Posted on March 9, 2020
Hospitals around the world continue to struggle with preventing and managing Clostridioides difficile (C. diff) infections. They are seeking better ways to predict which patients are at risk for C. diff infections (CDIs) sooner, and to improve prevention and treatment.
We sat down with three Wolters Kluwer experts—who have been working on developing, testing and integrating artificial intelligence (AI) tools into Sentri7—to discuss why AI is such an important advance for managing CDIs.
Matt: C. diff is an opportunistic infection, often healthcare-associated, that is debilitating for patients because it can cause symptoms that range from diarrhea to life-threatening inflammation of the colon. It’s a very hardy bacterium that is spore-forming and is very difficult to remove from the environment, dramatically increasing the risk of transmission.
C.diff is also somewhat difficult to treat because it tends to recur over time. And as hospitals know all too well, patients that do become infected experience sub-optimal outcomes. In the U.S. alone, there are about a half-million C. diff infections each year, responsible for about 29,000 deaths. To make things worse, C. diff infection increases the length of stay and treatment costs. In fact, the CDC estimates that C. diff is responsible for $4.8 billion in costs each year. Financial penalties are also in play because C. diff is monitored within hospital pay-for-performance programs. Those organizations that are not proactively focusing on preventing C. diff will typically find themselves among the poorest performers, and their bottom lines will suffer due to penalties. That’s why so much work goes into controlling and preventing CDIs.
Steve: From a pharmacy and antimicrobial stewardship perspective, C. diff is concerning because treatment options are limited to three antibiotics, and here is also the high risk of recurrence that Matt referenced. Even if we successfully treat the infection, the C. diff bacteria hangs around in the gut or the patient’s surroundings, waiting for the next flora disturbance, and we have to go back to the same three antibiotics. In the past 6-8 years, some people have gotten so desperate to address recurring infections that they’ve opted for a fecal microbiota transplantation. That has shown some promise, but the treatment is not without risk, especially for acquiring other types of infections.
Matt: We saw an opportunity to leverage AI, identify at-risk patients sooner, and allow clinicians the opportunity to address modifiable risk factors and proactively prevent C. diff infection using known, evidence-based prevention strategies. We know our hospital and health system clients are incredibly concerned with how C. diff tends to increase mortality risk, length of stay, and introduce potential for financial penalties and pressures. Our customers already rely on Sentri7 as a trusted clinical surveillance platform, and this enabled us to provide extended value across members of the care team while addressing this worrisome infection.
Steve: Well, there’s a consensus, for example, that some antibiotics, like fluoroquinolones, predispose patients to C. diff infections. If we identify those patients, the physician and pharmacist can assess whether they need that particular antibiotic and then change it, if possible, to lower-risk antibiotics. Another example is the proton pump inhibitor. It is also highly associated with C. diff infection. Again here, with more information, pharmacists can have a meaningful conversation with prescribers to do a better-informed risk-benefit analysis. We see AI as an opportunity to introduce many other risk factors and dimensions to the risk-benefit analysis so providers can make more informed clinical decisions.
John: I agree that the clinical concerns were the most important drivers behind our effort, and I would add, from that perspective, that early identification also enables you to isolate patients more promptly. We’ve statistically shown that when CDI happens in a hospital, it typically occurs in clusters. That is, as soon as the first patient contracts a C. diff infection, usually a number of other people in the same hospital area contract it as well. So if you can detect C. diff earlier, or even predict, you can not only begin treatment, but you can also isolate patients to try to prevent the spread and implement other necessary protocols. With early action, you can keep the infection to one person rather than five.
Matt: Clinical policies and protocols for environmental cleaning recognize that routine disinfectants are often not sporicidal. If the bacteria are on a bedrail, glove, telephone, etc., it may be transmitted to other patients or providers and may contribute to the clustering effect. At a high level, I view AI-enabled technologies for C. diff—and, ultimately, for other HAIs—as a way to bridge the prevention aspects of public health with the diagnosis and treatment components of clinical medicine. With CDI, we aim to address things earlier in the care process by looking upstream at hospitalized patients to determine if they’re at risk of developing CDI. Then, clinicians may reassess their choices of antimicrobial therapy or other medications that could increase that risk. That’s what I mean by bringing the prevention aspect to clinical medicine. We can create proactive workflows that target the prevention of adverse events.
John: The simplest way to put it is that AI deals with more than the binary risk indicators assessment that humans commonly perform in their heads. The reality is that we can only look at a small set of variables and how they might factor with the other. Thus, systems that merely assess binary risk indicators are brittle because they are so limited in terms of the complexity of situations they can address. In contrast, the power of AI is its ability to consider many different interaction effects. Working with our clinicians, we incorporate a tremendous number of inputs: lab results, white blood cell count, bilirubin, neutrophils, vital signs, medicine administration, concentrations and durations of meds, duration in hospital, demographics about the patients and demographics on a hospital—to name a few. The AI model can look at hundreds of these things – and the hundreds of correlations among them. To give just one simple example: We know older people are more at risk, and if they are on a particular broad-spectrum antibiotic, there can be a multiplicative effect – that is, the effect is greater than simply adding each of the factors independently. Now expand that to looking at hundreds of things. We’ve trained all of the statistics and variables and have shown that we can be incredibly precise. When you compare them head-to-head, machine learning and AI crush current risk assessment systems.
John: A few different things should ease those concerns. First, working with bedside providers in a clinical setting, we’ve shown we can identify a higher percentage of C. diff cases, predict them significantly sooner, and therefore, empower clinicians to respond both more quickly and more effectively.
In addition, one of the features of this model is that it provides a visual picture of its output that explains why it made the predictions it did: what features it identified as most important and why they put the patient at greater risk. For example, it might note it is predicting a CDI in a hospitalized older adult who has a high white blood cell count and fever lasting two days. This information allows physicians to examine the output and decide if the logic makes sense. If there are other considerations rooted in the provider’s clinical experience, he or she can always disregard the tool’s assessments and recommendations.
Finally, and maybe most importantly, there is the time component—the ability to track when all of these factors occur and understand how they figure into the various interaction effects. Time is complicated. Many algorithms become popular because they are easy to use and deploy, in part because they treat each feature independently and don’t perceive an order between input features. But the order of things matters. Knowing when there was an increase or decrease in the white blood cell count, for example, is critical to making accurate predictions.
Steve: Well, the time element is pretty rare, if not unique, and it is essential for accuracy. Then there’s the data. The power and reliability of AI tools depend on having the right data and lots of it. A recent survey commissioned by Wolters Kluwer—Mending Healthcare in America 2020—found that nearly 90 percent of hospital executives feel they need more comprehensive patient data to deliver better care, and attribute incorrect or poor quality data as a cause of increased costs. Clinical data is a strength of Wolters Kluwer, as is our ability to integrate a tool like this into an established clinical surveillance solution like Sentri7—one that has a familiar workflow for users. Sentri7 monitors over 500,000 patients at any given time, as well as processing 4 billion lab orders and 677 million drug orders each year. That amount of information dramatically increases the predictive power of the tool.
John: My team at Wolters Kluwer employs rigorous data science approaches, including cross-validation, in which we hold out some data, so that we can do multiple evaluations of the holdout data versus the AI-trained data. We have also completed extensive analysis to understand how different demographic factors, such as age, contribute to predictions to control for bias. Last of all, and maybe the most important difference, is the way we have always combined data science expertise and clinical expertise at Wolters Kluwer. While building the model, clinicians strongly influenced the feature engineering—that is the way we determined how to transpose raw data from the electronic medical record in a way that gives the AI algorithm traction to learn from the data and statistics. Clinical expertise is critical both at that point and, of course, in the validation process.
Steve: Another difference is that the tool offers evidence-based, customizable recommendations on how to treat these high-risk patients. This customizable risk score threshold enables each facility to decide at which point they want to act. We always give our hospital partners the control to fine-tune the alerts for their specific settings.
Matt: I would reiterate that having this on a trusted clinical surveillance platform already embraced and used by hundreds of hospitals and health systems extends the value of that solution to clinical teams that are hungry for a more progressive and forward-thinking approach to their problems. C. diff is a prime example of that. What I find especially exciting is that we provide transparency to clinicians and hospitals in terms of their patients’ risk and empower them to consider the next steps that are relevant to their circumstances. For example, hospitals can implement a pre-authorization protocol for a certain class of antimicrobials or protocols that avoid inappropriate CDI testing in combination with laxative use, a combination that can artificially inflate CDI rates. Lastly, it’s not a black-box approach, and that’s very, very important for making sure clinicians feel that they can embrace this, interpret it, and use it to its fullest potential within the confines of their facility.
Written for clinicians