Out of the 782 papers that were considered, only 14 studies on AI models in the area of UTIs met the criteria for inclusion (Fig. 1). These can be grouped according to the scenarios the AI models were developed for, namely, (1) diagnosis of uncomplicated UTI and symptoms checkers, (2) diagnosis of complicated UTI, and (3) diagnosis of UTIs in specific population groups. Among models, 12 and two papers described machine and deep learning approaches, respectively. The most popular machine learning model was the artificial neural network (ANN) described in six studies, followed by extreme gradient boosting (XGBoost) (n = 3), support vector machine (SVM) (n = 1), CatBoost (n = 1), and ensemble learning model (ELM) (n = 1). Among predictive inputs, demographic parameters were used in 10/14 (71.4%) studies and mostly in the view of age (n = 9), gender (9), race, and weight (n = 1). Notably, the latter are included in papers with pediatric patients. Anamnesis was considered in 7/14 (50%) papers, namely, history of previous UTIs (n = 4), history of previous antibiotic treatment failure (n = 1), history of previous urine culture results (n = 1), and invasive urethral procedures (n = 1). Comorbidities were used in 3/14 (21.4%) studies: diabetes (n = 2), pneumonia (n = 2), classification of stroke (n = 1), and the presence of mixed cerebrovascular disease (n = 1). Logically, the last two were used in developing AI decision support for stroke patients. UTI-associated symptoms were included in 7/14 (50%) papers: dysuria (n = 4), fever (n = 3), suprapubic pain (n = 3), frequency and urgency (n = 1), pollakiuria (n = 1), and urine incontinence (n = 1). Urinalysis was used, and prognostic input was provided in 7/14 (50%) papers; two of them included dipstick tests only. When urine microscopy was used, red blood cells (RBC), white blood cells (WBC), bacteria’s presence, nitrites, epithelial cells, and glucose were analyzed in 2, 3, 2, 1, 1, and 1 studies, respectively. The study used urine cloudiness as one of its features. Imaging data were used in 3/14 (21.4%): one study analyzed the cystoscopic appearance of the lower urinary tract, and two papers described ultrasound imaging usage (for estimation of hydronephrosis and vesicoureteral reflux grades, respectively). Also, there were other inputs not related to the abovementioned groups: length of stay (LOS) (n = 3), length of urethral catheterization (n = 2), immunological urine markers (n = 1), ward (n = 1), serum creatinine and albumin (n = 1 and n = 1), glucocorticosteroid use (n = 1), and duration of immobility (n = 1). Performance metrics, validation type as well, and the abovementioned data arranged to include studies are discussed and presented in the review.

Research on AI-based models for uncomplicated UTI diagnosis and symptom checking is listed in Table 1. The study by Ozkan et al. [12••] sought to determine the accuracy of several artificial intelligence models in predicting the likelihood of cystitis and non-specific urethritis disorders, given similar symptoms from the urinary system. Anamnesis, urinalysis, and ultrasound results from 59 individuals were gathered as a training and validation dataset for the study. Four distinct artificial intelligence techniques were applied: decision trees (DT), random forests (RF), support vector machines (SVM), and artificial neural networks (ANN). When these models were compared, it became evident that ANN had the greatest accuracy for UTI detection, with a result of 98.3%. This ANN model only requires the variables pollakiuria, erythrocyturia, and suprapubic pain to acquire a diagnosis with comparable accuracy to a clinical-based diagnosis (Fig. 2).

PRISMA flowchart: AI-based approaches for UTI diagnosis

It was demonstrated that the ANN-based model structure could categorize UTIs without the requirement for expensive laboratory testing, ultrasounds, or invasive methods. Hence, it results in a cheaper diagnostic cost and a quicker decision-making process.

The motivation behind Gadalla et al.’s [13] paper is that women with uncomplicated UTI symptoms are frequently treated with empirical antibiotics, leading to antibiotic misuse and the development of antimicrobial resistance. The authors looked into 17 clinical and 42 immunological potential predictors for bacterial culture using a random forest or support vector machine (SVM) paired with recursive feature removal (RFE). The most effective clinical predictor to rule in and rule out UTI was urine cloudiness. Interestingly, adding the selected immunological biomarkers to the model with clinical features (including cloudiness or turbidity) did not improve the predictive properties. Dhanda et al. [14] described the NoMicro model, which does not take into account urine microscopy. Instead, the results of the urine dipstick test are used. Moreover, the authors generated NoMicro models based on several machine learning classificators, namely XGBoost, RF, and ANN, and compared their efficiency. The primary outcome was a pathogenic urine culture growing ≥ 100,000 colony-forming units. Predictor variables included age; gender; dipstick urinalysis nitrites, leukocytes, clarity, glucose, protein, and blood; dysuria; abdominal pain; and history of UTI. According to the results, the AUC of the NoMicro approach reached 0.85 in external validation and did not statistically differ from the version considering urine microscopy results. Arches et al. [15] described an application providing an analysis of the urine test strip using smartphones. According to the results, among the 65 participants, the confirmed UTI AI model achieved an overall accuracy rate of 96.03% and an overall reliability rate of ≥ 0.9, which is interpreted as excellent.

Research describing AI-based models for complicated UTI diagnosis is listed in Table 2. Møller et al. [16] aimed to develop two predictive models, using data from the index admission as well as historic data on a patient, to predict the development of UTI at the time of entry to the hospital and after 48 h of admission (HA-UTI). The ultimate goal was to assess the individual patient’s risk. The methodology included developing five machine learning models using features such as demographic information, laboratory results, past medical history, and clinical data. The unstructured features, such as the narrative text in electronic medical records, were preprocessed and converted to structured form by natural language processing. The area under the curve ranged from 0.82 to 0.84 for the entry model (t = 0 h) and 0.71 to 0.77 for the model predicting HA-UTI.

Taylor et al. [17] performed a single-center, multi-site, retrospective cohort analysis of adults who visited the emergency department based on urine culture results, clinical symptoms, and blood tests. Using both laboratory and clinical data, models for UTI prediction were created using six machine learning algorithms: RF, XGBoost, SVM, adaptive boosting, elastic net, and ANN. A full set of 211 variables and a reduced set of 10 variables (age, gender, history of UTI, dysuria, the presence of nitrites in urine, white blood cells (WBC), red blood cells (RBC), bacteria, and epithelial cells) were both used to develop the models. Comparisons between the UTI predictions and previously recorded UTI diagnoses were made. XGBoost, which has an area under the curve of 0.904, was found to be the best-performing method. It was also shown to have greater sensitivity when compared to the documentation of the UTI diagnosis. According to the results obtained, in practical application, approximately 1 in 4 patients will be re-classified from false positive to true negative, and 1 in 11 patients will be re-categorized from false negative to true positive on account of implementing the algorithm. Mancini et al. [18] created a machine learning model that can forecast a patient’s likelihood of developing a multidrug-resistant (MDR) UTI after being admitted to the hospital. The paper added a user-friendly cloud platform called DSaaS (Data Science as a Service), which is ideal for hospital organizations where healthcare operators might not have specialized programming language skills but need to analyze data, via machine learning techniques including CatBoost, SVM, and ANN. The paper employed DSaaS on a real antibiotic stewardship dataset. The development of an MDR UTI was predicted using data related to 1486 hospitalized patients, namely, sex, age, age class, ward, and time period. According to the results obtained, CatBoost exhibited the best predictive results, with the highest value in every metric used. Cai et al. [19] described two models based on ANN for predicting fluoroquinolone-based therapy failure (model 1) and fosfomycin-based therapy failure (model 2) among patients with recurrent UTI. Input data mostly consisted of previous urine culture profiles as well as types of antibiotic therapy failures. After the completion of the ANN learning and prediction processes, our neural network showed a sensitivity of 87.8% and a specificity of 97.3% in predicting the clinical efficacy of empirical therapy. Interestingly, the previous use of a specific class of antibiotic was not a risk factor for developing bacterial resistance to the same class (except for the fluoroquinolones), but instead, the most important risk factor for predicting resistance is the use of other classes of antibiotics.

Chen et al. [20•] compared models based on LR and ANN in defining UTI risk after cystoscopy to reduce antibiotic overuse. As input data, previous UTI history as well as cystoscopic findings such as benign prostatic hyperplasia (BPH), diverticulum, trabeculation, blood clot, cystocele, stone, and tumor was selected. The neural network model had a high accuracy of 85%, sensitivity of 80%, and specificity of 88%. Hong et al. [21] constructed a prediction model for urosepsis risk for patients with upper urinary tract calculi with the use of a machine learning ANN model. Several clinical and laboratory features, as well as a hydronephrosis degree based in the USA, were taken as predictive inputs. The area under the receiver operating curve in the validation set was 0.95. According to the results, the proposed model could provide risk assessments for urosepsis in patients with upper urinary tract calculi.

Papers describing AI-based models for uncomplicated UTI diagnosis in susceptible subgroups are listed in Table 3. Pregnant women and children represent a separate subgroup of patients more susceptible to UTIs and requiring specific diagnostic flow and treatment. Pregnancy immunologic and urinary tract alterations predispose women to UTIs. Progesterone-induced smooth muscle relaxation and gravid uterine compression cause ureter and renal calyces dilatation. Also, vesicoureteral reflux may occur. These modifications exacerbate urinary tract infections [22]. In turn, UTIs are among the most prevalent bacterial pediatric infections. They are equally prevalent in males and girls during the first year of life but become more prevalent in girls following the first year [23]. This high susceptibility makes the development of decision support models based on AI even more relevant. Bertsimas et al. [24] developed a machine learning model to better stratify pediatric patients with vesicoureteral reflux complicated by UTI according to the effect of continuous antibiotic prophylaxis. The authors used the following data as input: vesicoureteral reflux grade, serum creatinine, race/gender, fever, dysuria, and weight, and achieved an AUC of 0.82. The described model allows better identification of patients for whom continuous antibiotic prophylaxis will be more effective, thereby providing a personalized approach, while minimizing use in those with the least need. A study by Burton et al. [25] aimed at introducing a way to increase the efficiency of urine culture results among pregnant women and children by reducing the number of query samples to be cultured and enabling diagnostic services to concentrate on those in which there are true microbial infections.

This research discussed two methods of classification to test: one is a heuristic approach using a combination of features such as urine WBC and bacterial counts, and the second is testing typical machine learning models such as random forest, neural network, and XGBoost using independent features such as demographics, previous urine culture results, and clinical details as well. The most optimal solution found was three separate XGBoost algorithms trained separately for pregnant patients, children, and the rest of the categories. Combining the three models yielded a workload reduction of 41% and a sensitivity of 95% for each patient group. The work shows the possibility of using supervised machine learning models to improve service efficiency in situations where demand exceeds the number of resources available to public healthcare providers.

Immobile stroke patients also represent a highly susceptible patient subgroup. The prevalence of urinary tract infections is approximately 19%. In addition, the occurrence of an infection can exacerbate the physical harm caused by a stroke, forming a vicious circle with the stroke [26]. Zhu et al. [27] aimed to develop a prognostic model to define the risk of UTI among immobile stroke patients. Six machine learning models and an ensemble learning model were derived and evaluated. The latter achieved the best performance metrics both in internal and external validation sets, with an AUC of up to 0.82. Xu et al. [28] created an effective prediction model for identifying UTI risk in immobile stroke patients and compared its prediction performance to establish machine learning algorithms. They addressed this issue by developing a Siamese network that employed commonly used clinical criteria to identify patients at risk of UTIs. The model was developed and validated using a countrywide dataset of 3982 Chinese patients. A Siamese network is a deep neural network architecture with two or more identical subnetworks that are commonly employed in object detection. With an AUC of 0.83, the Siamese deep learning network did better than all the other machine learning–based models at predicting UTIs in stroke patients who were unable to move.