What if doctors can swiftly and accurately prescribe a personalised antibiotics regimen to patients, thereby reducing resistance to the drug?
In Singapore, this could soon become a reality, with the help of artificial intelligence. Singapore General Hospital (SGH), in partnership with DXC Technology and national health tech agency Synapxe, developed a solution called Augmented Intelligence in Infectious Diseases (AI2D), meant to target two common infections treated in hospitals — pneumonia and urinary tract infection.
Andrea Kwa, Associate Professor, Centre for Clinician-Scientist Plus Development, SGH, and Valerie Nathan, Director, Consulting & Analytics, ASEAN, DXC Technology, spoke to Frontier Enterprise to elaborate on details of the innovation.
Cutting-edge
According to Kwa, AI has a lot of potential to aid doctors in making the correct diagnosis, and in the context of antibiotics, ensure that the right medicine is prescribed.
“We have envisioned that if we can augment the doctors’ clinical judgement by helping to rule out bacterial pneumonia with a level of certainty by using patients’ clinical features, symptoms, biochemical results, and clinical history, chest x-ray reports, then antibiotics prescribing can be appropriate right from the start. If the patient has very low risk for bacterial infection, then the doctor may not need to prescribe antibiotics,” she said.
In the U.S., the Centers for Disease Control and Prevention (CDC) reported that up to 50% of all antibiotics prescribed in the country are unnecessary or inappropriate, the majority of which were prescribed in in-patient settings.
During the pilot validation study, SGH, led by its Division of Pharmacy, showed that AI2D has a 90% accuracy in determining whether antibiotics are necessary in the first place.
“AI2D is more specific than doctors in predicting bacterial pneumonia, while doctors are more sensitive in predicting the disease. If the doctors use AI2D in their clinical judgment, their high sensitivity will be augmented by high specificity contributed by AI2D in bacterial pneumonia diagnosis and prediction. When taken together, the final clinical judgement will be very much sharpened with high accuracy,” Kwa explained.
According to SGH, the pneumonia model leveraged retrospective deidentified clinical data like x-rays, clinical symptoms, periodic vital signs, and trends of common body responses to infection, of around 8,000 SGH patients between 2019 and 2020.
The model was then validated against another 2,000 cases in 2023 using a design that would simulate real-life usage when deployed.
“Besides showing a high level of accuracy, the pilot validation study also revealed that almost 40% of antibiotics prescribed in those cases to treat pneumonia at the onset may not have been necessary. This situation is not unique to Singapore,” SGH said in a statement.
Integration challenges
In order to maintain accuracy and reliability during clinical data integration, several technical issues had to be overcome.
According to Nathan, these included data quality and conformity due to diverse data sources and silos, which required data sets to be processed and accurately categorised.
“Fortunately, these challenges could be addressed through data engineering and data cleansing strategies. Data engineering efforts helped transform medical information collected at various points of the patient’s journey into useful components for analysis for the specific illness, while data cleansing strategies addressed missing information as well as data standardisation, which was needed by the AI model,” she said.
Nathan also cautioned against overreliance on AI models, which may result in unconscious biases in decision-making or unintentional machine override in human judgements. “There were cases which required manual reviews, and in these cases, the operational AI model identified treatment profiles that required additional reviews to narrow down the number of cases requiring immediate attention by the Antimicrobial Stewardship Programs (ASP) teams,” she recalled.
“The model reduced the number of cases for reviews by a third (from 2,012 to 624) and increased the possibility of identifying priority cases requiring intervention by three times,” Nathan added.
Finally, as regulations and compliance frameworks struggle to keep up with the speed and rapid advancement of AI technology, patients and healthcare providers need to be fully informed about how data will be used within AI systems to ensure ethical compliance. Likewise, participating organisations need to adapt to changing regulations around the topic of AI,” Nathan noted.
“Data governance that includes data integrity and security will be needed, and Singapore has put this very much in place,” Kwa stressed.
Ways forward
Globally, antibiotic pipelines are dwindling, and very soon, there will be no pharmaceutical company manufacturing antibiotics. This is why SGH is focused on fine-tuning AI2D, Kwa said.
“More than 50% of companies selling antibiotics have already fallen into bankruptcy. This also means that antibiotics which are newly launched will be rendered ineffective by antimicrobial resistance within a few years. We foresee AI2D will reduce unnecessary antibiotics prescription, thus retarding antibiotic resistance, thereby providing sustainability until we have better options,” she said.
Meanwhile, DXC is now working with healthcare providers on how to further integrate AI into other parts of the hospital workflow via digital twins.
“Valuable lessons have been learnt through this project, namely that AI has the power to enhance medical procedures through data collection and analysis. The pilot study not only demonstrated a 90% accuracy for the prescription of antibiotics for pneumonia, but also showed that nearly 40% of antibiotics prescribed for it might not have been necessary,” Nathan concluded.
