Contagion® Connect Episode 3: Can AI Cut Antimicrobial Resistance?

October 16, 2019
Contagion&reg Editorial Staff

Paul Abrams, MBA, director of Health Care Specialty Group at MJH Life Sciences, spoke to Vaibhav Singh from Circle of Life about Zevac, a tool that aims to help clinicians make more informed prescribing decisions.

Welcome to Contagion® Connect. This new podcast will bring you expert perspectives on trending infectious disease topics.

In this episode Paul Abrams, MBA, director of Health Care Specialty Group at MJH Life Sciences, spoke to Vaibhav Singh from Circle of Life about Zevac, a tool that aims to help clinicians make more informed prescribing decisions.

Paul Abrams: Can you tell us a little bit about your product?

Vaibhav Singh: So, we are an AI software product company, which is using advanced machine learning to predict culture results at the zero hour. The idea is that when a clinician suspects an infection, they send the relevant sample to the lab for processing the sample. And then for creating a culture or susceptibility report, our machine-learning tool uses data from the past and triangulates it with certain features of the patient and instantly predicts a proxy culture report or a personalized antibiogram for the patient, so that the frontline clinician is nudged into the right direction as to which drug might be right for this patient at this time.

The idea is that as per CDC data, 1 in 2 patients admitted in a hospital in the US is administered an antibiotic and 1 out of those 3 antibiotic prescriptions are either in inaccurate or inappropriate. Therefore, there is a huge problem of antimicrobial resistance.

What this product does best is that by giving you an insight into what could be the right drug to give at a certain point in time, and by looking into the future and kind of predicting the culture, it helps you undertake the right therapy for the patient.

Paul Abrams: So rather than making it a guessing game or trading up and down, you can make a better clinical decision by using data, machine learning and AI to say, “Hey, this patient is at the hospital and they have an outbreak of MRSA,” now you know what the best combination of therapy and antibiotic is, whether it's broad spectrum or isolated, to treat that patient and hopefully stop that infection.

Vaibhav Singh: So, the idea is that there is a lot of information which the data holds, but it's not available at the point of care for the benefit of the clinician and for the patient. The beauty of this product is that for one, with a very simple interface with the extent hospital management information system and the lab management information system, at the press of a button, the clinician effectively gets a personalized, predicted antibiogram at the level of the specific patient, as if pulling out a page of the institution-level, cumulative anti bio gram, which is generally prepared at a periodicity of 6 to 12 months by AMSD or even the hospital right now. This is extremely inefficient because those antibiograms at the level of the institution are not accessible at the point of care, they are very generic, they are broad, 3 by 4 matrices at best, and the clinician typically does not use them.

Now, with machine learning at our disposal, the machine can actually go in and triangulate for the specific features of the patient, what would be the right combination of drug susceptibility, which would provide the best therapy, and therefore it's a revolution, which is waiting to happen now in antibiotic therapy.

The best part is that the software is non-intrusive and extremely cheap because what we are simply doing is using raw computing power of the machine to go into past data. You're not doing any costly, rapid diagnostics. You're not getting into the business of changing any workflow for the clinician. I think huge benefits can come from it.

Paul Abrams: So, the friction that would be there when adapting to new technology is not there, so that's huge. Uptake is usually delayed by friction in anything you do, whether you're buying something online — if I have to go through 5 extra steps to buy a product, if I can do it by asking Alexa to buy me something and it shows up at my house, it’s the same exact thing. You can say, “Okay, we know who this patient is,” and enter their demographics.

Vaibhav Singh: You don't even need to do that, actually. Because of the high-level interfacing through an API that exists for our software, with the systems inside the place in the hospital — generally Epic, Cerner, Allscripts — you simply press the UHID or enter the identifier of the patient and everything else get pulled forward from the data. The machine-learning algorithm works on the back and presents the output in an instant — less than two seconds.

Paul Abrams: Can you tell me some of the data points you have to show the benefits and efficacy of your tool?

Vaibhav Singh: We’ve set our research and development center in one of the top 10 engineering schools in India because of the fantastic availability of privileged patient information on the back of which we could build — it’s very open platform. Second, we had great machine-learning talent there. We've used feature sets or explanatory variables in the model, which include a mix of patient level variables, like demographics, age, gender and what community they are coming from, because infections are either community acquired, or hospital acquired.

Then you have features like the co-morbidities for the patient, device insertions, which part of the hospital they are in, what kind of care type. Then there are certain other weather-related factors, because drug bug susceptibilities vary according to season. We also have some other factors around what past antibiotics were administered on the patient, immunocompromised status of the patient, if at all. All of those go into the mix and this is done through an extensive pre-training schedule of the models where we evaluated more than 600 variables before we hone down on a parsimonious model, which takes into account such factors which are available at the point of care and are easily accessible or in-portable into the system. That number is no more than 20-25. By using those variables, we’re able to predict with almost 93% accuracy the real ground truth, which is a cultural report.

Paul Abrams: That's huge, right? So, you can really change the way decisions are made in hospitals and clinics so that clinicians themselves are making better and more informed decision. They don’t have to worry about antibiotic resistant strains at all, right?

Vaibhav Singh: Not at all. And I think the benefits are multifaceted. You have straight benefits for the hospital because antibiotic costs come down, readmission rates, infection rates and length of sheer optimization revenue for the hospital goes up, because now you don't have patients who are sitting using a bed merely recovering from a wrong drug given to them two days back. Now they get the right drug at the right time and then you can actually send them home quickly and you can get the next patient in and increase your revenues.

Also, it's fantastic for patient outcomes, patient satisfaction, fighting EHR resistance and it’s great for payers because now they're not paying for an unnecessary stay.

We’re very excited with the large opportunity, especially considering that infections are massively pervasive — they're here to stay.

I think for the next several years, the utilitarian value of something like this only increases because the quality and integrity of the data improves. So, as we get more and more data, the bugs are getting smarter. So, it's incumbent upon us to fight them more intelligently. We have to show them that we are the smartest species on the planet. They can mutate to beat us, but we have to keep coming up with novel ways or new guerrilla warfare to fight.

So, I think this is our start point. Infections are just the start.

Paul Abrams: Where do you see the pivot for something like this?

Vaibhav Singh: I think there's a huge opportunity just in in what we've built at present, but I think as we move ahead, different kinds of virus infections and different kinds of bacterial complications are on our radar. We also want to get into the business of predicting with a high degree of accuracy, propensity for a patient to acquire an infection inside of a hospital.

As we all know, the biggest adverse event of health care is health care. So, therefore, it's extremely important that we could put to good use the power of the information that rests within the data. Now with great democratization of information and machine-learning capability, technology and computing power, the next battle needs to be fought with data, not with new medicine.

Paul Abrams: This is a small example, but it’s like not using as much Purell and just washing your hands with soap and water. It’s just as powerful and the difference with Purell is that the alcohol smells and it's killing the cell walls, but again, it's also making bacteria stronger. This is another way where you're using the data that you have at-hand in the algorithms to make more informed decisions, which is really what we need in health care. That's where — if you look at the technology with robotics, AI and big data, and IoT and how it can affect healthcare — it’s huge.

Vaibhav Singh: I think that's that was the inspiration behind us on commencing on this journey and we are very happy with the response that we've received globally. We’ve been in the Singaporean market, the Middle East market, we've been in Asia, in India where we built the product and now in the US.

The problem is pervasive globally, it's of equal magnitude, if not worse, in every part of the world. Some markets will have to subsidize some other markets but I think still the opportunity is immense and that's what excites us.

Thanks for tuning in to this Contagion Connect podcast. Be sure to share, like and subscribe. And for the latest infectious disease news, be sure to check out