Artificial Intelligence (AI) has been widely regarded as the future of the twenty-first century. AI is used in every primary industry, including the healthcare industry, to advance health care and improve patient outcomes. Literature has often debated the definition of AI as there is a lack of clarity around what AI entails. Today we refer to the umbrella term of AI as “machines that respond to stimulation consistent with traditional responses from humans, given the human capacity for contemplation, judgment, and intention” [16]. There are six main subsets of artificial intelligence, as described in Table 2 [17] .

Artificial intelligence has become a revolutionary intelligence system in the healthcare industry in recent years—with improvements being made in diagnosis, patient monitoring, and precision-based medicine. Table 3 describes certain AI-based healthcare technologies that utilize specific subsystems of AI—as mentioned by Wahl (2018) and Davenport (2019) [18] [19].

A study conducted in 2016 revealed that physicians spend 27% of their office days on direct patient interaction and spent nearly 50% of their days keeping up with electronic hospital records (EHR) and clerical desk work [20]. AI has been implemented to help primary care physicians in this scenario by helping take notes, analyze discussions with patients and enter required information directly into EHR systems, which has reduced manual labor for physicians, and as an effect, increased meaningful patient-physician interactions [21]. AI is rapidly growing, and investment in AI technology for the medical industry is expected to grow from $2.1 billion to $36.1 billion by 2025 as more advancements are made in this field [22]. AI will expand in the healthcare industry to support physicians, increase accuracy, and reduce medical errors in various ways.

In women’s health, including obstetrics and gynecology (OBGYN), the use of artificial intelligence has allowed improvements in protocol efficiency and delivery. Examples of general applications of AI in women’s health include fetal heart monitoring, gestational diabetes mellitus, and in vitro fertilization. A summary of these applications is given in Table 4.

In diagnosis and management, populations in low socioeconomic backgrounds do not have the means to afford treatment or have an unawareness of why they are suffering from dysfunction in the bladder and pelvic regions that cause urinary incontinence. Enabling artificial intelligence to diagnose patients with UI would improve healthcare access and ensure the quality of diagnoses for the underrepresented populations. The AI system would rely on current UI diagnostic modalities for its input system.

In Garcia del Salto’s (2014) study, they noted that pelvic floor prolapses were frequently misidentified and proposed to use magnetic resonance imaging (MRI) that would visualize the pelvic floor’s compartments to achieve more accurate diagnoses [34]. In Wei et al. (2017), they argued that although MRI would deliver high-quality images and data for patient-specific 3D reconstruction of the female pelvic floor for diagnoses, its utilization may be restrained due to the high expense and limited availability [35]. Ultrasound (US) imaging however is a low-cost modality that can be used to help build 3D models of the pelvic floor. Wei et al. (2017) proposed to utilize the high-resolution endo-vaginal ultrasonography to create accurate pelvic floor models to evaluate the pelvic floor dysfunctions and reduce injuries [35]. Emerging technologies utilizing ultrasound were highlighted by Shobeiri and Manonai (2014) including transperineal ultrasound (TPUS) to view the pelvic floor region with anatomical access, and the endocavitary ultrasound and real-time elastography (RTE) that uses the differences in the radiofrequency signals and differences of elastic properties in tissues to yield additional information about anatomical structures to enhance the information of the image [36]. Other emerging techniques were mentioned by Kim, Egorov, and Shobeiri (2017) including using Ultrasound Elasticity Imaging to provide visual information of the elasticity and anatomical structure, Acoustic Radiation Force Impulse (ARFI) and Ultrasonic Shear Wave Elasticity Imaging (SWEI) that are operator-independent and provide images of relative difference in tissue difference, and Functional Tactile Imaging (FTI) which would translate muscular activity for dynamic pressure patterns with respects to time [37]. An example of a device using FTI is the Vaginal Tactile Imaging (VTI) probe which acquires pressure response from the vaginal wall [37]. Instances, where the FTI modality including the VTI probe can be used, is to study the voluntary and involuntary reflex contractions, involuntary relaxation, or a Valsalva maneuver [37]. Table 5 compares the practicality of UI diagnosis methods to use with artificial intelligence in low-resource locations.

Ultimately, after diagnosis, the physician would provide a treatment plan to the patients, which can be challenging considering the limited availability of resources in areas with low socioeconomic status. To assist the physicians to deliver specific treatment and management protocols (whether it be drugs, therapeutic devices, or surgery), precision or personalized medicine using AI systems can be used using the patient’s data against large datasets to individualize care – from the diagnosis to treatment selection and control, which would be optimized when new therapies and diagnostic methods come about [38].

For an AI System intended to be used on the diagnosis and management of UI, it needs to be able to parse the input data for it to accurately report diagnosis and potentially provide intervention and therapeutics to the patient. This means that the system needs to be sensitive and adaptable to different situations, such as in the case of patients residing in low socioeconomic regions where certain factors, either patient specific or availability of medical techniques and knowledge can influence the decision outputs of the system.

In a study by Burton (2019), three different machine learning algorithms (random forest, neural network, and extreme gradient boosting) were used to screen samples of urine for the detection of urinary tract infection (UTI) while factoring in demographics, historical urine culture results, and clinical details in which 212,554 reports were generated with classification sensitivity above 95% [39]. The study demonstrated that AI systems could reduce the diagnostic workload without compromising UTI detection; the same machine learning algorithms can be adapted to the detection of urinary incontinence. Barzegari’s (2018) study investigated urethral pressure with particular parameters on patients with stress UI and ultimately found that if the trained artificial neural network (ANN) model were utilized, it would be an effective means to diagnose patients with stress UI [40]. In Geramipour’s (2015) study with neural networks, real-time estimation of the bladder pressure using ANN was proposed and measurements using electroneurograms (ENG) on three rats, in which the neural network model can provide accurate assessment and prediction of bladder pressure and then stimulate it to increase bladder capacity to reduce incontinence [41]. In Luarikkala (1999), a system based on genetic algorithms, Galactica, was developed to devise accurate diagnostic rules for stress, mixed and urge UI. Being trained on 1105 datasets with testing performed on 485 datasets, the Galactica had a mean prediction accuracy of UI diagnosis of 90%, with mean descriptive accuracy of UI diagnosis of 88% [42]. Galactica was compared against other models such as k-means cluster analysis, C.4.5 (rule-based and decision tree-based), and a random bit climber. The C.4.5 (rule-based) had the highest mean prediction and descriptive accuracy with 91% and 92%, respectively. Galactica was compared with C.4.5 (rule-based) and was similar in their mean accuracies [42]. The Galactica system, although performed well, hasn’t been used recently in clinical studies, which suggests that this system would not be of much use in an AI system being deployed in low-resource locations [42]. Lopes (2009) referenced three main categories of AI systems for diagnosis of UI including rule-based logic systems such as ALTURIN.EXP using EXPERTMD and ALTURIN.SDD (sensibility and specificity upper to 98%), genetic algorithm-based systems such as Galactica (mentioned above from Laurikka, 1999), and fuzzy logic-based approved by NANDA (determine diagnosis in total accordance with panel for 79.5% of cases with 1.5% error) [43]. Hung (2019) utilized DeepSurv’s deep learning model to predict UI after a robot-assisted radical prostatectomy (RARP) with performance metrics coming from 100 patients, with a mean absolute error of 85.9 and confidence interval of 0.6 [44]. Karam et al. (2016) developed a fast, adaptive, and accurate real-time method, called Context-Aware Thresholding (CAT) for classifying bladder events with pressure data acquired from 14 human subjects [45]. The CAT algorithm was tested against other algorithms for accuracy—Static Detrusor Thresholding (SDT), Global Detrusor Thresholding (GDT), Global Vesical Thresholding (GVT), Hybrid Detrusor Thresholding (HDT), and Adaptive Vesical Thresholding (AVT) [45]. In their results, the CAT algorithm displayed high accuracy of 97% for overactive bladder event classification with a 1.34% false positivity rate. AVT also displayed accurate results with 95% classification and a 1.39% false positivity rate. The other algorithms tested (SDT, HDT, GDT, and GVT) had an average of 74.25% accuracy with high false positivity rates of 20.88% (SDT) and 18.59% (HDT) [45]. In Huang (2007), a computer-aided vector-based perineal ultrasound was used with an MLP Neural Network to diagnose urodynamic stress incontinence (USI) amongst 48 women with 36 of whom had USI. The model was trained with a backpropagation algorithm to minimize output distortion and then validated using the k-fold cross-validation for performance analyses [46]. The system attained accurate results with 91.7% accuracy of diagnosis USI with the computer-aided perineal ultrasound [46].

The studies above suggested that these systems can be trained over time, which means that the speed of the system could improve, the number of false negatives could decrease, and therefore an increased accuracy in diagnosis and determination of proper intervention methods [47]. Table 6 represents a summary of most of the AI systems mentioned above which may be useful to deploy to low-resource locations due to overall good performance results and the ability that these systems have to train over time.