Multidrug-resistant Escherichia coli has become a serious obstacle in treating urinary tract infections (UTIs). These infections are now harder to cure, more likely to come back, and cost more to manage. Traditional treatments often fall back on broad-spectrum antibiotics or guesswork, approaches that not only risk treatment failure but also contribute to rising antimicrobial resistance.
At CMKL University, student Chanbormei Suy and her advisor, Dr.Prompong Sugunnasil, are turning to machine learning for a better answer. Their work uses real-world hospital data to
develop a smarter, more personalized approach to choosing antibiotics, one that goes beyond resistance prediction and aims to optimize entire treatment regimens.
“We’re trying to optimize the entire treatment plan,” Chanbormei explained. “It’s not just about saying this drug works or doesn’t. It’s about what works best, in combination, for a particular patient with a particular history.”
The project pulls from the MIMIC-III critical care database, which contains anonymized health records from over 40,000 ICU patients. From this data, the team filtered cases involving E. coli, then built a dataset combining microbiology test results, treatment histories, and basic demographics like age and gender. This allowed them to see which antibiotics were effective (or not) for each individual patient.
They then trained multiple machine learning models, including Random Forest, LightGBM, a Stacked Ensemble, and a custom neural network using PyTorch. Each model was designed for multi-label classification, meaning it could recommend several potentially effective antibiotics at once, not just a single drug.
Because the task is complex, the team used a range of evaluation methods to measure model performance. They looked at how often the models correctly predicted effective antibiotics, how well the predicted sets matched the actual treatments, and how many incorrect suggestions were made. These metrics included Hamming Loss, Jaccard Similarity, and the Macro F1 score, which balances precision and recall across all possible antibiotic labels.
They also applied SHAP (Shapley Additive Explanations) to help explain why a model recommended certain antibiotics, which is an essential step for building trust with clinicians.
“Interpretability is key,” said Chanbormei. “Doctors need to understand not just what’s being recommended, but why.”
Looking forward, the team plans to expand the models to consider treatment duration, recurrence risk, and richer clinical data such as comorbidities, lab values, and prior antibiotic use. The ultimate goal is a clinical decision support tool that helps doctors pick the most effective and resistance-aware treatment plan for each patient.
“This is just the beginning,” said Dr.Prompong. “With more data and better models, we can help doctors make faster, safer, and more personalized choices, hopefully before resistance becomes an even bigger crisis.”
In a world where antibiotic resistance is growing rapidly, this research offers a hopeful direction. By combining patient data, microbiology, and machine learning, the team is building a smarter, more personalized approach to treating UTIs, one that could help patients recover faster and limit the spread of resistance.
