Harnessing AI to generate patterns of antibiotic resistance in real time

“There is a shared mechanism of resistance between antibiotics, we already know. Usually to do that, people use genomics, but that’s an expensive proposition,” explains Dr. Sethi. “We have proposed a way, which is inexpensive, because it uses these routine data sets from hospitals. We show that by using routine data effectively, we can discern relationships between different antibiotics pairs and the direction AMR is taking – whether it is rising or not. Say, for instance, if resistance to one specific antibiotic is going up, some months down the line, it is quite likely that resistance to an antibiotic pair might also shoot up. With these connections, we generated actionable pieces of evidence.” 

Dr. Sethi adds: “We have tried to go beyond the traditional way of looking at AI – asking how can it enable better decision-making for a given patient in a clinical setting or a public health setting. We think AI can also be used to understand AMR stewardship and surveillance aspects, from the hospital level, upwards. Hospitals already routinely send out patient isolates, for example, blood, sputum, urine, pus, etc., for culture sensitivity testing in order to make informed decisions on treatment courses. We are saying that these reports can be used to construct AI-based pipelines and methods that can lead to AI-driven or AI-enhanced antimicrobial stewardship.” 

The AMROrbit Scorecard that the team developed also won an award at the 2024 AMR Surveillance Data Challenge. Can we use these scorecards to make it more timely? Dr. Sethi explains: “It plots the orbit of resistance, say of every hospital or department, alongside a global median of resistance and a global rate of change. So around those global values, how well does a department, a hospital, or a certain country fare? That is what the scorecard will be able to provide real time data for.”  

The ideal quadrant for any hospital or country to be in is where there is low baseline resistance and low rate of change as well, explains Jasmine Kaur, of IIIT-D, and lead author of the paper. Orbits spiral in or out, but the AI tool can offer information facilitating timely interventions that can bring it to a desirable range of resistance.  

How accurate and reliable are these AI models? “In our paper, we have shown that our models did capture the trends as observed in the period we collected data for. However, unless we have future data, we can’t really say, like, for example COVID- 19 upended things, right? The only evidence we have currently is that globally it seems that our models are capturing the increasing rate of resistance in various studies.” 

Clinicians can make informed decisions based on the visual image that OMROrbit provides them using the data generated by the hospital, explains Ms. Kaur. It has been proven that it can augment ongoing surveillance at various levels. Various kinds of comparisons can be done using the tool, she adds. For instance, if it is a chain of hospitals, then the tool can be used to compare AMR rates between different departments, cities and centres across the country. “The only possible limitation would be in circumstances and settings that do not have consistent, granular surveillance data. Then the AI model will not make sense. This could occur in countries where surveillance data is not digitally accessible.,” she adds. 

“We know there are other environmental factors such as anibiotics being used as growth factors in the poultry industry or leachates in the soil, that can also lead to AMR. The ideal would be, if at the public health level, we should be able to use the data we have from the hospitals, matching it with antibiotic sales, and community-level data, and study the environmental factors too. We hope to do that soon, Dr. Sethi explains.