honorary assistant professor at the University of Nottingham.

Article courtesy of 360info.

You may also like