A quiet revolution is occurring. The machine learning revolution. It is pervasive and we have all been swept up in it. From the smart speaker that wakes you in the morning, to the route guidance that gets you to work, to the social media feed that knows what you want to read, our lives have – for better or worse – been changed by it.
This revolution is now beginning to transform our healthcare. Machine learning systems can review and report chest X-rays, recognise diabetic retinopathy and anticipate kidney injury. We wondered whether they could help choose your next antibiotic treatment.
The selection of empirical antibiotic treatments – those chosen before there is any clear evidence of the cause of a patient’s infection – remains guided largely by the judgement of the assessing physician, perhaps with reference to a clinical guideline. These judgements must be made days before full microbiological results are available and the prescriber will account for such factors as the presumed source of the infection, the patient’s medical history and previous microbiology results before selecting treatment. They must balance the survival benefit that may result from the prompt initiation of effective antibiotic therapy against the risk of adverse side-effects, complications and increased costs that may follow the use of unnecessarily broad-spectrum agents.
The modern health system has access to enormous amounts of information about each patient including previous admissions, microbiology results, diagnostic coding, electronic prescribing data and basic demographics. Yet this resource is under-used when making empirical treatment decisions.
There has been some success at developing electronic decision support systems for the purposes of guiding individualised prescribing. These may use scoring tools or decision trees based on a few previously identified risk factors. More sophisticated machine learning techniques have been shown to have great potential, accurately predicting antibiotic resistance in those with urinary tract infections or children with bacteraemia. They are yet to enter widespread use perhaps due to their proprietorial nature, acceptability to physicians and patients or uncertainties regarding their performance in different clinical and geographical settings.
Recent years have seen the development of a number of open-source machine learning algorithms. Such tools could be applied to the clinical and microbiological data held by individual healthcare institutions, producing computerised decision support systems tailored to each hospital and capable of responding to changes over time as new data is gathered. In a proof-of-concept study published in the Journal of Antimicrobial Chemotherapy we assessed the accuracy of the open source “XGBoost” machine-learning algorithm. We trained it to predict antibiotic resistance within three selected bacterial species grown from samples obtained from patients within the first 48-hours of admission. We compared its performance to that of medical staff and demonstrated that such systems have the potential to improve antibiotic selection, potentially reducing costs, complications and ultimately, perhaps, improving patient outcomes.
The medical machine learning revolution is just beginning. Soon the same technological principles so effectively deployed in running our smart speakers, choosing our adverts, and filling our junk mail boxes will be improving our healthcare.
Ed Moran, Consultant in Infectious Disease