Developing antimicrobial stewardship programs using AI models to ensure optimal selection of antibiotics
Overuse or misuse of antibiotics in hospital settings is a global problem, fueling one of the most serious public health threats: antimicrobial resistance. An estimated 20 per cent to 50 per cent of acute care hospitals worldwide are not prescribing antibiotics appropriately. As a result, micro-organisms become resistant to antimicrobials, including antibiotics, and can become superbugs that don’t respond to any drugs.
In Singapore, up to 30 percent of infections in hospitals are resistant to third-generation cephalosporins (a large group of antibiotics derived from the mold acremonium), which are widely used broad-spectrum antibiotics. In 2018 alone, approximately 12,000 audits performed for broad-spectrum antibiotic use at Singapore General Hospital (SGH) showed that 20 per cent to 30 per cent of these prescriptions were inappropriate.
One way to tackle the problem of antimicrobial resistance (AMR) is the practice of antimicrobial stewardship programs (ASP).
These initiatives seek to prevent the overuse of antimicrobial agents and ensure optimal selection of antibiotics. A key priority at acute care hospitals in Singapore, led by ASP teams and supported by precise, comprehensive data, is to establish a clear prescribing methodology. However, the development of an effective ASP comes with several challenges and considerations, as Dr Vinod Seetharaman, Chief Medical Officer, Asia at DXC Technology shared from their recent collaboration with SGH.
“Disease trajectories can widely vary owing to a variety of factors such as, patient genotype, phenotype, pathogen, co-morbidities, prescribed medication and other social factors. Therefore, it is important to realize that all of these factors can be variably relevant for different disease groups and solving for a disease group at a time would be a more prudent strategy that allows for progressive refinement of technique with in a well-defined scope,” said Dr Seetharaman.
One of the main challenges was to have a line of sight on all influencing factors. Additionally, the chosen clinical condition would also need to be one where the lifecycle of the patient journey from incidence to resolution would have a data trail, for example, inpatient conditions that typically occur in the hospital setting and also most likely to get resolved in a hospital setting.
“This is important so that we ensure all influencing factors of a patient’s disease trajectory are appropriately factored in an AI model.”
According to Dr Seetharaman, the data elements used in ASPs include:
a) Antibiotic treatment guidelines around recommended dose, duration, route of administration
b) Patient encounter data that included length of stay and outcome information
c) Clinical data around observations, antibiograms, biomarkers, vitals, allergy, recent medications, and compounding conditions such as systemic infections or immunocompromised states, current treatment (dialysis) and
d) Antimicrobial Stewardship Unit (ASU) case resolution states (approved / rejected), ASU new antibiotic recommendations
These data elements are typically aggregated from systems of record such as electronic medical records (EMR), registries and other repositories of clinical and other transactional data.
Good quality data is an essential ingredient in the creation of a robust algorithm for the ASP. In order to achieve good data quality, close collaboration with clinical teams at SGH led to a framework which contained data elements and their interrelationships specific to hospital acquired pneumonia. According to SGH’s own studies, approximately 20 per cent of all infections treated in the hospital setting are due to pneumonia, and more antibiotic prescriptions are written for pneumonia than any other condition.
Data cleansing was also part of the process of achieving good quality data and records that were incomplete with respect to data elements in points c) and d) mentioned above were excluded.
After the data was cleansed, it was segregated into training and test data. The ASP algorithms were built using a variety of machine learning algorithms such as decision trees, XGBoost, SVM, SVC, logistical regression and SMOTE.
As the ASP development progresses, a phased approach is taken with a gradual expansion to include more antibiotics and more disease conditions. As that happens, the AI insights are offered at the point of care in an offline controlled environment for key stakeholder group which can greatly enhance workflow efficiencies.
In summary, AI based decision support for Antimicrobial prescriptions relies on a few fundamental building blocks:
- Sourcing of data from various hospital systems for record – i.e. Electronic medical Record (EMR), Lab Information systems (LIS), Radiology Information Systems (RIS) and patient administration systems (PAS); as well as ensuring good data quality around the input factors – patient demographics, lab/rad tests, observations, episode details, clinical diagnosis and treatment plan and clinical outcomes. DXC Enterprise Management is an EMR that helps maintain high quality, standardized clinical data at an organizational/health system level. to enable the standardization of the data.
- The ability to integrate and standardize disparate sources of data from within and across organizational boundaries, additionally is a capability DXC Open Health Connect provides also provides a robust platform for investigative analytics, which helps quantify quality of care and outcomes at an individual and population level. Further, the platform can also employ actionable intelligence to drive intelligent workflow that result in timely management and interventions.
- Finally, to close the loop, it is key to be able to provide actionable insights in context, at the point of care, such that appropriate and timely interventions may be triggered via intelligent workflows; a core value proposition of DXC Care Personas.
For more information on the joint ASP development between DXC and SGH, click here to download the White Paper.