AI spots colistin-resistant bacteria in one hour

A new machine learning workflow cuts the detection time for superbugs from days to sixty minutes, shifting the battle against drug-resistant hospital infections.

When a patient contracts a pan-drug-resistant infection, doctors reach for colistin as their absolute last resort. But waiting days for standard lab cultures to confirm if the bacteria resist this final drug is a luxury modern medicine cannot afford.

This delay forces clinicians to prescribe blind. A new study challenges this slow diagnostic paradigm by using machine learning to spot colistin-resistant *Enterobacter cloacae* complex (ECC) in just **one hour**.

The shift from biological culturing to algorithmic pattern matching changes how we manage hospital outbreaks. Instead of waiting for bacteria to grow in the presence of drugs, we can now read their chemical fingerprints instantly.

How the math works

The researchers bypassed traditional slow testing by pairing negative-ion mode MALDI-TOF mass spectrometry with a deep learning model. They focused specifically on lipid A-enriched signals in the **m/z 1500–3000** range, where resistance mechanisms leave distinct chemical marks.

The team trained a one-dimensional convolutional neural network (1D-CNN) equipped with a squeeze-and-excitation module to analyze these complex spectra. They processed the raw data by binning it at **0.1 Da** and smoothing it with a Savitzky–Golay filter to eliminate background noise.

This approach targets the physical structure of the bacterial cell wall rather than waiting for the organism to multiply. This chemical-first strategy is crucial as global reports highlight the rise of highly resistant strains, such as those documented in Senegal and emerging multi-drug resistant strains in Italy.

The performance metrics

The study analyzed **267** clinical ECC isolates, verifying species identification through whole-genome sequencing. The neural network was trained on **217** isolates and then tested against an independent external cohort of **50** isolates.

The results show that the algorithm did not just match human expertise; it outperformed standard machine-learning baselines.

• The model achieved **95.5%** accuracy during internal testing, with an AUROC of **0.986** and an F1-score of **0.943**.
• External validation on the independent cohort yielded **96.0%** accuracy and an F1-score of **0.960**.
• SHAP analysis identified **30** key lipid-associated features that drove the model’s predictions, offering a transparent look at the underlying biology.

The reality check

This diagnostic speed matters because it stops the unnecessary use of toxic last-line antibiotics. However, the workflow still requires specialized mass spectrometry equipment, which is not yet universal in low-resource clinics.

Additionally, while the **30** SHAP features offer clues to resistance mechanisms, translating these computational patterns into actionable biological targets remains a hurdle.

Even with these limits, the trial proves that physical chemistry and neural networks can bypass the slow biological clock of bacterial growth.

Read the full study in ACS Omega.