A team led by IIITD-AIIMS researchers has developed a machine-learning algorithm to detect and predict shock using a single photo captured through thermal imaging. It has the ability to predict shock with 69% accuracy 12 hours before it can be clinically recognised by doctors using the current gold standard. The researchers plan to deploy it for clinical use in the paediatric ICU in AIIMS within six months.
Predicting shock (less blood and oxygen supply to major organs leading to death in some cases) even 12 hours before it can be clinically recognised by doctors using the current gold standard (intra-arterial blood pressure) is now possible, thanks to the work by an AIIMS-led multi-institutional team of researchers. Shock can arise from loss of blood volume, inefficient pumping of the heart or infection (sepsis).
The machine-learning algorithm to detect shock at the time a single photo is taken using thermal imaging has an accuracy of 75%. The ability of the algorithm to forecast the probability of a shock happening three, six and 12 hours before clinical recognition can be done using the gold standard method is 77%, 69% and 69% respectively. The algorithm was used in conjunction with pulse rate to both detect and predict shock. The results were published in the journal Scientific Reports.
In paediatric intensive care units, 70-90% babies develop signs of sepsis. Almost 30% paediatric ICU patients suffer from sepsis shock and 30% of them end up dying due to multiorgan failure. “This number will be much higher at district hospitals. Sepsis shock is one major killer in paediatric ICUs,” says Dr. Tavpritesh Sethi from the Department of Paediatrics at AIIMS, New Delhi, who led the team. In principle, the model can be used for predicting shock in adult patients too. But the model has to be tested on adults as the current study was limited to 539 thermal images of paediatric patients.
It is possible to prevent organ failure and death by adopting simple measures such as fluid management through transfusion and raising the blood pressure using certain drugs. Body starts responding to shock very quickly but takes some time for clinical recognition. This is where the machine-learning algorithm comes handy in saving lives with its ability to detect and predict shock.
“Due to noise, thermal images are fuzzy and so it is difficult for the computer to identify body parts. So the machine-learning algorithm was trained to identify body parts, capture body surface temperature and calculate the temperature difference between abdomen and feet to detect and predict shock,” says Aditya Nagori from the Institute of Genomics and Integrative Biology (CSIR-IGIB), New Delhi and first author of the paper.
How it works
When there is less blood and oxygen supply, blood starts to flow away from the hand and feet to important organs such as heart and brain. As a result, the temperature of the limbs falls compared with chest and abdomen. “This temperature difference is not measured all the time even in an intensive care unit. And there is human judgement as well to call it a shock. But ours is a quantitative method without human judgement to measure the temperature difference between peripheral and centre of the body and predict shock,” says Dr. Sethi, who is a Fellow of the Wellcome Trust DBT India Alliance. “Most importantly, ours is a non-contact, non-invasive method.”
Just like temperature difference between peripheral and centre of the body, the pulse rate too increases when shock sets in. The increase in pulse rate happens in response to reduced blood and oxygen supply. The heart starts beating faster to circulate the smaller volume of blood leading to higher pulse rate.
Since the current study used a single image to detect and predict shock, images have to be taken at regular intervals for continuous monitoring. “We are expanding the scope of the work to use video for continuous monitoring,” Dr. Sethi says.
Since a single image along with pulse rate is all that is required for detecting and predicting shock, children admitted in ICUs at remote locations can be monitored using the model as a tele-diagnostic decision support system. “We are expanding the scope of the work to track patients remotely at the district level hospitals and primary heath centres. We hope to start this before the end of the year,” he says. “Clinical use of the model in the safe ICU at AIIMS will start within the next six months.”
“The team is excited to launch a smartphone application which will incorporate the model capability to predict shock,” he adds.
The researchers were able to use the machine-learning algorithm to detect difference in body temperature at AIIMS once the safe ICU with big data warehousing, the first of its kind in India, started functioning since February 2016. “Here data of every patient in the paediatric ICU at AIIMS is captured every second,” Dr. Sethi says.