ID is having a ‘Wild West moment’ with AI
Click Here to Manage Email Alerts
Artificial intelligence has emerged as a potentially important tool for patient care, including in infectious diseases.
Although AI is not currently being used widely in ID clinical care, experts anticipate that it will eventually be able to support — but not replace — the work of ID clinicians, including in the areas of antimicrobial stewardship, infection prevention and identifying patients at increased risk for infection.
It could also be used to facilitate and speed up the development of new drugs, according to Healio | Infectious Disease News Editorial Board Member Gitanjali Pai, MD, AAHIVS, FIDSA, chief medical officer for the state of Oklahoma and an infectious disease physician at Memorial Hospital and Physicians Clinic in Stilwell, Oklahoma.
“In our ongoing fight against emerging infectious diseases and in combating antimicrobial resistance, AI can be a crucial tool in helping to examine and analyze massive datasets of molecular structures to identify potential drugs and expedite the pipeline of new antibiotics and vaccines,” Pai told Healio | Infectious Disease News.
Pai called AI “a powerful ally that can significantly improve our ability to combat infectious diseases” but also noted that it is complementary to clinical expertise and dependent on the quality of its data.
Research has also shown that, in its current iterations, AI could be harmful to patient care if used alone.
“Anything we use AI for now needs to be thoroughly fact checked. In the future, I imagine it will be a much more evidence-based, standard part of care, but right now, there are dangers to it,” said Healio | Infectious Disease News Editorial Board Member Keith S. Kaye, MD, MPH, division chief of allergy, immunology and infectious diseases and professor of infectious diseases and medicine at Rutgers Robert Wood Johnson Medical School.
Given its ascendence, we asked clinicians about their expectations and concerns regarding AI and its conceivable impact in the field of ID, including what it could mean for the future of patient care.
“Right now,” Kaye said, “we’re at a Wild West moment.”
‘The good, the bad’
In a recent article published in Open Forum Infectious Diseases, Gonzalo M. Bearman, MD, MPH, and colleagues addressed the potential for AI to both innovate and disrupt ID care.
“AI models may help ensure earlier detection of disease, more personalized empiric treatment recommendations, and allocation of human resources to support higher yield antimicrobial stewardship and infection prevention strategies,” they wrote, but they also acknowledged that AI has limitations, including a lack of contextual awareness and transparency and a potential to “perpetuate existing biases.”
“AI is unlikely to replace the role of ID experts but could instead augment it,” they wrote.
Bearman, a professor of internal medicine and chief of infectious diseases at Virginia Commonwealth University, told Healio | Infectious Disease News that he and his colleagues wanted to respond to the “hype and excitement” of AI, but they also wanted to discuss some of their concerns about the technology.
“I wanted that response to be comprehensive in terms of the good, the bad, what we can do, and the role that we can play,” he said.
In the paper, the authors reviewed different types of AI, including large language models like ChatGPT, which they noted could potentially be used to distill electronic medical charts, help with billing and assist clinicians in making decisions about antimicrobials.
According to the review, other potential benefits of incorporating AI in ID include using machine learning to combine genomic data and electronic health records to identify unseen outbreak clusters, and using robots to disinfect surfaces or monitor the donning and doffing of personal protective equipment.
“While AI is likely to disrupt infectious diseases and health care in general, clinicians who cautiously embrace these emerging technologies can help improve patient care and enrich their career,” another one of the authors, Bradley Langford, PharmD, BCIDP, an antimicrobial stewardship specialist and assistant professor at the University of Toronto, told Healio | Infectious Disease News.
Langford said ID specialists and antimicrobial stewards should participate in AI training and educational opportunities and advocate for ID's involvement in “identifying, deploying and evaluating AI solutions to help patients with infectious diseases.”
AI applications are ‘not sci-fi at this point’
Research has demonstrated the potential strengths of AI for ID.
One study by Timothy Wiemken, PhD, MPH, associate professor of ID, allergy and immunology at the Saint Louis University School of Medicine, and Ruth M. Carrico, PhD, DNP, an associate professor of medicine at the University of Louisville, assessed the ability of two AI platforms, including ChatGPT, to identify two health care-associated infections: central line-associated bloodstream infections (CLABSIs) and catheter-associated UTIs (CAUTIs) from six training scenarios.
Overall, both platforms accurately identified CLABSIs and CAUTIs 100% of the time when given clear prompts — questions or tasks assigned to the platform — but struggled when given ambiguous prompts, such as when the researchers used numerals instead of identifying months by their names. The overuse of abbreviations and special characters also caused occasional inaccuracies, the researchers reported.
Wiemken and Carrico wrote that the study “provided the first indications of the benefits that may be gained using AI assistance” for the surveillance of hospital-associated infections.
“AI is here to stay,” Wiemken told Healio | Infectious Disease News. “What if you had another person or another 100 people — what would be the thing that you would have them do? Take that and then work with somebody try to build systems that can facilitate this for you. Don't be scared that these systems are going to take your job — they’re not going to take your job — but they can help you do your job and focus more on patients.”
AI has also been shown to accurately analyze medical images or scans, offering earlier and potentially more precise diagnoses, and is a promising new tool for personalized medicine, with the potential to analyze a patient's health information, medical history and genetic data to identify patterns and predict potential health risks, Pai explained.
In one example of this, a study showed that by using an AI-based risk assessment tool, researchers were able to create infection risk scores to identify patients at higher risk for HIV, chlamydia, gonorrhea and syphilis.
Phyu M. Latt,MBBS, MPH, a PhD candidate and research assistant at the Melbourne Sexual Health Center, explained why the results are so important.
“In practice, knowing one’s risk of HIV and STIs enables targeted and better HIV testing and prevention,” Latt told Healio | Infectious Disease News after the study was published.
A 2018 study demonstrated that AI could be used to speed up the identification of patients at higher risk for Clostridium difficile by generating daily risk scores for each patient. This could enhance prevention strategies, enabling cost-effective interventions that could reduce exposures to high-risk antibiotics, the researchers said.
“A lot of recent AI research in ID focuses on prediction,” Langford said. “But prediction is not the same as human decision-making. We need a human in the loop to ensure such models are properly implemented, user friendly, validated, and an expert still makes the final decision, taking into account all the nuances of each individual patient.”
AI has the potential to offer more than just predictions, Wiemken said.
“When we talk about AI and where we're at, we’re at the precipice of moving to multimodal models and not just language models,” he said. “You can imagine a scenario where you hold your phone up to someone's heart and it’s able to tell you that they don’t have endocarditis. These are things that are not sci-fi at this point. It's coming very soon.”
Studies have already shown that smartphone app-based programs may have a place in clinical care.
One example is a smartphone app that records coughs and can diagnose respiratory conditions in children, such as asthma, pneumonia and bronchiolitis, with accuracy that is comparable to a clinician, according to study results published in 2019. The researchers recorded cough sounds from more than 500 children using a smartphone’s microphone and developed a diagnostic algorithm that incorporated the cough sounds with five patient- or parent-reported symptoms.
More recently, a team led by researchers and clinicians at the University of Pittsburgh developed an app that uses an AI algorithm to diagnose ear infections in children. According to study results published in JAMA Pediatrics, the app, which assesses videos of the tympanic membrane, was more accurate than pediatricians, primary care physicians and advanced practice clinicians in correctly identifying acute otitis media in young children, which is frequently diagnosed, although often inaccurately, the researchers said.
‘Ask a specialist’
Experts also have concerns about the widespread use of AI in ID.
One concern is AI “hallucination” — when large language models produce “fluent but factually incorrect or off-topic output,” researchers explained last year in an article in Clinical Infectious Diseases.
“Large language models present a risk of confabulation or hallucinations where the output seems confident and plausible but is incorrect,” Langford said.
Another study published last year showed that relying on AI by itself without input from an ID clinician could be hazardous for patients.
Alexis Maillard, MSc, a member of the infectious diseases stewardship team at the Paris Centre University Hospital, and colleagues assessed the ability of ChatGPT to manage patients with positive blood cultures in a real-life setting by prospectively providing data from ID consultations to a chatbot over a 4-week period. ChatGPT used the data to generate management plans, which the researchers compared with plans suggested by ID consultants that were based on literature and guidelines.
In total, 44 cases of a first positive blood culture were included in the study. Overall, the researchers found that ChatGPT provided “detailed and well-written” reports in all cases and generated identical diagnoses as the ID consult 59% of the time. However, the study also showed that ChatGPT’s definitive antibiotic therapies were “appropriate and optimal” for only 36% of patients — and harmful in two patients.
“So far, do not use ChatGPT for infectious disease advice — ask a specialist,” Maillard told Healio | Infectious Disease News after the study was published.
“However,” he said, “given its ability to write clear and informative medical reports, why not use generative AI as an assistive tool to improve the formal quality of our reports in the future?”
Questions of equity and privacy
There is also the potential for large language models to propagate pre-existing biases against race and gender that are present in the data used to train them, Langford said.
These biases “can skew AI outputs, leading to questionable recommendations and processes,” according to Pai.
“By no means is AI perfect, and it is not devoid of limitations,” she said. “AI is going to be as good as the data that it is based off of, so any data incompleteness, accuracy issues and biases will be carried over into the final recommendations — unless these are identified and addressed early on.”
According to Pai, although it may be difficult to eliminate biases in AI, “taking concrete steps beforehand — such as the use of diverse data sets to train AI models, strict quality control and regulatory oversight — is paramount to minimizing the potential for bias.”
In their article, Bearman and colleagues said ID experts should be involved in validating machine learning algorithms to help minimize bias “and advocate for equitable access to AI models to support underserved and/or marginalized communities.”
“For example, data from insurance billing claims illustrate that the use of medical AI is skewed toward higher income, urban regions, and academic medical centers,” they wrote.
Pai noted that AI could be an expensive proposition for rural practices that might not be able afford the cost of developing and maintaining it. Even maintaining a basic EHR system can be financially draining for some practices, Pai said.
“This may play a crucial role in equity — facilities that may not be able to afford the cost of AI may be left behind, magnifying health care inequities and disparities further,” she said.
Resistance to technology in general could also make it daunting to roll out AI in rural areas, Pai said — all of which raises concerns that the areas that might benefit the most from AI may not be able to use it.
“Rural facilities can also tend to be older, and less acquainted or interested in technology,” Pai said.
Careful consideration and care also must be taken to protect patient data and avoid a collision between AI data sharing and HIPAA law, according to experts.
“While the convenience of AI in health care may seem very appealing, and rightfully so, the imminent impending question of HIPAA violation remains,” Pai said.
HIPAA compliance, she reiterated, is non-negotiable, and ensuring that there are multiple quality checkpoints embedded within AI systems is critical.
“These include explicit patient consent regarding the scope and use of AI in managing patient data, measures to prevent data breaches, and dynamic regulatory oversight,” she said.
Much like any other technology, AI use in health care can be a double-edged sword, according to Pai.
“Responsible and ethical development and implementation of AI in infectious disease management is our collective responsibility. By acknowledging its limitations and actively working toward solutions, AI can evolve into a powerful and trustworthy ally in our fight against infectious diseases,” she said. “Collectively as a field, we will have to work to strike this balance in order for AI to be applied across the board for the greater good.”
‘The ultimate tool in our arsenal’
Experts are clear that AI should not be seen as a replacement for clinician consultation but as an aid. Right now, it is best thought of as a helper, something that can reduce a clinician’s workload, Wiemken said.
“No matter what field you’re in, [AI is going to] allow us to focus more on why everyone is in health care to begin with — to improve patient safety and to improve patient outcomes,” he said.
Pai said AI has the potential to be “the ultimate tool in our arsenal” — but not a substitute for the human touch.
“Humans are social animals,” she said. “All said and done, patients would probably still prefer to speak directly to a human health care provider rather than getting their diagnosis or treatment recommendations completely from a machine or robot or an application like Siri or Alexa — pleasant as they might sound.”
References:
- Langford BJ, et al. Open Forum Infect Dis. 2024;doi: 10.1093/ofid/ofae053.
- Latt PM, et al. Open Forum Infect Dis. 2024;doi: 10.1093/ofid/ofae011.
- Maillard A, et al. Clin Infect Dis. 2023;doi:10.1093/cid/ciad632.
- Oh J, et al. Infect Control Hosp Epidemiol. 2018;doi:10.1017/ice.2018.16.
- Porter P, et al. Respir Res. 2019;doi:10.1186/s12931-019-1046-6.
- Schwartz IS, et al. Clin Infect Dis. 2023;doi:10.1093/cid/ciad633.
- Shaikh N, et al. JAMA Pediatr. 2024;doi:10.1001/jamapediatrics.2024.0011.
- Wiemken TL, Carrico RM. Am J Infect Control. 2024;doi:10.1016/j.ajic.2024.02.007.
For more information:
Gonzalo M. Bearman, MD, MPH, can be reached at gonzalo.bearman@vcuhealth.org.
Keith S. Kaye, MD, MPH, can be reached at kk1116@rwjms.rutgers.edu.
Bradley Langford, PharmD, can be reached at brad.langford@utoronto.ca.
Gitanjali Pai, MD, AVVHIVS, FIDSA, can be reached at dr.paigg@gmail.com.
Timothy Wiemken, PhD, MPH, can be reached at Tim.wiemken@gmail.com.
Collapse
Healio Interviews
Read more about
Click Here to Manage Email Alerts