Machine learning helps predict resistance to common CF antibiotics
Models map electronic health record data against 5 common treatments
Written by |
A study found machine learning models could help predict whether chronic lung infections in people with cystic fibrosis (CF) will be resistant to commonly used antibiotics, helping doctors make more informed treatment choices.
By linking a decade of information on patients’ antibiotic resistance with previous antibiotic use and other clinical information collected in electronic health records (EHRs), researchers developed models that could predict resistance to five commonly used antibiotics with reasonable accuracy. A person’s previous antibiotic use and longer-term resistance history helped inform the predictions.
“These findings highlight the potential value of EHR-derived prediction tools as a proof-of-concept approach to support earlier, more individualized antibiotic decision-making in chronic lung infection,” the researchers wrote. They added that the results require “external validation before clinical implementation.”
The study, “Prediction of Antimicrobial Resistance in People Living With Cystic Fibrosis Using Machine Learning,” was published in MedComm.
Turning to AI for help with challenge
Persistent lung infections remain a major challenge for people with CF, with Pseudomonas aeruginosa (P. aeruginosa) being a particularly common cause of chronic infection in adults. Treating these infections often requires repeated or prolonged courses of antibiotics, which can promote the emergence of bacteria that are resistant to treatment.
Doctors need to know which antibiotics will work against the bacteria in order to prescribe the right one. Antibiotic susceptibility testing (AST) involves testing bacteria grown from a sputum sample against different antibiotics, a process that can take several days, so treatment is frequently started before resistance profiles are available.
Predicting antimicrobial resistance before treatment begins could help doctors choose antibiotics that are more likely to work. Researchers in the U.K. set out to investigate whether routinely collected EHR data could be used to make such predictions before current AST results become available.
To do so, they turned to machine learning, a type of artificial intelligence in which computers use mathematical rules, called algorithms, to identify patterns in large amounts of data that can then be used to make predictions.
The team retrospectively analyzed EHR data from 209 adults with CF, with a median age of 33, treated at a specialist center in the U.K. between 2012 and 2022. About one-third (34.9%) carried the Liverpool epidemic strain, a highly transmissible strain of P. aeruginosa associated with chronic lung infections.
The analysis involved 12,618 sputum cultures, which were linked to 63,823 days of intravenous (into-the-vein) antibiotic use, as well as previous AST results, demographic characteristics, and clinical data.
The researchers then trained and tested five machine-learning models to predict resistance to five commonly tested antibiotics: ciprofloxacin, ceftazidime, meropenem, piperacillin/tazobactam, and tobramycin.
Of the models tested, extreme gradient boosting performed most consistently. Its ability to distinguish resistant from susceptible cultures, measured by the area under the curve, ranged from 0.75 to 0.8 across the five antibiotics, suggesting “reasonable discrimination for predicting resistance to the five commonly tested antibiotics,” the researchers wrote.
Previous resistance history stood out as a key piece of information for the predictions. Longer-term resistance patterns were generally more informative than recent AST results, suggesting that looking further back in a patient’s history may provide useful information about current resistance.
Previous antibiotic use also played a role. In some cases, exposure to one antibiotic helped predict resistance to another. For example, among cultures containing P. aeruginosa, prior piperacillin/tazobactam use was important for predicting tobramycin resistance, while previous tobramycin use helped predict resistance to piperacillin/tazobactam.
The researchers said the findings support “the feasibility of using EHR-derived data to estimate [antimicrobial resistance] before culture results are available,” which could potentially help doctors make more informed initial treatment decisions.
Still, they cautioned that the study was retrospective and used data from a single CF center. The records also captured intravenous antibiotic use, but did not provide a complete picture of patients’ exposure to oral and inhaled antibiotics. Changes in CF care and laboratory practices during the decade covered by the study could also have affected the results.
“External validation, broader antibiotic exposure data, and assessment of temporal dataset shift are needed before clinical use,” the team concluded.
Leave a comment
Fill in the required fields to post. Your email address will not be published.