Table 1.

Summary of studies using machine language for prediction of antibiotic resistance for clinical decision-making.

Task	Input Data	Primary Models Used in the Study	Comparisons Models Used/Tested	Peformance/Accuracy	Comments/Limitations	References
Antibiotic resistance prediction in patients	11,496 antimicrobial susceptability datasets from laboratory information system, internal medical ward, public hospital, Greece	stack ensemble (Microsoft Azure AutoML)	VotingEnsemble, MaxAbsScaler, LightGBM, Sparse Normalizer, XGBoost Classifier	AUCW is 0.822 and 0.850 Accuracy rate is 0.770	Study uses 499 patients’ data. Data scientist is needed at this stage for pre-processing, feature selection, and final analysis. More data training needed to obtain more accurate results.	[107]
Prediction of antimicrobial resistance in Acinetobacter baumannii, Mycobacterium tuberculosis, and Streptococcus pneumoniae	K-mer representation of bacterial genomic data	random forest	Adaboost, logistic regression, deep learning	80–92%	Lower data give less accuracy and higher number of data give elevated accuracy. Not possible to use by laboratories and hospitals having small datasets.	[109]
Predicting antibiotic resistance in hospitalized patients	5590 instance datasets with different variables from general hospital laboratory, Greece	WEKA framework ML, run under Java platform	J48 algorithm, random forest, multinomial logistic regression, kNN algorithm, multilayer perceptron (MLP)	ROC area of 0.758 and accuracy of 75.8%	Low accuracy, limited dataset, and less clinical attributes. Increased dataset can give accurate results.	[110]
Prediction of antibiotic resistance in hospitalized patients	16,000 antibiotic resistance tests of electronic medical record (EMR)	Ensemble of 3 models: Lasso logistic regression, neural networks, gradient boosted trees	Independent algorithms:Lasso logistic regression, neural networks, gradient boosted trees	Combined algorithm: 0.821, Xgb: 0.82, Lasso: 0.82, dnn: 0.803, auROC score was 0.8–0.88	Bacterial details can increase AUROC score if included, additional information needed to improve accuracy such as antibiotics used prior to admission, microbiome composition, diet, and exercise.	[111]
Prediction of antimicrobial resistance from ICU patients in Pseudomonas and Enterococcus, Stenotrophomonas	Dataset of 32,997 collected from health information system of 2630 patients, University Hospital of Fuenlabrada, Spain	Logistic regression, K-nn, decision trees, random forest, multilayer perceptron		AMG: 82.2 ± 1.7, CAR: 79.6 ± 2.1, CF4: 74.9 ± 2.1, PAP: 77.1 ± 1.7, POL: 68.5 ± 7.0, QUI: 88.1 ± 1.6	Accuracy differs based on various antibiotics and bacterial species; upgradation yet to undergo based on mechnical ventilation, bed; sepsis patients in ICU are to be considered.	[106]
Prediction of Carbapenem-resistant Klebsiella pneumoniae	46 Carbapenem-resistant Klebsiella pneumonia (CRKP) isolated from hospital patients	random forest	Logistic regression, Naïve Bayes, nearest neighbors, support vector machine	Accuracy: 97%, Carbapenem-resistant identification: 93%	Small sample size and limited data of CRKP.	[112]
Prediction of antimicrobial phenotype resistance of Staphylococcus aureus	K-mer representation of bacterial genomic data	Random forest, SVM, XGBoost		Among 10 anitibiotics used in the study, AUC was recorded between 82.02% for Linezolid and 96.13 for vancomycin. Cefoxitin registers AUC of 92.65% with sensitivity of 94% and major error 6.82%	Study uses 466 whole genome sequencing results to predict the antimicrobial resistance.	[108]