Liverpool, UK – A new study suggests that machine learning could help predict whether chronic lung infections in people with cystic fibrosis (CF) are resistant to commonly used antibiotics. By identifying resistance patterns before laboratory results are available, such tools could eventually help doctors make more informed and individualized treatment decisions.
The study, “Prediction of Antimicrobial Resistance in People Living With Cystic Fibrosis Using Machine Learning,” was published in MedComm. Researchers in the U.K. used information from electronic health records (EHRs) to determine whether previous antibiotic use, resistance history, and other clinical factors could be used to predict antimicrobial resistance.
A challenge in treating chronic lung infections
Persistent lung infections are a major concern for people with CF. Pseudomonas aeruginosa (P. aeruginosa) is a particularly common cause of chronic infection in adults with the disease.
Managing these infections often involves repeated or prolonged antibiotic treatment. Over time, this exposure can contribute to the development of antibiotic-resistant bacteria, making it increasingly difficult to identify treatments that will be effective.
Doctors typically rely on antibiotic susceptibility testing (AST) to determine which antibiotics are likely to work. AST involves testing bacteria grown from a patient’s sputum sample against different antibiotics, but results can take several days. As a result, treatment often begins before the patient’s current resistance profile is known.
A tool capable of predicting resistance before AST results become available could help doctors select an initial antibiotic that is more likely to be effective.
Researchers use EHR data to build prediction models
The research team retrospectively analyzed EHR data from 209 adults with CF who were treated at a specialist U.K. center between 2012 and 2022. Participants had a median age of 33, and 34.9% carried the Liverpool epidemic strain of P. aeruginosa, a highly transmissible strain associated with chronic lung infections.
The analysis included 12,618 sputum cultures. Researchers linked these samples to 63,823 days of intravenous antibiotic treatment, previous AST results, demographic characteristics, and other clinical information.
They then developed and tested five machine-learning models to predict resistance to five commonly tested antibiotics:
- Ciprofloxacin
- Ceftazidime
- Meropenem
- Piperacillin/tazobactam
- Tobramycin
Of the models evaluated, extreme gradient boosting produced the most consistent results. Its ability to distinguish between antibiotic-resistant and susceptible cultures, measured using the area under the curve (AUC), ranged from 0.75 to 0.80 across the five antibiotics. The researchers described this as “reasonable discrimination” for predicting resistance.
Previous resistance and antibiotic exposure were important
A patient’s history of antibiotic resistance was one of the most informative factors in the models. Interestingly, longer-term resistance patterns were generally more useful than recent AST results, suggesting that a patient’s resistance history over an extended period may provide valuable information about their current infection.
Previous antibiotic exposure also contributed to the predictions. In some cases, treatment with one antibiotic appeared to provide information about resistance to another.
For example, among cultures containing P. aeruginosa, previous use of piperacillin/tazobactam was an important predictor of tobramycin resistance, while previous tobramycin use helped predict resistance to piperacillin/tazobactam.
These findings suggest that combining a patient’s broader treatment history with previous resistance results may improve the ability to anticipate antimicrobial resistance.
Potential to support earlier treatment decisions
The researchers said their findings demonstrate the feasibility of using routinely collected EHR data to estimate antimicrobial resistance before current culture and AST results are available.
Such prediction tools could potentially give doctors additional information when choosing an initial antibiotic, particularly when waiting several days for laboratory results is not practical.
However, the researchers emphasized that the findings are preliminary. The study was retrospective and relied on data from a single CF center, so the models may not perform in the same way in other patient populations or healthcare settings.
The study also recorded intravenous antibiotic exposure but did not provide a complete picture of patients’ use of oral or inhaled antibiotics. In addition, changes in CF treatment practices and laboratory procedures over the 10-year study period could have influenced the results.
The researchers concluded that external validation, more comprehensive antibiotic exposure data, and evaluation of changes in the underlying data over time are needed before these models can be used in clinical practice.
Contact
Freddy Frost
University of Liverpool