Machine learning tool outperforms existing methods in detecting drug-resistant bacterial infections

Carbapenem-resistant Klebsiella pneumoniae has become a fixture in hospital infections worldwide, and it kills with a particular ruthlessness: the bacteria produce enzymes that break down the last-resort antibiotics doctors reach for when everything else has failed. A team of researchers has now built a machine learning system that can identify these resistant strains faster and more reliably than the diagnostic tools currently in use, potentially changing how clinicians choose treatment for patients facing these infections.

The bacterium develops resistance through several mechanisms. It manufactures carbapenemases—enzymes with names like KPC and NDM that dismantle the antibiotics meant to kill it. It also thickens its outer membrane by mutating the proteins that normally allow drugs to pass through, a change in structures called OmpK35 and OmpK36. And it shuffles its genetic material, copying genes or trading plasmids with other bacteria to accumulate new defenses. This combination of strategies has made K. pneumoniae a moving target for antibiotic therapy.

The researchers trained their model, called PredKlebAMR, on nearly 6,000 bacterial genomes pulled from public databases. They used a technique called Random Forest, which builds many decision trees and combines their predictions, and they fed it 38 genetic features extracted using a tool called Kleborate. The system learned to recognize patterns in the genome that signal resistance—fluoroquinolone mutations, high-risk bacterial lineages, and the specific porin mutations that wall off the cell. When tested on a held-out set of genomes the model had never seen, it achieved 91.92 percent accuracy, with sensitivity of 94.67 percent and specificity of 89.47 percent.

But the real test came when researchers compared PredKlebAMR to two existing diagnostic tools: ResFinder 4.1 and AMRFinderPlus 4.2.7. They used three separate cohorts—a global dataset, samples from Africa, and samples from Thailand—to see how each tool performed. The stakes here matter: a "Very Major Error" (VME) occurs when a diagnostic tool misses resistance that is actually present, leading a clinician to prescribe an antibiotic the bacteria can defeat. In the global cohort, PredKlebAMR caught resistance in 95.45 percent of cases and made only one VME, compared to two VMEs for AMRFinderPlus and three for ResFinder. In the African cohort, the advantage was starker: PredKlebAMR achieved 95.56 percent sensitivity with two VMEs, while ResFinder made eight errors. In Thailand, PredKlebAMR identified resistance in 95.62 percent of samples and made seven VMEs; ResFinder made 59. The improvement in Africa reached statistical significance.

What makes this tool different is not just its accuracy but how it communicates that accuracy to the people who need it. The researchers built an interactive dashboard that shows clinicians not just a prediction but the biological reasoning behind it—which genes are present, which mutations matter, what the resistance profile actually is. This transparency matters when a doctor is deciding whether to use a particular antibiotic or switch to something else, knowing that the choice could mean the difference between a patient recovering and a patient dying from an infection that should have been treatable.

The tool is now available online, accessible to any laboratory or hospital system that wants to use it. The real work ahead is adoption—whether clinicians will integrate it into their workflows, whether laboratories will run genomes through it before reporting results, whether the speed and accuracy it offers will actually change how patients are treated. For now, the system exists as a proof that machine learning can outperform human-designed diagnostic algorithms at a task where the cost of failure is measured in lives.