A deep learning approach for predicting antimicrobial resistance in urinary tract infections.

backgroundDrug resistance is a crucial issue in patients with urinary tract infection (UTI). This study aimed to detect drug-resistant (DR) UTI cases using suitable machine learning classifiers for early prediction of DR UTI.

methodsThree databases were used. The first database was from UTI patients at Alexandria University Hospital in Egypt. The second database was from a general clinic in Northern Mindanao, Philippines. The third was from Al-Rafa Laboratory, IBB University, IBB City, Yemen. Novel mathematical methods were compared for the detection of DR UTI patients in database 1, positive UTI patients or not in database 2, and multi-drug resistance (MDR) or not in database 3. These methods were compared to extract features from the used databases based on novel engineering calculation methods. After that, several classifiers were compared using these features to determine whether the UTI patients were DR to the antibiotics or not. The classifier used was based on a deep learning method, which was known as bidirectional long short-term memory (BILSTM). The logistic regression and gradient boosting classifiers were also compared. Each database was classified based on holdout or K-fold cross-validation methods.

resultsThe selected approach for extracting values of the UTI patients was based on the fractional discrete cosine decomposition method (FrdcDM). The obtained accuracy rate and the area under the curve (AUC) of this method were the highest when it was applied to the three databases, based on hold-out and K-fold cross-validation methods, using deep learning. For example, the accuracy rate and AUC after applying this method on database 1, using deep learning as a classification method based on the holdout validation method, were 97.4% and 95%, respectively. If the same model was applied to database 2, the accuracy and the AUC rates were 99.1% and 99.5%, respectively. The accuracy and AUC rates after applying the method for database 3 were 98.2%, 98.6%, respectively, based on the deep learning of the holdout cross-validation method. The classification parameter rates of the deep learning were the highest compared with the other classifiers, such as logistic regression and gradient boosting, especially using databases 1 and 2.

conclusionThe classification parameters of the selected method were the highest compared to other methods. Therefore, this method could be used to detect DR UTI in patients. This model showed potential for assisting in DR prediction. Therefore, physicians are helped to choose the right plan and improve patient health.