Designing Tools for Real World Impact: Using Machine Learning to Personalize Antibiotic Treatment
The rise of antibiotic resistance is a major threat to the practice of medicine and is driven in large part by overuse of antibiotics. In his latest study, Institute Lecturer Sanjat Kanjilal, MD, MPH, and team take on this issue by showing how large-scale machine learning models can be applied to observational electronic health record data to predict antibiotic resistance, make treatment recommendations, and build patient-level and public health models to aid in decision support. We spoke with Dr. Kanjilal to learn more about the study, “A decision algorithm to promote outpatient antimicrobial stewardship for uncomplicated urinary tract infection,” which appears in Science Translational Medicine.
Q: Tell us about antibiotic resistance – why is it a problem, in both the short- and long-term? Has your experience as a clinician dealing with antibiotic resistance influenced your research on the matter?
A: Modern medicine rests on the assumption that we will be able to treat bacterial, fungal, and viral infections when they occur. The rise of antimicrobial resistance threatens that assumption and impacts the care we provide to patients across all healthcare settings. We already know that infections from drug-resistant pathogens are associated with increased rates of complications, economic costs, and mortality. Less appreciated but equally concerning is the prospect that continued antibiotic use will lead to the emergence of new strains that combine the deadly triad of high virulence, high transmissibility, and high-level resistance.
As an infectious diseases doctor, I have cared for multiple patients with untreatable infections due to drug-resistance as well as those who have suffered from complications and drug side-effects associated with increasingly powerful antibiotic therapy, necessitated by the presence of pre-existing resistance. In my mind, this is an existential threat to the very foundations of medical practice, in resource-poor and resource-rich settings alike.
Q: How has your past research influenced this current work?
A: My prior research focused on understanding the rise and spread of methicillin-resistance Staphylococcus aureus (MRSA) in the Boston area over a 17 year period. We found through a combination of whole genome sequencing and analysis of electronic health record data that resistance is a highly dynamic and complex phenomenon. My current work builds upon my experience using passively collected data in the electronic health record but uses far more sophisticated models to pull in more data as well as analyze it in a more rigorous manner. Additionally we move beyond simply describing the epidemiology of antibiotic resistance to building a tool with the potential for real-world impact.
We move beyond simply describing the epidemiology of antibiotic resistance to building a tool with the potential for real-world impact.
Q: Algorithms have been used for clinical decision support for infectious diseases since the 1970s but have not yet been widely adopted. You point out that outpatient settings are particularly neglected when it comes to effective tools to optimize treatment. Why is that, and how does your work with machine learning potentially change that?
A: Up to 70% of all antibiotics for humans are prescribed in outpatient clinics. In contrast, efforts to curb antibiotic use are largely focused on inpatient settings, reflecting a bias to work with ‘data-rich’ environments and for areas that are easier to access physically. Up until recently, relatively little attention has been paid to providing similar tools in a decentralized and scalable fashion for primary care and other clinics scattered throughout the community. Unfortunately, such clinics may be precisely where such efforts are needed the most.
Our goal was to fill this unmet need by developing a preliminary version of a tool that will eventually embed into clinical workflows at any workstation and provide personalized antibiotic treatment recommendations that are personalized to the patient and available in real time.
Q: How did your team conduct this particular study, and what did the findings show? Why did you choose urinary tract infections (UTI) as the infection to focus on?
A: We developed machine learning models trained on data present in the electronic health record to predict the probability of resistance to first and second line antibiotics in women who visit their healthcare provider for uncomplicated UTI. We then developed an algorithm that uses those probabilities to make treatment recommendations that guide providers towards the antibiotic that is of the narrowest-possible spectrum that is still predicted to be effective. We chose UTI because it is a very common reason for prescribing antibiotics in outpatient settings and is a syndrome where providers often use second line agents despite the presence of national practice guidelines that discourage their use.
The main takeaway from our study is that personalized antibiotic treatment using machine learning models is feasible.
Q: What are the potential implications of this study? What future studies are you and your team considering to expand on this work?
A: The main takeaway from our study is that personalized antibiotic treatment using machine learning models is feasible. This opens up an entirely new area of work where the goal is to adapt our current models for deployment into clinical workflows. Such tools would then have to be evaluated using randomized controlled trials to establish whether they can truly impact antibiotic prescription and ultimately reduce antibiotic resistance without harming patients. We are also expanding our prediction models to examine patients with bloodstream infection and pneumonia, as well as to predict the future risk of developing resistance occurring in response to antibiotic treatment.