Prediction of antibiotic resistance at the patient level using deep learning

Rapid and accurate diagnostics of bacterial infections are necessary for efficient treatment of antibiotic-resistant pathogens. Cultivation-based methods, such as antibiotic susceptibility testing (AST), are limited by bacterial growth rates and may not produce results before the treatment starts. This increases patient risks and antibiotic overprescription. Here, we present a deep-learning method that merges patient data with available AST results to predict antibiotic susceptibilities that have not yet been measured. The method is combined with conformal prediction (CP) to enable the estimation of uncertainty of the predictions at the patient level. After training on three million AST results from thirty European countries, the method can predict susceptibility with a major error rate below 2.5% for quinolones, cephalosporins, and aminoglycosides, and below 12% for penicillins. Furthermore, the model predicts resistance of cephalosporins and fluoroquinolones with an average very major rate of 1.5% and 3.2%, respectively, but with higher very major error rates for penicillins, nalidixic acid, and aminoglycosides. We also show that the method reflects empirical error rates, even when limited diagnostic information is available. We conclude that decision support based on deep learning may offer new means to meet the growing burden of antibiotic resistance. IMPORTANCE Improved diagnostics tools are vital for maintaining efficient treatments of antibiotic-resistant bacteria and for reducing the overconsumption of antibiotics. In our research, we introduce a new deep learning-based method capable of predicting untested antibiotic resistance phenotypes. The method utilizes transformers – a powerful technique also used in large language models – which can efficiently take advantage of antibiotic susceptibility tests (AST) and patient data simultaneously. The model produces computational predictions that can be used as time- and cost-efficient alternatives to results from additional cultivation-based diagnostic tests. Significantly, our study highlights the potential of AI technologies for meeting the growing burden of antibiotic-resistant bacterial infections. ### Competing Interest Statement The authors have declared no competing interest.