Open access peer-reviewed chapter - ONLINE FIRST
Levels of autonomy in surgical robots.
Abstract
As the quantity and complexity of data and information needed to care for patients increases, the need for efficient and complex computing solutions in healthcare rises. Being heralded as the answer, artificial intelligence (AI) continues to make inroads into various aspects of medicine and patient care. The surgical specialty of urology is one of the earliest adopters of new technology and AI, particularly in the area of surgical robotics. The introduction and growth of novel technological applications promises to transform urology, making patient care safer and more efficient. At the same time, technological evolution is not “risk-free” and may be controversial. This chapter will review the advances and developments in AI as it specifically pertains to the field of urology.
Keywords
- artificial intelligence
- urology
- robotic surgery
- machine learning
- deep learning
- neural networks
1. Introduction
An English mathematician and pioneer of computer science by the name of Alan Turing is credited for sparking initial assessments of artificial intelligence (AI) with the simple question: “can machines think [1]?” Mr. Turing’s imitation game, the so-called “Turing test,” pits man against machine to determine whether machine could convincingly disguise itself as human. Now, in more modern times, AI appears more ubiquitous with each passing day. The question has since shifted from the philosophical, “can machines think?” to that of mundane tasks, “can machines paint me a picture?,” and “can machines do my work?” Regarding healthcare, patients, physicians, and hospital administrators alike seem to pose a different question: “can machines do a doctor’s job?”
There have been numerous developments in computer and information science which have allowed what was once considered “science fiction” to become “science fact.” These advancements, such as machine learning (ML) and deep learning models, have helped establish AI in healthcare with various applications. The utility of these machines may range from organization of clinic appointments to pattern recognition and disease detection on radiographic imaging or pathology slides, and even providing fine-tuned, highly nuanced assistance to surgeons in the operating room [2]. However, the adoption of AI is not uniform across all institutions, and the utilization may vary greatly between subspecialties even within the same hospital or health network. In the field of urology, there has been widespread implementation of machine applications to improve patient care, outcomes, and overall efficiency [3].
While principles/components of AI are used in technology we have today, some of what is included in AI is better classified as cybernetics or intelligent automation or augmentation (IA). Cybernetics is “the science of control and communications in the animal and machine,” functions mainly on feedback and involves interaction between actors (i.e., a robot and surgeon), while intelligent automation is the use of algorithms to automate tasks [4, 5]. Intelligent augmentation improves human ability using AI systems [5]. Indeed, the AI aspects of learning, reasoning, problem solving, and decision making are inherent within the existing technology, yet currently there is no product within the medical field that functions fully independently of human interaction or oversight. The present state of the art is more of a hybrid model. Further in the future would be the potential for fully autonomous technologies and devices.
Here, we explore the possibilities of AI in healthcare and take a finer look at its implementation in the field of urology. We will discuss the application of AI in clerical and administrative duties, diagnostic programs, prognostic predictors, and the operating room itself (see Summary Table in Appendix). Throughout, we will provide relevant synopses of up-to-date studies of applied AI, highlighting both success and failure. Afterward, we will consider common concerns against AI and attempt to navigate a path forward despite overarching uncertainties.
Advertisement
2. Administrative uses
AI is revolutionizing administrative roles in healthcare by automating and optimizing key processes, reducing costs, and improving efficiency. AI-driven systems, such as chatbots and rule-based expert systems (RBES), streamline workflows in divisions such as patient care, billing, and resource management. AI has been utilized in a domain known as Expert Systems (ES), which has been used since the early 1960s and continues to be utilized due to its consistency and completeness [6]. From here, there have been branches of computer systems specifically created for medical billing: simple systems designed to work with standard software such as Microsoft™ Excel (Microsoft Corporation, Redmond, Washington, USA), specialized systems that often integrate databases, and Expert-like systems programmed with the knowledge component. AI’s integration in billing practices through natural language processing (NLP) and ML enhances accuracy in generating procedural codes, decreasing errors, improving claim acceptance rates, and mitigating revenue loss [7]. Specifically, codes have already been developed that utilize AI to complete a first pass of submitted documents to search for specific errors and filter results based on the particular edit code. For example, 3M Audit ESTM (3M, Beirut, Lebanon) is a modernized tool for analyzing medical claims and has developed an ability to streamline auditing, reporting, and management. It is designed to evaluate consistency with medical coding guidelines and cross-reference of different aspects of a patient claim, including lengths of stay and its related charges [8]. AI has also been utilized to optimize billing practices by interpreting operative notes to generate and classify current procedural terminology (CPT) codes (American Medical Association, Chicago, Illinois, USA) and accelerate the process, complemented by ML. Specifically, CPT code descriptions are often complemented with details from insurance policy documents. This may include details regarding a specific procedure as well as support for its reimbursement based on medical evidence [9]. NLP has been used to analyze clinical documentation in electronic medical records (EMR) to create codes for diagnoses and comorbidities, thus smoothing out the billing process and workflow. NLP-based systems improve revenue cycle management and have even been reported to have accuracy as high as 96% [7]. Through this, AI continues to not only revolutionize the efficiency of such systems, but also advance its functionality and applicability within healthcare, estimated to have saved $300–$450 billion annually in 2013 [10].
The use of AI does not stop at billing. In fields like urology, chatbots can address administrative burdens by triaging patients, managing follow-up care, and handling symptom prioritization. One of the key advantages currently being evaluated is the ability of AI-driven chatbots to provide individualized medical advice to patients. AI-based chatbots in urology such as PROstate cancer Conversational Agent (PROSCA) (Systems Applications and Products in data processing Societas Europaea, Walldrof, Germany) and Sneha AI (SnehAI) (Population Foundation of India, New Delhi, India) may aid in improving decision-making by cross-referencing patient data and medical records, providing evidence-based recommendations for the diagnosis and treatment of patients [11]. Chatbots may be deployed as tools for urological symptom checking, health screening, patient education, counseling, lifestyle change, and post-treatment follow-up care [12]. These tools reduce unnecessary in-person visits, saving patients and providers valuable time and resources while ensuring continuity of care. By optimizing these processes, AI empowers administrators to shift focus toward strategic planning and improving the patient care experience.
It is no secret that there is still much room to grow. AI systems are not without flaws, and, especially in healthcare, it is of utmost importance to ensure that patient safety is not compromised for efficiency. Ensuring reliability and accuracy remains one of the most crucial factors for the success of AI and its chatbots. Equally important is ensuring patient privacy and security. With much of the current AI software storing information in cloud storage and within its memory system, patients must be assured their personal information remains restricted and inaccessible to unauthorized users. Their successful implementation requires meticulous design, seamless workflow integration, and ongoing system updates to maximize functionality and user adoption.
Advertisement
3. Diagnostics
The key to successful treatment of any disease is an accurate diagnosis. This is often not as straightforward as desired, given the nuances and variations in patient presentation coupled with genetic and environmental influences. Multiple pieces of data must be interpreted and combined to arrive at a diagnosis, and as a result this process can be time consuming. Using AI solutions such as ML, deep learning (DL), and convolutional neural networks (CNNs), these data can be rapidly processed in an accurate manner, leading to a precise and quick diagnosis. These solutions are being applied throughout urology to improve oncologic, stone disease, and renal pathologic diagnoses.
Within uro-oncology, the detection and depiction of prostate, renal, and bladder cancers have shown improvement when AI solutions/models are implemented. AI algorithms combined with prostate imaging-reporting and data system (PI-RADS) led to improved sensitivity of prostate and prostate capsule anatomy detection on magnetic resonance imaging (MRI) by 78% [13]. This increased sensitivity aids in fusion biopsy and brachytherapy planning [13]. Wang et al. conducted a multicenter trial of AI ultrasound of the prostate (AIUSP) versus transrectal ultrasound-guided systematic biopsy (TRUS-SB) and multiparametric MRI (mpMRI) and found that AIUSP significantly increased the detection rate of prostate cancer on subsequent biopsy. The detection rate was 49.6% using AIUSP, compared to 34.6% with TRUS-SB and 35.8% using mpMRI [14]. AI algorithms have also been used in Gleason grading. The PANDA trial showed that AI algorithms performed at the level of expert pathologists with agreement levels of 0.862 and 0.868 on US and European datasets [14]. DL algorithms have been found to outperform general pathologists in Gleason scoring, with an accuracy of 0.70 compared to a 0.61 rate from the general pathologists. The accuracy was validated using review by expert uropathologists [15]. CNNs are also used to increase the accuracy of Gleason score assignment.
Similarly, the detection of bladder and urinary tract cancers is positively impacted by the use of AI. Cystoscopy is a prominent diagnostic tool for bladder neoplasms; however, it is difficult to determine flat tumors with this modality [16]. The use of blue light as opposed to white light during cystoscopy has been touted to increase accuracy, but this is currently under scrutiny [16]. Various AI solutions have been applied to both methods with encouraging results. CNNs have been used to detect flat lesions seen in white light cystoscopy. Specifically, the cystoscopy AI diagnostic system (CAIDS) was applied to bladder tumor identification and achieved an AUC of 0.94 [16]. Yoo et al. created an AI-enhanced platform able to predict the grade of bladder neoplasms with ≥98% accuracy for low grade or benign tumors and >90% accuracy compared to white light cystoscopy [16]. A separate platform piloted by Ali et al. predicts malignancy, invasiveness, and tumor grade of bladder cancer from blue light cystoscopy images [16]. The system exhibited 95.77% sensitivity and 87.84% specificity for cancer diagnosis, along with 88% sensitivity and 96.56% specificity for invasiveness [16]. There is even a DL platform for smartphones named EasyDL which diagnoses bladder cancer using cystoscopic images uploaded to the app with 96.9% accuracy [16]. CNNs are also in development for use in diagnosis of upper tract urothelial carcinoma on flexible ureteroscopy [14]. The lymph node metastases diagnostic model (LNMDM) created by Wu et al. outperformed junior and senior pathologists in detecting lymph node metastases from bladder cancer on histologic slides. AI assistance also helped to increase sensitivity and diagnostic accuracy by detecting micrometastases the pathologist had not seen [14].
AI applications for radiology and radiomics are aiding the diagnosis of renal cancers and nodal metastases on cross-sectional imaging. The Houshyar group developed a CNN capable of distinguishing renal tumors on single-phase contrast CT scans and calculating tumor volume [14]. It achieved Pearson correlation coefficients of 0.998 and 0.993 for kidney segmentation and tumor segmentation, respectively, relative to human estimates [14]. One important component of diagnostics is distinguishing between benign and malignant conditions, which prevents unnecessary interventions. Pedersen et al. developed CNNs that differentiate benign oncocytoma from renal cell carcinoma (RCC) on CT scans with 93% accuracy and 94% specificity [14]. A predictive model containing a random forest ML algorithm made by Erdim et al. was able to correctly determine between benign and malignant tumors on CT imaging with 91% accuracy and a 0.915 AUC [14]. CT radiomics with ML algorithms has been applied to identifying lymph node metastases. Baessler et al. found the application to be 81% accurate, 88% sensitive, and 72% specific in discriminating between malignant and benign lymph nodes [14].
Not all AI solutions are viable or superior to physicians or current diagnostic methods. Although there are a myriad of DL systems and AI algorithms for diagnosing bladder cancer on CT and MRI, their widespread implementation is hindered by low sensitivity and lack of clearly defined radiologic features necessary for reaching a correct diagnosis [13, 16]. DL and ML models created for automatic histopathologic slide reading of bladder cancers were found to perform worse than expert uropathologists in detecting non-muscle invasive bladder cancer (NMIBC) [16]. AI applications for histologic detection of renal cell carcinoma also perform worse than pathologists [14]. These technologies are also limited in availability and most of the histopathologic research thus far has been retrospective [16].
Advertisement
4. Prognostics and clinical prediction
With access to large datasets, AI models will be able to provide powerful, individualized predictions for various urological applications, including providing clinicians and patients with improved prognostication for common urologic diseases.
4.1 Infertility
Infertility is a condition affecting approximately 8–12% of couples worldwide, with an estimated 40–50% of cases involving male factor infertility [17, 18]. As such, the work-up of infertility warrants evaluation of the male partner
The term “omics” is defined as the investigation and assessment of substantial quantities of data that represent the make-up as well as the function of a biological system at a set level [27]. More specific to medicine are large data measurements of molecular markers (genes/genomics, proteins/proteomics) and processes (transcription/transcriptomics, metabolism/metabolomics) to gain insight into physiological function [27, 28]. There is increasing interest in the use of omics models to evaluate male fertility, such as through the measurement of proxies for DNA structural integrity. For instance, terminal deoxytransferase-mediated deoxyuridine triphosphate nick end labeling assay (TUNEL), 7,8-dihydro-8-oxo-2′-deoxoguanosine (8-OHdG), and sperm chromatin compaction have all been investigated as possible markers for male fertility with varying degrees of efficacy [28, 29, 30, 31]. Tang et al. used whole genome sequencing to identify specific gene variants commonly found in patients with NOA [32]. Another study used gene microarray testing to identify 215 genes differentially expressed between patients with and without NOA [33]. Other omics fields like epigenomics, transcriptomics, and proteomics studies have investigated the use of various epigenetic markers, RNA expression, and protein levels, respectively, as ways to gain insight into sperm function on the cellular level for fertile and infertile patients [20, 34]. Thus, omics models show promise in the evaluation of male fertility and may be able to elucidate the proportion of viable and nonviable sperm in infertile patients better than traditional semen analysis. However, various shortcomings still exist with the current state of omics in the evaluation of human fertility. While studies have investigated many individual markers in the evaluation of fertility, literature looking at the interplay between markers in different omics fields—and even within the same field—is in its infancy [20]. Additionally, omics models are currently cost- and time-prohibitive, though this is likely to change in the future with continued innovation [34]. Omics represents a field ripe for optimization with AI models.
Machine learning may be useful not only for evaluating for male infertility, but these modalities are also being investigated in the treatment of male infertility. One of the mainstays of treatment for NOA is testicular sperm extraction (TESE), a procedure that directly extracts sperm from the testis for use in assisted fertilization [35]. Microdissection TESE (mTESE) is a form of TESE involving 20× magnification direct visualization of spermatozoa in the testis during the procedure. However, shortcomings of surgical sperm extraction exist, resulting in extraction rates of approximately 38–50% [36, 37]. Because of this, there is interest in the use of AI/ML to stratify patients that may respond to treatment. Bachelot et al. created ML models trained on retrospectively collected demographic, hormonal, and fertility characteristics to predict response to TESE in patients with NOA. The best performing model identified serum inhibin B concentration, serum prolactin concentration, age, and history of varicoceles as factors most associated with failure of TESE; the model had AUC-ROC of 0.90, sensitivity of 100%, and specificity of 69.2% [38]. In the future, the discriminatory ability of similar models may be improved with the use of omics. Various biomarkers have shown promise in stratifying successful sperm extraction
Besides assisting in the stratification of patients with NOA looking to undergo surgical sperm extraction, ML models will be able to assist the extraction procedure itself. Computer vision models trained on videos captured during mTESE were able to identify viable sperm with 1000-fold greater detection speed compared to manual identification by an embryologist [42]. Although precision was higher for manual identification compared with the model (98.18% vs. 89.58%, p < 0.0001), this study indicates that computer vision models may be able to assist manual detection of viable sperm, and, with improvements in the model, they may be able to help automate detection and extraction [42].
4.2 Nephrolithiasis
Kidney stone disease incidence has increased over time and AI may prove useful in predicting the development of nephrolithiasis. Models using patient demographics, lifestyle factors, and health data have been able to predict the development of kidney stones with 88% accuracy [43]. Other investigators have analyzed whether AI models can be used to predict spontaneous stone passage. Solakhan et al. used a neural network model trained on demographics, imaging characteristics, and disease characteristics from patients that passed stones compared with patients that were unable to spontaneously pass their stone. Based on their highest performing model, they were able to predict whether a patient would have spontaneous stone passage with 92.8% accuracy [44]. In patients requiring operative management of nephrolithiasis, AI programs show promise in predicting successful outcomes over 95% of the time for percutaneous nephrolithotomy in internal validation studies [45]. AI has also been applied to prediction of outcomes for ureteroscopy (URS) in the treatment of urolithiasis. Sepsis due to a urinary source is a grave potential complication of ureteroscopy that affects between 0.5% and 11% of patients [46]. Unfortunately, while risk factors such as positive preoperative urine culture and female sex have been identified for the development of sepsis following ureteroscopy, clinical tools helping risk stratify patients for developing sepsis are lacking [47]. Pietropaolo et al. created a multivariate model to risk stratify the development of sepsis in patients undergoing ureteroscopy. Their model was accurate 81.3% of the time, and proximal stone location, stent time, size of stone, and total procedure time were most predictive of developing sepsis after URS [48]. Using these models, clinicians will be better equipped to plan for postoperative complications and improve disposition planning for patients at higher risk of sepsis after endourologic intervention.
4.3 Bladder cancer
Bladder cancer is a urologic cancer with significant morbidity and mortality. Five-year recurrence rates have been reported from 50% to 70% for non-muscle-invasive bladder cancer (NMIBC) [47]. Muscle-invasive bladder cancer (MIBC), characterized by invasion into the muscle layers of the bladder, is less common yet more aggressive than NMIBC. Even after definitive treatment with radical cystectomy/urinary diversion or chemoradiation therapy, the 5-year survival rate for MIBC is estimated to be between 50% and 70% [47]. Given the complex course of bladder cancer, there is increasing interest in the use of AI to help predict bladder cancer outcomes. Bhambhvani et al. developed a neural network model trained using demographic, tumor, and treatment variables to predict 5-year survival rates for patients with urothelial bladder cancer. Their model was able to predict 5-year overall survival with AUC of 0.70, and 5-year disease-specific survival with AUC of 0.81 [48]. Instead of using clinical data, other authors have used AI to risk stratify bladder cancer based on immunohistochemical data. Gavriel et al. created a ML random forest model trained on the presence and distribution of cancerous urothelial cells, tumor-infiltrating lymphocytes (TILs), tumor-associated macrophages (TAMs), and programmed death ligand 1 (PD-L1) on multiplexed immunofluorescence slides [49]. Their model was able to predict 5-year survival with AUC of 0.89, compared with AUC of 0.64 using standard tumor, nodes, metastasis (TNM) staging criteria [49]. Given the high recurrence rates of bladder cancer, AI has been implemented to help predict bladder cancer recurrence following treatment. Hasnain et al. created multivariate models using various anatomic, histological, and pathological variables to predict 5-year recurrence and survival. Based on their highest performing model, the variables most associated with bladder cancer recurrence within 5 years were higher pathologic stage subgrouping (i.e., whether the bladder cancer was organ confined, extravesical, or node positive), higher pathologic stage (pT) stage, and greater number of positive lymph nodes [50]. Their model had slightly improved overall performance in predicting 1-, 3-, and 5-year recurrence and survival compared to single factor predictors such as pathologic stage subgrouping or pT stage; for instance, the F1 value for 5-year recurrence was 0.636, compared with 0.623 and 0.583 for pathologic stage subgroup and pT stage, respectively [50]. Other models have been created which show greater predictive capabilities, though they have not been extended to 5-year recurrence. A radiomics model trained on preoperative MRIs of bladder cancer was able to predict 2-year recurrence with AUC of 0.81 [51]. Another random forest model trained on immunohistochemical stains of NMIBC was able to predict 1-year recurrence with AUC of 0.86 [52]. AI enhanced cystoscopy, as discussed earlier in the chapter, can also be applied to post-operative evaluation for bladder cancer recurrence.
While AI has shown promise in stratifying bladder cancer outcomes, it has also been used to predict the outcomes for the specific treatments of bladder cancer. Among the treatments for bladder cancer is radical cystectomy, a complex procedure with high complication and readmission rates [53]. Because of the highly variable postoperative course of radical cystectomy, there is interest in the use of AI to assist in prediction and risk stratification of immediate postoperative outcomes, such as infection, minor complications, and major complications involving cardiac arrest, venous thromboembolism, reoperation, and death [54]. Efforts to broadly categorize patients as high risk for minor or major complications only showed moderate predictability, with the highest performing model demonstrating AUC of 0.67 [54]. Other studies have used similar models to predict in-hospital and 90-day postoperative mortality with reasonable success, achieving AUC of 0.811 [55]. Zhao et al. investigated the role of AI in predicting whether patients undergoing radical cystectomy would be discharged home or to higher care facilities. Their model had AUC of 0.80, which outperformed prior attempts at similar models [53]. Though radical cystectomy for the treatment of bladder cancer is a complicated procedure with a similarly complicated postop course, attempts to use ML to assist in prognostication of postoperative care have shown some success.
Chemotherapy is often used as a neoadjuvant therapy to improve survival after radical cystectomy. AI has been investigated to improve patient selection for neoadjuvant therapy to ensure tumor response and decrease potential adverse effects in patients with non-responsive bladder cancer. Cha et al. created a radiomics model to discriminate between bladder cancers that responded neoadjuvant chemotherapy, defined as stage T0 posttreatment, and non-responders to chemotherapy, defined as any residual tumor. Based on pre- and post-neoadjuvant chemotherapy CT images, they predicted pT stage at the time of cystectomy. Their model showed equivalent predictive power compared with the expert radiologists in the study (AUC = 0.73 vs. 0.76/0.77) [56]. Omics models have also helped identify non-responders to neoadjuvant chemotherapy for bladder cancer. For instance, Hepburn et al. utilized a genomics model to identify genes that were upregulated in non-responders. In their study, CNGB1 was most associated with non-responders, which was corroborated in an animal model experiment with CNGB1 knock-out cells showing significantly decreased growth in the presence of cisplatin [57, 58, 59]. AI shows promise in the prediction of bladder cancer outcomes and treatment response.
4.4 Prostate cancer
Prostate cancer is one of the most common causes of cancer in men and is the second leading cause of cancer deaths in men [60]. Identifying molecular markers may help clinicians predict progression of prostate cancer. For instance, ERG gene fusion is associated with more aggressive prostate cancers, higher recurrence rates, and lower survival rates [61, 62]. However, current methods involving genetic sequencing are expensive and time consuming [63]. The use of AI to help bridge this gap has been investigated using machine learning models in predicting gene mutations and expression using pathology slides. Dadhania et al. developed a CNN that was able to predict ERG gene fusion status based on hematoxylin and eosin (H&E) stained prostate cancer samples with AUC of 0.823–0.851 depending on image magnification [64]. Other authors have validated the ability predict ERG gene fusion status using histopathological imaging, as well as predict the loss of the tumor suppressor gene PTEN, which is associated with worse outcomes in prostate cancer [65].
Radical prostatectomy is one of the first-line treatment options for patients with localized prostate cancer [66]. Given the intimate relationship between the prostate, the external urinary sphincter, and the cavernous nerve bundles, potential risks of radical prostatectomy include erectile dysfunction and urinary incontinence. Several studies have investigated the use of AI in risk-stratifying patients for the development of these potential outcomes. Ma et al. developed a model based on surgical gesture timing and sequencing during robotic radical prostatectomy (such as spreading tissue, hooking tissue, burn dissecting) that was able to predict development of erectile dysfunction better than demographic and clinical features alone (AUC 0.77 vs. 0.69) [67]. Other studies have focused on the use of clinical data to predict postoperative outcomes. A deep learning model created by Hung et al. identified that operative data, such as the length of time without movement of the third robotic instrument during vesicourethral anastomosis and total procedure length, was significantly correlated with return of urinary continence following robotic-assisted radical prostatectomy [68].
Radiotherapy is another treatment option for localized prostate cancer, and its effectiveness in some cases is complimented by androgen deprivation therapy (ADT) to decrease serum testosterone levels [69, 70]. However, ADT comes with its own risks, including hot flashes, muscle loss, osteoporosis, and cardiovascular events [71]. To date, a validated, individualized method for determining which patients would benefit the most from ADT therapy remains lacking. However, AI may be able to provide individualized predictions about ADT use for the treatment of prostate cancer. Spratt et al. created an ML model trained on pretreatment histological data and clinical data from five clinical trials to predict which patients would have decreased metastasis rates and prostate cancer-specific mortality with the use of ADT. Their model was then tested on patients from another clinical trial. They found that their model was able to successfully predict which patients would show clinical benefit with ADT alongside radiotherapy compared with radiotherapy alone. In patients identified by their model to benefit from ADT, the trial data showed that the hazard ratios for the development of distal metastasis and mortality were 0.34 and 0.28, respectively [72].
Despite treatment, up to 20–50% of prostate cancers develop biochemical recurrence (BCR) as measured by prostate-specific antigen (PSA) [73]. As such, improved detection or prediction of prostate cancer recurrence may help to decrease prostate cancer mortality. Huang et al. created a CNN trained on H&E histological images of prostate biopsies to characterize regions of interest (ROIs) patterns that were associated with BCR. Their model showed an AUC of 0.78, which indicated superior performance compared with Gleason Grade Group (AUC = 0.62), pT stage (AUC = 0.64), and margin status (AUC = 0.51). When also incorporating these clinical characteristics, the model’s performance was further improved to an AUC of 0.84 in the combined model [74].
4.5 Renal cancer
Renal cell carcinoma (RCC) comprises approximately 5% and 3% of all cancers in men and women, respectively [75]. AI models show promise for renal carcinoma. Schulz et al. created an AI model trained using histochemical and radiological images of clear cell renal cell carcinoma to predict survival time post diagnosis, achieving a mean C-index of 0.779. They also used their model to predict 5-year survival, achieving an AUC of 0.916 [76]. Interestingly, the addition of genomic parameters did not significantly improve model performance compared with histochemical and radiographic data alone [76]. Other studies have investigated the use of omics to predict survival in renal cell carcinoma [77]. A multimodal K-nearest neighbors (KNN) and support vector machine (SVM) model were able to use gene, isoform, exon, and junction expression data from RCC sample RNA-sequences to predict 5-year survival with moderate accuracy (AUC = 0.644) [78]. Improvements in predictive capabilities were shown by Tang et al. who created a multivariate regression model that used gene methylation data from RNA-seq to predict 1-, 3-, and 5-year survival rates with AUC of 0.794, 0.752, and 0.731, respectively [79].
For localized RCC, first-line treatment involves partial or radical nephrectomy, which is associated with 5-year survival rates between 97% and 78% depending on stage for localized RCC [80]. Despite high 5-year survival rates, identifying patients at risk of disease progression is vital for postoperative management of disease, including the decision to pursue adjuvant immunotherapy. Zhao et al. developed a hierarchical clustering model that identified five genomic subgroups of RCC. Using gene expression profiles of RCC samples taken after nephrectomy, they used a scoring system to determine high-, medium-, and low-risk patients. While low- and medium-risk patients according to scoring did not have significantly different survival after 12 years, high-risk patients had significantly decreased survival with a hazard ratio of 2.62 [81]. Other attempts to identify high-risk patients for RCC progression post-nephrectomy have focused on different ML modalities. For instance, Chen et al. developed a multimodal pathomics and radiomics model that was able to differentiate between high- and low-risk patients for decreased disease-free survival with hazard ratio of 13.4 [82].
Unfortunately, RCC may present as an aggressive cancer, with an estimated 17–30% of cases showing synchronous distant metastasis at the time of diagnosis [83, 84]. Several studies have used AI to predict the presence of distal metastases. Bai et al. created a model combining clinical risk factors and MRI radiomics data that was able to predict synchronous distal metastases with AUC 0.731 and 0.677 in the internal and external validation groups, respectively [85]. Wen et al. created a similar model to predict distal metastases using CT imaging data, with their model achieving an AUC of 0.821 in their validation cohort [86]. In the treatment of metastatic RCC, immunotherapy may help some patients by decreasing tumor burden and prolonging survival. However, overall response rates may be as low as 25% for common immunotherapy agents for RCC [87]. Thus, some authors have investigated the use of AI in helping predict treatment response to immunotherapy agents in patients with RCC. Khene et al. created a radiomics logistic regression model that was able to predict partial response or stable disease on nivolumab with AUC 0.91 [88]. Similar studies have been conducted for other immunotherapy medications. Negreros-Osuna et al. developed a logistic regression model combining patient demographic and clinical data with radiomics variables to create a model that was able to predict response to tyrosine kinase inhibitors with AUC of 0.94 [89]. Although the sample size in this study was small limiting the availability of a testing group for the model, these parameters may prove useful in predicting response to tyrosine kinase inhibitors in the future.
Though these do not represent an exhaustive list of the uses of AI in prediction of disease progression and outcomes in urological disease, AI has the potential to change how clinicians manage these conditions in the future.
Advertisement
5. Operative
Adoption of AI in the intraoperative setting is a natural progression for urology, as this field was an early embracer of robotics platforms and other technologies. Thus far, the application of AI involves adjunctive imaging and decision-making algorithms, along with augmented reality modeling. Ventures into autonomous robotics remain in their infancy and are, at the time of this writing, experimental [90].
Several applications are in the development for the management of stone disease. AI algorithms have been created to assist with access for percutaneous nephrolithotomy (PCNL). Percutaneous access to the renal collecting system can represent a hurdle to PCNL as there is a significant learning curve to the mastery of this technique [91]. Created by Taguchi et al., the automated needle targeting with X-ray device (ANT-X) facilitates accurate needle trajectory for the bullseye percutaneous access technique [91]. The puncture success rate (1.82 vs. 2.51 punctures, p = 0.025) and median puncture time (5.5 vs. 8 minutes, p = 0.049) were significantly improved with this technology in comparison with unassisted methodologies [91]. Neural networks are being used to predict the ability of a semirigid ureteroscope to fit within a patient’s ureter [91]. This information can help surgeons determine whether active dilatation of the ureter in a single stage procedure or passive dilatation with a stent in a staged approach is more appropriate.
Imaging algorithms and augmented reality (AR) show promise for improved determination of surgical margins in the intraoperative setting. Multiple modalities are being applied to robotic prostatectomy as well as to partial nephrectomy. CNNs can predict bleeding in the surgical field using images taken from the endoscope with a 98% true positive rate [92]. The images are transmitted in 3-second intervals and a confidence interval is then reported back to the surgeon’s display, with any interval below 100 being indicative of potential bleeding [93]. Porpiglia and Checcucci have created MRI-generated prostate models that display tumor properties like extracapsular extension and fused 3D models within the visible surgical field of the da Vinci robot [94]. Using this technology, the rate of extracapsular extension identification improved from 47% to 100% in robotic prostatectomies [92]. There are also experiments involving three-dimensional augmented reality (3D-AR) in robotic-assisted partial nephrectomies, with a trial led by Amparore et al. yielding successful overlaying of models using computer vision technology or CNN [92, 95]. Near-infrared fluorescence was also used to increase the visibility of the kidney. The challenge to applying 3D-AR to partial nephrectomies in particular is the constant need to manually adjust the model with patient anatomy [92]. Hailfler’s group was able to use ML analysis of Raman spectroscopy to differentiate renal cell cancer from benign renal tissue in a laboratory study [92]. The accuracy was 92.5%, with a 95.8% sensitivity, and an 88.8% specificity. If successfully translated to the operating room, the technology could potentially facilitate real-time intra-operative determination of surgical margins [92].
The desire to decrease variation in surgical technique and outcomes, as well as advances in AI, has led to robotic surgical platform development and consideration of autonomous robotic surgery. While there are several robotic systems, few of them possess meaningful autonomy, and all require human control or oversight. Autonomy has been defined in terms of levels, ranging from nil to full [96, 97]. The highest level that has been achieved is Level 3, conditional autonomy (Table 1). Within the realm of urology, 11 surgical robots have FDA approval for use. Of these, there is one Level 2 and a single Level 3 machine. One platform has been marketed for its ML capabilities; however, these are not included in its FDA clearance [96]. One obstacle to full autonomy is the challenge of programming multi-step complex procedures during which the surgical field is continuously changing as the surgery progresses. Automation of surgical assistant tasks and eventually soft tissue surgery represent the most feasible steps forward in the path to full autonomy. Furthermore, continued autonomous surgical trials are necessary to determine safety and efficacy versus current methods. These efforts raise interesting medicolegal questions.
| Autonomy level | Definition* | System | Use |
|---|---|---|---|
| Level 0: no autonomy | No autonomy. Devices with no robotic equipment. | ||
| Level 1: robot assistance | The robot provides mechanical guidance/assistance during a task but the human operator has continuous control of the robot. | Da Vinci X, Xi Da Vinci SP (single port) Soloassist II Artemis (ML) AmaKris SR1-A Monarch Platform Hand X Maestro Platform Canady Flex RoboWrist | Urologic surgical procedures Urologic surgical procedures Laparoscopic surgery/MIS Prostate biopsy, interventional prostate procedures TPPB, prostate brachytherapy Ureteroscopy, ULL Laparoscopic surgery/MIS Laparoscopic surgery/MIS Laparoscopic surgery/MIS Laparoscopic surgery/MIS |
| Level 2: task autonomy | Task specific robot directed by a human operator. The surgeon gives parameters for the task and the robot then performs the task. | PROBOT1 EUCLIDIAN STAR1 AquaBeam | TURP Prostate brachytherapy Soft tissue surgery Prostate resection |
| Level 3: conditional autonomy | Robot generates task strategies but the human operator selects from various strategies and/or approves the strategy suggested by the robot. The robot performs tasks without close oversight. | MrBot1 iSR’obot | Prostate brachytherapy Prostate biopsy |
| Level 4: high autonomy | Robot creates surgical plan and executes the plan once approved by the surgeon. | N/A | N/A |
| Level 5: full autonomy | Robotic system makes decisions about and performs the whole operation independently. | N/A | N/A |
Table 1.
Definitions adapted from levels of autonomy in surgical robotics [96]. System acronyms and companies provided in the Appendix.
Not FDA cleared or approved.
MIS = minimally invasive surgery, ML = machine learning capabilities, TPPB = trans-perineal prostate biopsy, ULL = ureteroscopy with laser lithotripsy, TURP = trans-urethral resection of the prostate, N/A = not applicable.
Other than the technical and programming aspects of creating robots with Level 4 or 5 autonomy, the FDA clearance process is challenging. To obtain approval of a new device, either a 510(k) must be granted or a De Novo approval given. A 510(k) approval is based on predicates; however, machines with such high levels of autonomy would not have any predecessors or comparable devices to achieve this [96, 97]. FDA classification is risk-based, with Class III being the highest risk and possessing the strictest approval guidelines [98]. Current Level 2 and 3 systems use surgeon oversight and review to reduce risk, but a Level 4 or 5 robot would not have such oversight, making risk-reduction difficult [96]. Given this, it has been suggested that Level 4 and 5 robots would be categorized by the FDA as Class III devices, which necessitates premarket approval (PMA) [96]. Class III approval requires substantially more time and monetary investment, which could have the unintended consequence of excluding smaller robotics companies as well as creating a bottleneck in advancements [96]. Fees include annual registration, PMA, and annual reporting fees, which can altogether total $568,990 [99]. There are discounts available for small businesses, which would lower the costs to $149,208 [99]. Lastly, the level of decision making that would be afforded to a Level 4 or 5 robot raises the idea that the device itself is practicing medicine. Should that hold true, medical societies would need to be heavily involved in regulating these robots [96]. Indeed, a significant amount of collaboration between medical societies, regulatory bodies, and engineers is necessary to appropriately determine the proper path forward.
Advertisement
6. Ethics and pitfalls
The use of AI in urology, and in medicine, is not without potential sticking points. This is not to say that AI has no place in the clinic or the operating room. AI can be properly adopted and accepted by patients, physicians, and the healthcare system when concerns regarding ethical standards, energy infrastructure, and medical guidelines are addressed.
For the patient, the prospect of a healthcare provider relying on AI is unsettling. According to a 2022 survey of 11,004 U.S. adults, 60% of respondents were uncomfortable with the idea and 33% believed it would lead to worse outcomes for patients [100]. When explored further in a survey of 600 Floridians, the anxiety appeared to stem from fear of losing the “human touch” in medicine. The public was rather confident in the ability of AI to efficiently complete clerical tasks like scheduling appointments (84.2%). However, tasks of extensive interaction, such as administering prescribed medications (33.7%) or assisting with surgical procedures (46.2%), produced greater discomfort [101]. A thorough review of patient perspectives on 37 different articles concluded that patients were not fully considered in the utilization and development of AI in healthcare. Of the articles reviewed, negative patient attitudes toward AI were derived from the “lack of human supervision in care” and “potential risk of job loss” [102].
For the physician, AI may represent technological breakthrough or their own professional obsolescence. Large Language Models (LLMs), such as the now familiar ChatGPT, respond to medical prompts with United States Medical Licensing Examination (USMLE) accurate results that are seemingly indistinguishable from that of their human counterparts [103, 104]. Newer LLMs specifically trained on biomedical data have demonstrated human parity in text generation and data mining [105]. On one hand, this opens doors for greater overall healthcare coverage, such that anyone with a Wi-Fi connection may find answers to pressing medical questions and concerns. On the other hand, the value of primary care and consultative services may be devalued.
In truth, LLMs do not comprehend the language that they themselves speak. Rather, these programs regurgitate digestible responses from a massive bank of prepared information [103]. The term “stochastic parrot” has been used to describe LLMs for their ability to echo pre-existing literature [106]. By simply parroting the information that is fed to them, they are at risk of spreading imbedded misinformation. In at least one reported instance, ChatGPT was successfully coaxed into assembling a fraudulent, yet seemingly genuine, scientific paper [107]. The current situation of AI is not dissimilar to the early years of WebMD, which led to medical personnel distrust and an increase in self-diagnosis (and self-misdiagnosis) [108]. While ChatGPT has shown high diagnostic accuracy, it was found to have a significantly higher unsafe triage rate when compared to other programs like WebMD. At least 2 of the 3 emergency medicine physicians surveyed advised a more urgent level of care (home care vs. PCP follow-up vs. emergency care) than ChatGPT in 41% of cases, compared to 19% for WebMD [109]. To realize the full potential that LLMs offer the world of medicine, healthcare workers, LLM engineers, and governmental bodies must remain vigilant to minimize misinformation [110, 111].
The header of the American Urology Association (AUA) webpage reads:
This statement, and the blocking of ChatGPT and Google’s Gemini
The development and deployment of LLMs throughout healthcare has significant monetary and energy costs. The training alone for ChatGPT-3 emitted 552 tons of CO2 and consumed 1287 megawatt-hours, or approximately the energy use of an average American home over 120 years [117, 118]. The overall ecologic impact and carbon footprint of several LLMs were evaluated using Software Carbon Intensity (SCI), a metric released by the Green Software Foundation to calculate the rate of carbon emissions for a particular software system. During this assessment, the total amount of energy required by a single LLM (GPT-J 6B) over its server’s lifespan was estimated to be 1285.47 kilowatt-hours. Each instance of the GPT-J 6B model was then calculated to emit 8.2 g carbon dioxide equivalents (CO2e) [119]. Compare that to a single Google search, which in 2009 was said to be equivalent to 0.2 g CO2e [120]. It should be noted that energy- and resource-intensive LLMs are not necessarily better tools depending on the task. With each iteration of GPT, the opportunity to develop more energy efficient models arises.
The financial strain of implementing AI in medicine is much more nuanced and difficult to assess. When considering the infrastructure, training, and maintenance of AI, fiscal estimates range between tens of thousands to half a million dollars per institution [121]. This being true, potential savings and increased revenue may be orders of magnitude greater. Some estimates place widespread implementation at $200–$360 billion dollars saved annually for the United States [122]. As we move closer toward universal AI adoption, focus on cost versus value continues to evolve. As essentially living entities, algorithms may decay or change over time as new situations arise (e.g., the COVID-19 pandemic). To maintain LLMs and AI in healthcare, trained personnel must monitor for such perturbations and seek to account for them. The need for highly specialized teams that are technologically equipped may prove costly. In an early 2025 interview, a chief data scientist at Stanford Health Care debated the viability of implementing AI in the event that it results in higher overall costs for patients [123].
Advertisement
7. Conclusions
Since the inception of the concept of AI, there have been and will continue to be substantial strides made, which are applicable to urology and medicine. The strictest definitions of AI are yet to be realized within medicine, but there is progression and effort toward this. As with any tool or technology, safety and ethical concerns will arise. Vigilance is needed to explore and answer the concerns and questions outlined above. While uncertainty remains regarding how AI will shape the practice of urology and the medical field as a whole, this is an exciting time to be alive.
Advertisement
Appendices and nomenclature
See Table A1.
| Urologic disease | Applications of AI |
|---|---|
| Infertility | |
| Urolithiasis | |
| Bladder cancer | |
| Prostate cancer | |
| Renal cancer |
Table A1.
Summary of potential applications of AI in various urologic diseases.
Da Vinci: Intuitive, Sunnyvale, California, USA.
Soloassist II: AKTORmed, Neutraubling, Germany.
Artemis: Eigen Health, Grass Valley, California USA.
amaKris: Augment Intelligent Medical System (China), Nanjing, China.
Monarch: Ethicon (Johnson & Johnson), New Brunswick, New Jersey, USA.
Hand X: Human Xtensions, Netanya, Israel.
Maestro: Moon Surgical, Paris, France.
Canady Flex: US Medical Innovations, Takoma Park, Maryland, USA.
Robowrist: US Medical Innovations, Takoma Park, Maryland, USA.
PROBOT: Prostatectomy ROBOT.
EUCLIDIAN: Endo-Uro Computer Lattice for Intratumoral Delivery, Implantation, and Ablation with Nanosensing.
STAR: Smart Tissue Autonomous Robot.
AquaBeam: Procept Biorobotics, San Jose, California, USA.
MrBot: Philips Medical Systems, Cleveland, Ohio, USA.
iSR’obot: Biobot Surgical, Singapore.
EUCLIDIAN and STAR are not mass produced by a major company.
References
- 1.
Turing AM. I.—Computing machinery and intelligence. Mind. 1950; LIX (236):433-460 - 2.
Kaul V, Enslin S, Gross SA. History of artificial intelligence in medicine. Gastrointestinal Endoscopy. 2020; 92 (4):807-812 - 3.
Chen J et al. Current status of artificial intelligence applications in urology and their potential to influence clinical practice. BJU International. 2019; 124 (4):567-577 - 4.
The Editors of Encyclopaedia Britannica. Cybernetics. In: Encyclopedia Britannica. Chicago, IL, USA: Encyclopædia Britannica, Inc.; 2025 - 5.
Edwards J. In: Smartbridge, editor. AI vs IA: What’s the Real Difference? Houston, TX, USA: Smartbridge; 2024 - 6.
Abdullah U, Ligęza A, Zafar K. Performance evaluation of rule-based expert systems: An example from medical billing domain. Expert Systems. 2017; 34 (6):e12218 - 7.
Zhu C et al. Current applications of artificial intelligence in billing practices and clinical plastic surgery. Plastic and Reconstructive Surgery. Global Open. 2024; 12 (7):e5939 - 8.
3M Health Information Systems. 3M™ Audit Expert System. St. Paul, MN, USA: 3M; 2017 - 9.
Wu K et al. Characterizing the clinical adoption of medical AI devices through U.S. insurance claims. NEJM AI. 2024; 1 (1):AIoa2300030 - 10.
Groves P et al. The “Big Data” Revolution in Healthcare: Accelerating Value and Innovation. Chicago, IL, USA: Center for US Health System Reform Business Technology Office; 2013 - 11.
Talyshinskii A et al. Potential of AI-driven Chatbots in urology: Revolutionizing patient care through artificial intelligence. Current Urology Reports. 2024; 25 (1):9-18 - 12.
Calvo RA et al. Assessing the feasibility of a text-based conversational agent for asthma support: Protocol for a mixed methods observational study. JMIR Research Protocols. 2023; 12 :e42965 - 13.
Rabaan AA et al. Artificial intelligence for clinical diagnosis and treatment of prostate Cancer. Cancers. 2022; 14 (22):5595 - 14.
Froń A et al. Artificial intelligence in urooncology: What we have and what we expect. Cancers. 2023; 15 (17):4282 - 15.
Nagpal K et al. Development and validation of a deep learning algorithm for improving Gleason scoring of prostate cancer. npj Digital Medicine. 2019; 2 (1):48 - 16.
Rossin G et al. Artificial intelligence in bladder cancer diagnosis: Current applications and future perspectives. BioMedInformatics. 2023; 3 (1):104-114 - 17.
Vander Borght M, Wyns C. Fertility and infertility: Definition and epidemiology. Clinical Biochemistry. 2018; 62 :2-10 - 18.
Agarwal A et al. Male infertility. The Lancet. 2021; 397 (10271):319-333 - 19.
Penzias A et al. Fertility evaluation of infertile women: A committee opinion. Fertility and Sterility. 2021; 116 (5):1255-1265 - 20.
Llavanera M et al. A systematic review identifying fertility biomarkers in semen: A clinical approach through omics to diagnose male infertility. Fertility and Sterility. 2022; 118 (2):291-313 - 21.
Guzick DS et al. Sperm morphology, motility, and concentration in fertile and infertile men. New England Journal of Medicine. 2001; 345 (19):1388-1393 - 22.
Nallella KP et al. Significance of sperm characteristics in the evaluation of male infertility. Fertility and Sterility. 2006; 85 (3):629-634 - 23.
Swan SH. Semen quality in fertile US men in relation to geographical area and pesticide exposure. International Journal of Andrology. 2006; 29 (1):62-68 - 24.
Haugen TB, Egeland T, Magnus Ø. Semen parameters in Norwegian fertile men. Journal of Andrology. 2006; 27 (1):66-71 - 25.
Lewis SE. Is sperm evaluation useful in predicting human fertility? Reproduction. 2007; 134 (1):31-40 - 26.
Esteves SC. Clinical management of infertile men with nonobstructive azoospermia. Asian Journal of Andrology. 2015; 17 (3):459-470 - 27.
Dai X, Shen L. Advances and trends in omics technology development. Frontiers in Medicine. 2022; 9 - 28.
Sergerie M et al. Sperm DNA fragmentation: Threshold value in male fertility. Human Reproduction. 2005; 20 (12):3446-3451 - 29.
Ribas-Maynou J et al. Comprehensive analysis of sperm DNA fragmentation by five different assays: TUNEL assay, SCSA, SCD test and alkaline and neutral Comet assay. Andrology. 2013; 1 (5):715-722 - 30.
Simon L et al. Clinical significance of sperm DNA damage in assisted reproduction outcome. Human Reproduction. 2010; 25 (7):1594-1608 - 31.
Chan PJ et al. A simple sperm DNA toroid integrity test and risk of miscarriage. BioMed Research International. 2015; 2015 (1):780983 - 32.
Tang D et al. Identification of deleterious variants in patients with male infertility due to idiopathic non-obstructive azoospermia. Reproductive Biology and Endocrinology. 2022; 20 (1):63 - 33.
Tang Q et al. Identifying potential biomarkers for non-obstructive azoospermia using WGCNA and machine learning algorithms. Frontiers in Endocrinology. 2023; 14 :1108616 - 34.
Jia X, Yin Z, Peng Y. Gene differential co-expression analysis of male infertility patients based on statistical and machine learning methods. Frontiers in Microbiology. 2023; 14 :1092143 - 35.
Kaltsas A et al. Predictors of successful testicular sperm extraction: A new era for men with non-obstructive azoospermia. Biomedicine. 2024; 12 (12):2679 - 36.
Pavan-Jukic D et al. Predictive factors for sperm retrieval from males with azoospermia who are eligible for testicular sperm extraction (TESE). Systems Biology in Reproductive Medicine. 2020; 66 (1):70-75 - 37.
Vahidi S et al. Success rate and ART outcome of microsurgical sperm extraction in non obstructive azoospermia: A retrospective study. International Journal of Reproductive BioMedicine. 2021; 19 (9):781 - 38.
Bachelot G et al. A machine learning approach for the prediction of testicular sperm extraction in nonobstructive azoospermia: Algorithm development and validation study. Journal of Medical Internet Research. 2023; 25 :e44047 - 39.
Jarvi KA et al. Semen biomarker TEX101 predicts sperm retrieval success for men with testicular failure. F1000Research. 2021; 10 :569 - 40.
Hashemi MS et al. Could analysis of testis-specific genes, as biomarkers in seminal plasma, predict presence of focal spermatogenesis in non-obstructive azoospermia? Andrologia. 2020; 52 (2):e13483 - 41.
Zhang Y et al. Circulating microRNAs in seminal plasma as predictors of sperm retrieval in microdissection testicular sperm extraction. Annals of Translational Medicine. 2022; 10 (7):380-392 - 42.
Goss DM et al. Evaluation of an artificial intelligence-facilitated sperm detection tool in azoospermic samples for use in ICSI. Reproductive Biomedicine Online. 2024; 49 (1):103910 - 43.
Sánchez C et al. Artificial intelligence in urology: Application of a machine learning model to predict the risk of urolithiasis in a general population. Journal of Endourology. 2024; 38 (8):712-718 - 44.
Solakhan M, Seckiner SU, Seckiner I. A neural network-based algorithm for predicting the spontaneous passage of ureteral stones. Urolithiasis. 2020; 48 (6):527-532 - 45.
Aminsharifi A et al. Predicting the postoperative outcome of percutaneous nephrolithotomy with machine learning system: Software validation and comparative analysis with guy's stone score and the CROES nomogram. Journal of Endourology. 2020; 34 (6):692-699 - 46.
Alves BM et al. Use of artificial intelligence for sepsis risk prediction after flexible ureteroscopy: A systematic review. Revista do Colégio Brasileiro de Cirurgiões. 2023; 50 :e20233561 - 47.
Xun Y et al. A preoperative nomogram for sepsis in percutaneous nephrolithotomy treating solitary, unilateral and proximal ureteral stones. PeerJ. 2020; 8 :e9435 - 48.
Pietropaolo A et al. A machine learning predictive model for post-ureteroscopy urosepsis needing intensive care unit admission: A case–control YAU endourology study from nine European centres. Journal of Clinical Medicine. 2021; 10 (17):3888 - 49.
Kamat AM et al. Bladder cancer. The Lancet. 2016; 388 (10061):2796-2810 - 50.
Bhambhvani HP et al. Development of robust artificial neural networks for prediction of 5-year survival in bladder cancer. In: Urologic Oncology: Seminars and Original Investigations. Amsterdam, NL: Elsevier; 2021 - 51.
Gavriel CG et al. Assessment of immunological features in muscle-invasive bladder cancer prognosis using ensemble learning. Cancers. 2021; 13 (7):1624 - 52.
Hasnain Z et al. Machine learning models for predicting post-cystectomy recurrence and survival in bladder cancer patients. PLoS ONE. 2019; 14 (2):e0210976 - 53.
Xu X et al. A predictive nomogram for individualized recurrence stratification of bladder cancer using multiparametric MRI and clinical risk factors. Journal of Magnetic Resonance Imaging. 2019; 50 (6):1893-1904 - 54.
Wang G-Y et al. Prediction of non-muscle invasive bladder cancer recurrence using deep learning of pathology image. Scientific Reports. 2024; 14 (1):18931 - 55.
Zhao CC et al. Machine learning decision support model for radical cystectomy discharge planning. In: Urologic Oncology: Seminars and Original Investigations. Amsterdam, NL: Elsevier; 2022 - 56.
Taylor J et al. Different models for prediction of radical cystectomy postoperative complications and care pathways. Therapeutic Advances in Urology. 2019; 11 :1756287219875587 - 57.
Wessels F et al. Comorbidity scores and machine learning methods can improve risk assessment in radical cystectomy for bladder cancer. Bladder Cancer. 2022; 8 (2):155-163 - 58.
Cha KH et al. Bladder cancer treatment response assessment in CT using radiomics with deep-learning. Scientific Reports. 2017; 7 (1):8738 - 59.
Hepburn AC et al. Identification of CNGB1 as a predictor of response to neoadjuvant chemotherapy in muscle-invasive bladder cancer. Cancers. 2021; 13 (15):3903 - 60.
Chen J et al. Translational bioinformatics for diagnostic and prognostic prediction of prostate cancer in the next-generation sequencing era. BioMed Research International. 2013; 2013 (1):901578 - 61.
Hägglöf C et al. TMPRSS2-ERG expression predicts prostate cancer survival and associates with stromal biomarkers. PLoS ONE. 2014; 9 (2):e86824 - 62.
Nam R et al. Expression of the TMPRSS2: ERG fusion gene predicts cancer recurrence after surgery for localised prostate cancer. British Journal of Cancer. 2007; 97 (12):1690-1695 - 63.
Zhu L et al. Harnessing artificial intelligence for prostate cancer management. Cell Reports Medicine. 2024:1-17 - 64.
Dadhania V et al. Leveraging artificial intelligence to predict ERG gene fusion status in prostate cancer. BMC Cancer. 2022; 22 (1):494 - 65.
Erak E et al. Predicting prostate cancer molecular subtype with deep learning on histopathologic images. Modern Pathology. 2023; 36 (10):100247 - 66.
Castellani D et al. Advancements in artificial intelligence for robotic-assisted radical prostatectomy in men suffering from prostate cancer: Results from a scoping review. Chinese Clinical Oncology. 2024; 13 (4):54-54 - 67.
Ma R et al. Surgical gestures as a method to quantify surgical performance and predict patient outcomes. npj Digital Medicine. 2022; 5 (1):187 - 68.
Hung AJ et al. A deep-learning model using automated performance metrics and clinical features to predict urinary continence recovery after robot-assisted radical prostatectomy. BJU International. 2019; 124 (3):487-495 - 69.
Jones CU et al. Adding short-term androgen deprivation therapy to radiation therapy in men with localized prostate cancer: Long-term update of the NRG/RTOG 9408 randomized clinical trial. International Journal of Radiation Oncology, Biology, Physics. 2022; 112 (2):294-303 - 70.
Kishan AU et al. Androgen deprivation therapy use and duration with definitive radiotherapy for localised prostate cancer: An individual patient data meta-analysis. The Lancet Oncology. 2022; 23 (2):304-316 - 71.
Nguyen PL et al. Adverse effects of androgen deprivation therapy and strategies to mitigate them. European Urology. 2015; 67 (5):825-836 - 72.
Spratt DE et al. Artificial intelligence predictive model for hormone therapy use in prostate cancer. NEJM Evidence. 2023; 2 (8):EVIDoa2300023 - 73.
Shore ND et al. Biochemical recurrence in patients with prostate cancer after primary definitive therapy: Treatment based on risk stratification. Prostate Cancer and Prostatic Diseases. 2024; 27 (2):192-201 - 74.
Huang W et al. A novel artificial intelligence–powered method for prediction of early recurrence of prostate cancer after prostatectomy and cancer drivers. JCO Clinical Cancer Informatics. 2022; 6 :e2100131 - 75.
Escudier B et al. Renal cell carcinoma: ESMO clinical practice guidelines for diagnosis, treatment and follow-up. Annals of Oncology. 2019; 30 (5):706-720 - 76.
Schulz S et al. Multimodal deep learning for prognosis prediction in renal cancer. Frontiers in Oncology. 2021; 11 :788740 - 77.
Giulietti M et al. The role of artificial intelligence in the diagnosis and prognosis of renal cell tumors. Diagnostics. 2021; 11 (2):206 - 78.
Schwartz M et al. Integration of multimodal RNA-seq data for prediction of kidney cancer survival. In: 2015 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). Washington DC, USA: IEEE; 2015 - 79.
Tang W, Cao Y, Ma X. Novel prognostic prediction model constructed through machine learning on the basis of methylation-driven genes in kidney renal clear cell carcinoma. Bioscience Reports. 2020; 40 (7):1-13 - 80.
Cheaib JG et al. Stage-specific conditional survival in renal cell carcinoma after nephrectomy. In: Urologic Oncology: Seminars and Original Investigations. Amsterdam, NL: Elsevier; 2020 - 81.
Zhao H et al. Gene expression profiling predicts survival in conventional renal cell carcinoma. PLoS Medicine. 2006; 3 (1):e13 - 82.
Chen S et al. Deep learning-based multi-model prediction for disease-free survival status of patients with clear cell renal cell carcinoma after surgery: A multicenter cohort study. International Journal of Surgery. 2024; 110 (5):2970-2977 - 83.
Najem EJ et al. Navigating advanced renal cell carcinoma in the era of artificial intelligence. Cancer Imaging. 2025; 25 :16 - 84.
Mikhail M et al. Role of metastasectomy in the management of renal cell carcinoma. Frontiers in Surgery. 2022; 9 :943604 - 85.
Bai X et al. MRI radiomics-based nomogram for individualised prediction of synchronous distant metastasis in patients with clear cell renal cell carcinoma. European Radiology. 2021; 31 :1029-1042 - 86.
Wen R et al. CT radiomics for the prediction of synchronous distant metastasis in clear cell renal cell carcinoma. Journal of Computer Assisted Tomography. 2021; 45 (5):696-703 - 87.
Motzer RJ et al. Nivolumab versus everolimus in advanced renal-cell carcinoma. New England Journal of Medicine. 2015; 373 (19):1803-1813 - 88.
Khene ZE et al. Radiomics can predict tumour response in patients treated with Nivolumab for a metastatic renal cell carcinoma: An artificial intelligence concept. World Journal of Urology. 2021; 39 :3707-3709 - 89.
Negreros-Osuna AA et al. Clinical-radiomic model in advanced kidney cancer predicts response to tyrosine kinase inhibitors. Oncology Letters. 2022; 24 (6):446 - 90.
Rivero-Moreno Y et al. Autonomous robotic surgery: Has the future arrived? Cureus. 2024; 16 (1):e52243 - 91.
Tano ZE et al. Surgical artificial intelligence: Endourology. Urologic Clinics of North America. 2024; 51 (1):77-89 - 92.
Bellos T et al. Artificial intelligence in urologic robotic oncologic surgery: A narrative review. Cancers. 2024; 16 (9):1775 - 93.
Checcucci E et al. Development of bleeding artificial intelligence detector (BLAIR) system for robotic radical prostatectomy. Journal of Clinical Medicine. 2023; 12 (23):7355 - 94.
Porpiglia F et al. Augmented-reality robot-assisted radical prostatectomy using hyper-accuracy three-dimensional reconstruction (HA3D™) technology: A radiological and pathological study. BJU International. 2019; 123 (5):834-845 - 95.
Amparore D et al. Computer vision and machine-learning techniques for automatic 3D virtual images overlapping during augmented reality guided robotic partial nephrectomy. Technology in Cancer Research & Treatment. 2024; 23 :15330338241229368 - 96.
Lee A et al. Levels of autonomy in FDA-cleared surgical robots: A systematic review. npj Digital Medicine. 2024; 7 (1):103 - 97.
Connor MJ et al. Autonomous surgery in the era of robotic urology: Friend or foe of the future surgeon? Nature Reviews Urology. 2020; 17 (11):643-649 - 98.
U.S. Food and Drug Administration. De Novo Classification Request. Premarket Submissions: Selecting and Preparing the Correct Submission. 2022. Available from: https://www.fda.gov/medical-devices/premarket-submissions-selecting-and-preparing-correct-submission/de-novo-classification-request - 99.
U.S. Food and Drug Administration. Medical Device User Fee Amendments (MDUFA). 2024/2025. Available from: https://www.fda.gov/industry/fda-user-fee-programs/medical-device-user-fee-amendments-mdufa - 100.
Tyson A et al. 60% of Americans Would be Uncomfortable with Provider Relying on AI in Their Own Health Care. Washington DC, USA: Pew Research Center; 2023. p. 46 - 101.
Witkowski K, Okhai R, Neely SR. Public perceptions of artificial intelligence in healthcare: Ethical concerns and opportunities for patient-centered care. BMC Medical Ethics. 2024; 25 (1):74 - 102.
Moy S et al. Patient perspectives on the use of artificial intelligence in health care: A scoping review. Journal of Patient-Centered Research and Reviews. 2024; 11 (1):51-62 - 103.
Harrer S. Attention is not all you need: The complicated case of ethically using large language models in healthcare and medicine. eBioMedicine. 2023; 90 :104512 - 104.
Singhal K et al. Large language models encode clinical knowledge. Nature. 2023; 620 (7972):172-180 - 105.
Luo R et al. BioGPT: Generative pre-trained transformer for biomedical text generation and mining. Briefings in Bioinformatics. 2022; 23 (6):1-11 - 106.
Bender EM et al. On the dangers of stochastic parrots: Can language models be too big? In: Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency. New York, NY, USA: Association for Computing Machinery; 2021 - 107.
Májovský M et al. Artificial intelligence can generate fraudulent but authentic-looking scientific medical articles: Pandora’s box has been opened. Journal of Medical Internet Research. 2023; 25 :e46924 - 108.
Kibbe DC. WebMD: The online future for physicians? Family Practice Management. 1999; 6 (7):48 - 109.
Fraser H et al. Comparison of diagnostic and triage accuracy of Ada health and WebMD symptom checkers, ChatGPT, and physicians for patients in an emergency department: Clinical data analysis study. JMIR mHealth and uHealth. 2023; 11 (1):e49995 - 110.
Chen C, Shu K. Can LLM-generated misinformation be detected? arXiv Preprint. arXiv:2309.13788. 2023 - 111.
Chen C, Shu K. Combating misinformation in the age of llms: Opportunities and challenges. AI Magazine. 2024; 45 (3):354-368 - 112.
American Urological Association. Restrictions on Use of AUA, AUAER, and UCF Content in Third Party Applications, Including Artificial Intelligence Technologies, Such as Large Language Models and Generative AI. Linthicum, MD, USA: American Urological Association; 2024. p. 1 - 113.
Omar M et al. Large language models in medicine: A review of current clinical trials across healthcare applications. PLOS Digital Health. 2024; 3 (11):e0000662 - 114.
Baker HP et al. ChatGPT's ability to assist with clinical documentation: A randomized controlled trial. JAAOS-Journal of the American Academy of Orthopaedic Surgeons. 2024; 32 (3):123-129 - 115.
Theodos K, Sittig S. Health information privacy laws in the digital age: HIPAA doesn’t apply. Perspectives in Health Information Management. 2020; 18 (Winter):1l - 116.
Edemekong PF, Annamaraju P, Haydel MJ. Health Insurance Portability and Accountability Act. Treasure Island, FL, USA: StatPearls Publishing; 2018 - 117.
Patterson D et al. Carbon emissions and large neural network training. arXiv Preprint. arXiv:2104.10350. 2021 - 118.
Mehta S. How much energy do LLMs consume? Unveiling the power behind AI. In: Deep Dives. The Association of Data Scientists; 2024 - 119.
Everman B et al. Evaluating the carbon impact of large language models at the inference stage. In: 2023 IEEE International Performance, Computing, and Communications Conference (IPCCC). Washington DC, USA: IEEE Computer Society; 2023 - 120.
Holzle U. Powering a Google search. In: Google Official Blog. Mountain View, CA, USA: Google; 2009 - 121.
Alkhaldi N. Assessing the cost of implementing AI in healthcare. In: Trending. Aliso Viejo, CA: itrex; 2024 - 122.
Alnasser B. A review of literature on the economic implications of implementing artificial intelligence in healthcare. E-Health Telecommunication Systems and Networks. 2023; 12 (3):35-48 - 123.
Tahir D. Health care AI, intended to save money, turns out to require a lot of expensive humans. In: KFF Health News. Washington, DC: KFF Health News; 2025