Bladder cancer is a disease characterized by the uncontrolled growth of abnormal cells in the lining of the bladder, a hollow organ in the lower abdomen responsible for storing urine. Non-muscle-invasive bladder cancer (NMIBC) is a subtype of bladder cancer that hasn’t penetrated the muscular wall of the bladder. This type of bladder cancer, often found early, comprises about 75% of all bladder cancer cases (21).

Bladder cancer staging follows the TNM system, detailing cancer progression within or beyond the bladder (see Figure 1 for TNM staging). Early stages (CIS, Ta, T1) are Non-Muscle-Invasive, with the tumor confined to the bladder’s surface or connective tissue yet not penetrating the deeper muscle layers. These stages can, however, progress to muscle-invasive (MIBC) stages. NMIBC carries a lower risk of metastasis compared to MIBC but has a high recurrence rate (21). NMIBC has a tendency to often recur. These recurrences can be either at the same stage as the initial tumor or at a more advanced stage (22). Treatment primarily includes transurethral resection of bladder tumor (TURBT) to remove cancerous cells (23). Depending on the risk of recurrence and progression, intravesical therapy, where medication is directly instilled into the bladder, may be applied post-TURBT. This can include chemotherapy agents, such as mitomycin C, or immunotherapy with Bacillus Calmette-Guerin.

Bladder cancer is associated with numerous risk factors, the most significant being tobacco smoking, which accounts for about half of all cases (24). Figure 2 highlights the overwhelming effect of smoking over other factors. Other factors include occupational exposures, chronic bladder inflammation, and genetic predisposition, which contribute to the complexity of predicting the disease’s occurrence and progression (25–27). Additionally, gender disparities exist, with men being about three to four times more likely to develop bladder cancer than women (28). Age, diet high in processed meat, and tumor characteristics like grade and size also play crucial roles in predicting the likelihood of recurrence in NMIBC (24, 26, 29). Other predictive factors include genetic alterations or mutations, the presence of bladder cancer markers in the urine, and findings from imaging studies and cystoscopy (30). This variability in risk factors highlights the challenges in developing personalized care plans, underscoring the need for sophisticated predictive tools.

Creating personalized care plans for bladder cancer is challenging due to its highly varied nature. Each case differs in terms of tumor biology, stage, grade, and genetic mutations, which means a treatment effective for one patient may not work for another. Despite advances in bladder cancer genetics, the lack of reliable predictive biomarkers underscores the need for more advanced tools. In the following sections, we explore how Artificial Intelligence and Machine Learning (ML)-based approaches, are poised to fill this gap by offering more precise and individualized predictions, ultimately improving patient outcomes and treatment strategies.

The European Organisation for Research and Treatment of Cancer (EORTC) and The Club Urológico Español de Tratamiento Oncológico (CUETO), are two popular clinical urology research organisations, that developed the tools for recurrence prediction used currently. EORTC developed the EORTC Risk Tables (31), and CUETO introduced the CUETO Scoring system. These tables use a variety of factors, such as the number and size of tumors, prior recurrence rates, T-stage, and grade to calculate risk scores (32, 33). Vedder et al. (34) conducted a study which revealed that both EORTC and CUETO’s risk scores were not accurate in predicting recurrence (found c-indices of 0.55-0.61, where 0.5 is random guess). Other studies have also concluded that CUETO and EORTC presented poor discriminative value in predicting clinical events. These models overestimated the risk, especially in highest-risk patients (35–37). This has significant implications for both physicians and patients alike.

Poor and unreliable recurrence prediction has far-reaching consequences. Unreliable predictions of recurrence can trigger unnecessary invasive procedures like repeated cystoscopies or biopsies, overburdening healthcare providers and depleting valuable resources (38). This, in turn, escalates the financial costs borne by the NHS (39, 40), diverting funds that could otherwise be allocated to patient care and research endeavors. For patients, the repercussions of unreliable recurrence prediction are substantial, entailing risks and complications. False-positive predictions cause unnecessary anxiety and subject patients to additional tests and treatments with inherent risks. Conversely, false-negative predictions can delay or miss disease progression, hindering timely intervention and adversely affecting patient health (41, 42).

Traditional mathematical and statistical tools use a limited set of variables. These tools assume straightforward, proportional relationships between patient variables like tumor size, number of tumors, prior recurrence, and T-stage with outcomes such as recurrence and progression. By assigning linear scores based on cancer grade (1, 2, or 3), they overlook the nuanced severity differences, such as the greater jump from Grade 2 to 3 compared to Grade 1 to 2. In reality, cancer prognosis is shaped by complex, non-linear interactions involving molecular markers, genetic factors, and evolving medical data—factors that static, linear models cannot adequately capture. These traditional methods also struggle with missing data and incorporating recent advancements. Emerging approaches aim to enhance tumor classification, discover novel biomarkers, and improve predictions for bladder cancer metastasis, while other promising areas include body composition analysis and biomarkers beyond traditional assessments (43–45). Additionally, novel prognostic models based on gene signatures show potential (46). However, these innovations cannot be integrated into traditional tools highlighting the urgent need for more sophisticated models that offer reliable, dynamic predictions for NMIBC recurrence.

Artificial Intelligences (AI) has emerged as a powerful tool in the medical field, with machine learning (ML) at its forefront, particularly in tasks related to cancer prognosis and recurrence prediction. AI, a broad category of computational methodologies designed to emulate human cognitive functions, has been increasingly deployed in the medical field. ML algorithms can be trained to learn from existing data, adjusting their mathematical parameters to predict outcomes with high accuracy. In the context of bladder cancer, algorithms like support vector machines (SVM), random forest (RF), artificial neural networks (ANN), and deep learning (DL) have been used to design models that enhance the prediction of cancer recurrence (47–52).

These models leverage a diverse range of markers — radiomic, clinical, pathological, and genomic — to build comprehensive and nuanced predictive models (53–57). They aim to improve risk stratification, anticipate recurrence, and optimize treatment planning, opening doors to more personalized patient management. Nonetheless, while advancements in AI and ML hold considerable promise, there is still much to learn, explore, and validate before these technologies can fully realize their potential in bladder cancer management.

Research in AI techniques to predict NMIBC recurrence is accelerating rapidly. As demonstrated in Figure 3 , the number of studies focusing on bladder cancer recurrence has shown a steady, linear increase over the past two decades (depicted by a line graph). However, the adoption of ML approaches has grown at a much faster, exponential rate (depicted by the bar graph). This surge in ML research highlights the recognition of its potential to improve predictive accuracy and model complex, non-linear relationships that traditional statistical methods often struggle to capture. With increasing accessibility and advancements in computational power, the relevance of ML approaches in bladder cancer research is expected to continue expanding.

An emerging trend in the development of ML models for NMIBC prediction is the integration of multiple feature types, such as clinical, genomic, and imaging data. Single-feature models often fail to account for the multifaceted nature of cancer progression, whereas combining multiple data sources allows for more comprehensive models and improved predictive performance. This shift can be attributed to advancements in data integration techniques, the growing availability of multi-modal datasets and ease of access of powerful GPUs, which enable researchers to leverage richer and more diverse data for more accurate and clinically relevant predictions.

Figure 4 depicts a typical bladder cancer prediction workflow using an ML model, consisting of data pre-processing and ML algorithm application. The process begins with secure data storage, followed by data pre-processing, where irrelevant or incomplete data is removed and medical scans are segmented to focus on key areas like tumor lesions. Redundant features are identified and eliminated, often using clustering techniques. Then, ML algorithms are trained with this processed data for cancer recurrence prediction. The model’s performance, evaluated by accuracy, precision, recall, and F1-score, may lead to retraining for improved standards. While accuracy measures the overall correctness of the model, precision indicates the proportion of positive cases correctly identified, avoiding false positives. Similarly, recall reflects how many actual positives were correctly identified, avoiding false negatives. The F1-score combines precision and recall into a single metric, balancing both false positives and false negatives. Finally, the model’s effectiveness is compared with existing methods to assess any advancements in bladder cancer recurrence prediction and clinical detection benefits.

Accurately assessing the risk and predicting the occurrence and recurrence of NMIBC early is crucial for effective treatment and management of NMIBC. ML-based diagnosis and predictive systems have been proven valuable in tumor detection, bladder segmentation, and NMIBC identification (48, 58). Moreover, AI approaches also have been applied to tumor staging, grading, survival rate prediction, response to chemotherapy, and recurrence rates, all of which are essential in personalized NMIBC management (59–61).

However, as of this date, no reliable AI algorithm is capable of accurately predicting NMIBC recurrence and improving management through a combination of the aforementioned markers. By reviewing these studies, we aim to pave the way for researchers to develop highly precise AI-based NMIBC recurrence prediction systems, facilitating optimal personalized management for early-stage NMIBC.

In this systematic review, we survey studies published up to October 2024, focusing on highly-ranked articles related to the implementation of ML and AI in bladder cancer prediction. We searched databases such as PubMed, IEEE Xplore, ScienceDirect (Elsevier), Springer, Nature, and MDPI. We used a combination of keywords that matched the scope of the survey such as “Bladder Cancer OR Non-Muscle Invasive Bladder Cancer OR NMIBC” AND “Artificial Intelligence” OR “Machine Learning” OR “Neural Networks” AND “Prediction” OR “Predictor” AND “Recurrence”. Our search yielded a total of 175 unique studies. We reviewed the 175 articles on ML for NMIBC, focusing on those used for recurrence prediction. Upon abstract evaluation, 98 articles were discarded according to the exclusion criteria defined below. A further 46 studies were excluded based on full-text analysis. A total of 25 studies were finally selected for in-depth analysis within this review. Our study adhered to the guidelines set forth by the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines (62). The PRISMA flowchart is depicted in Figure 5 .

The inclusion criteria for this systematic review were defined to ensure the relevance and quality of the studies selected for analysis. The criteria are as follows: (1) original research articles, including peer-reviewed journal papers and conference proceedings; (2) studies specifically focused on the application of ML techniques for the prediction of NMIBC recurrence; (3) research exclusively addressing NMIBC recurrence or survival outcomes; (4) publications available in the English language. (5) Articles that report specific performance metrics (e.g., accuracy, sensitivity, specificity, AUC) to evaluate the effectiveness of the proposed ML models.

The exclusion criteria for this systematic review were designed to eliminate studies outside the scope of NMIBC recurrence prediction. The following were excluded: (1) studies focused solely on diagnosis without addressing recurrence or survival prediction; (2) non-original research, including systematic reviews, meta-analyses, commentaries, editorials, and case reports; (3) articles lacking full-text access or sufficient methodological details; (4) studies using only traditional statistical methods without machine learning; (5) research on cancer types other than NMIBC (such as Gall Bladder Cancer, prostate cancer, lung cancer) or studies that did not differentiate between muscle-invasive and non-muscle-invasive bladder cancer.

Imaging-based ML models have become central to predicting NMIBC recurrence, leveraging data from pathology slides, radiomics, and CT scans. These approaches provide detailed insights into tumor morphology and cellular patterns, often revealing features that traditional methods may overlook. Studies in this category use deep learning and other ML techniques to enhance prognostic accuracy by analyzing both tissue samples and imaging modalities.

The analysis of pathology slides has been a cornerstone of many ML models for NMIBC recurrence prediction, offering insights into tissue morphology and cellular characteristics. Chen et al. (63) used machine learning on Hematoxylin and Eosin-stained images from 514 patients, a substantial sample size for bladder cancer research. Unlike other approaches that use nuclear extraction techniques, they applied LASSO with 10-fold cross-validation to identify 22 bladder cancer-related and 18 survival-related image features. Their diagnostic model achieved strong performance, with Area Under the Receiver Operating Curve (AUROC) values of 94.1%, effectively distinguishing bladder cancer from normal tissues and glandular cystitis. The ML-based risk score served as an independent predictor for survival, enhancing prediction accuracy for 1-, 3-, and 5-year overall survival by more than 10%. However, the use of non-uniform median cut-off values for high-risk scores limits consistency, indicating a need for further standardization and validation in future prospective trials before broad clinical adoption.

Building on this approach of using pathology slides, Tokuyama et al. (49) explored the predictive value of nuclear atypia in NMIBC, focusing on the morphological characteristics of cancer cells derived from transurethral resection specimen. They utilized ML to predict the recurrence of NMIBC based on nuclear atypia (abnormalities in the nuclei of cancer cells) extracted from transurethral resection specimens. Using SVM and RF algorithms on a dataset of 125 patients, the authors derived quantitative morphological features from regions of interest on Hematoxylin and Eosin-stained slides. This involved the application of a nuclear extraction process using software programs “Ilastik” and “YOLO v3” for the segmentation of individual nuclei. The SVM-based model achieved a 90% probability of predicting NMIBC recurrence within 2 years post-TURBT, while the RF-based model achieved 86.7%. Despite these promising results, the study’s limitations included a relatively small sample size, the potential for bias in the nuclear extraction process, and the limited generalizability due to the exclusion criteria.

Drachneris et al. (64) provided a new perspective by analyzing immune cell density gradients. They used CD8+ cell-density gradients to predict Recurrence-free survival in NMIBC patients post-BCG therapy. CD8+ are a type of white blood cell that is crucial for fighting infections or cancer. Traditional methods merely count the total number of immune cells, while this study analyses the spatial distribution (or the cell density gradients), which provides deeper insight into the cancer’s immune response. By combining the HALO AI Densenet v2 classifier (a deep learning-based image analysis platform designed for digital pathology) and multivariable Cox regression models, the study creates a robust model with a C-index of 0.74. However, the paper recognizes that broader, prospective validations and explorations of other immune factors are required to further validate the methods. This study shows that a fresh perspective on established methods can enhance bladder cancer prognosis accuracy.

The use of cellular features extends to urine cytology as well. The study by Levy et al. (65) used a Deep learning ML tool called AutoParis-X to predict bladder cancer recurrence risk from 1,259 urine cytology images. The prediction model achieved good accuracy (C-index 0.77), outperforming models using standard cytological assessment alone. It was found that the model worked best when it looked at samples from the first 6 months after the initial cancer diagnosis. This model was trained on retrospective data from 159 patients from only one hospital. The small number of patients, single-centered and retrospective analysis results in concerns about the model’s generalizability.

Pathomics, which integrates patch-level and whole-slide image analyses, represents another advancement in pathology-based models. Wang et al. (66) introduced a novel pathomics model using deep learning for predicting NMIBC recurrence with high accuracy. This study aimed to address the challenge of early recurrence prediction in NMIBC by leveraging deep learning to analyze pathology images. The model, developed using a two-phase approach—patch-level prediction followed by whole slide image (WSI)-level prediction—achieved a strong performance with an AUC of 0.860 in the test cohort. Transfer learning was employed to generalize the model across different datasets, while model interpretability was improved through visualization techniques, helping clinicians understand the predictions. However, the use of pathology images, while innovative, requires digital pathology infrastructure that may not be available in all clinical settings. The model’s excellent performance suggests it could be a valuable tool in clinical practice, but external validation in larger and more diverse patient populations is needed.

Shifting from pathology images to radiomics, Xu et al. (47) shifted the focus toward integrating radiomics features from MRI scans, offering a novel approach to personalized risk assessment. Xu et al. conducted a study which aimed to devise a personalized tool for estimating the two-year recurrence risk of bladder cancer. A model incorporating both radiomics features extracted from MRI scans and clinical factors was built using data from 71 patients. The model utilized ML methods like SVM-based recursive feature elimination. This nomogram, relying on muscle-invasive status and an radiomics-derived score, showed promising accuracy (80.95% in validation) and a high AUROC value of 0.838. Despite these strong results, potential bias due to retrospective design and single-center sampling remain as the limitations. Future investigations could explore the potential of additional factors, currently omitted due to incomplete data, in enhancing the predictive power of the model. Moreover, the role of different radiomics features in predicting lymph node status, crucial for prognosis, was recommended for future research.

Building on this, Huang et al. (67) demonstrated the potential of combining multiparametric MRI and Deep learning for predicting bladder cancer recurrence. In their study, the authors developed a clinical-radiomics deep learning model that integrated radiomics features from multiparametric MRI with deep learning and clinical data to predict the 5-year recurrence risk in NMIBC patients. With a dataset of 191 patients, the model outperformed traditional clinical models, achieving an AUC of 0.909. The use of SHapley Additive ExPlanations further enhanced the interpretability of the model, showing that radiomics features contributed significantly to prediction accuracy. However, the approach relies heavily on expensive and advanced MRI technology, which may not be accessible in many clinical settings. Furthermore, while SHAP values increase interpretability, the inherent complexity of radiomics and the deep Learning models still creates a “black box” issue, limiting clinician trust in the model’s outputs.

CT imaging has also been explored for recurrence prediction. Wang et al. (68) developed a deep learning model using multiphase enhanced CT images to predict bladder cancer recurrence, showing substantial clinical promise. This multi-center study involved 874 patients from four centers and used CNNs to develop a signature capable of predicting recurrence risk. The model demonstrated excellent performance, with an AUC of 0.889 and a concordance index of 0.869. It outperformed traditional clinical models and staging systems, indicating its potential utility in guiding personalized treatment strategies for NMIBC patients. Despite the robust performance of the deep learning model, it relies on high-resolution imaging data, which may limit accessibility in less technologically advanced medical centers. Further prospective validation and exploration of how the deep learning model integrates with clinical decision-making would be beneficial.

Finally, specialized clinical features, such as intravesical prostatic protrusion (IPP), have been studied as independent predictors of recurrence. Lee et al. (69) investigated the role of IPP in predicting NMIBC recurrence using imaging data obtained from preoperative CT urography. By analyzing the severity of IPP in 122 male NMIBC patients, the study demonstrated that structural bladder and prostate features significantly influence recurrence risk. Severe IPP (≥ 5mm) was observed in 27% of patients and was associated with a 2.6-fold increased risk of recurrence. Using SVM, incorporating IPP improved NMIBC recurrence prediction by 6%. Kaplan-Meier analysis showed that severe IPP negatively impacted recurrence-free especially in high-risk patients. However, limitations such as the small sample size, retrospective design, and lack of post-void residual data suggest further research is needed to explore the relationship between IPP, ageing, and bladder conditions. These findings position IPP as an independent risk factor for recurrence and its potential role in prognosis.

A multitude of studies have attempted to predict NMIBC recurrence, each employing unique yet complementary methodologies using biological insights combined with ML. The journey begins with Zhao et al.’s (70) paper, which focuses on ECM-related genes’ prognostic potential in Bladder Cancer. By using expression data from multiple datasets, six ECM-related genes were identified: CTHRC1, MMP11, COL10A1, FSTL1, SULF1, COL5A3. An AUC of 0.76 was achieved in recurrence prediction by training many ML models, including Generalised Linear Models, K-Nearest Neighbours, SVMs and Random Forests, on 675 non-recurrent and 285 recurrent bladder cancer patients. The study demonstrates the potential of ML and ECM gene signatures for recurrence prediction in bladder cancer, though larger sample sizes would help validate the approach. Limitations include few normal samples for comparison and lack of extensive validation of individual gene trends.

Another genomic study, this time by Cai et al. (71), analyzed loss of heterozygosity (LOH) on chromosome-18 in 65 patients with NMIBC and 43 controls. Loss of heterozygosity occurs when one of the two copies (alleles) of a gene, inherited from each parent, is lost or inactivated, which can lead to cancer if the affected gene’s role is to suppress tumors. At multivariate analysis, LOH on Chr 18 (P=0.002) and the number of lesions (P=0.03) were identified as independent predictors of recurrence-free probability. ANNs were used to confirm the multivariate analysis but the performance metrics were not mentioned. This paper, published in 2010, did not gain widespread adoption due to factors such as small study sample size; robustness of other recurrence predictors [FGR3 (72), Ki-67 (73) and NMP22 (74, 75)].

Urbanowicz (76) also aimed to uncover patterns of genetic associations with bladder cancer by applying an ML classifier system called AF-UCS (Attribute Feedback-sUpervised Classifier System). The algorithm aimed to validate findings that specific SNPs in DNA repair genes, such as XPD (Xeroderma pigmentosum group D) codon 751 and 312, along with SNPs in other DNA repair genes, are predictive of bladder cancer risk when considered alongside smoking pack-years. Their model had an accuracy of 0.66, indicating a low to moderate level of predictive performance. Since this publication in 2013, there has been limited follow-up potentially due to other ML methods having superior accuracy, and the other models not requiring specialized expertise. Additionally, large-scale Genome-Wide Association Studies are now more common than focused candidate gene approaches (77–81).

Building on the growing interest in genomic predictors, Maturana et al. (82) published a study three years later that focused on predicting outcomes of NMIBC patients using genomic SNP profiles in conjunction with clinico-pathological prognosticators. The study utilized Bayesian learning methods, including sequential threshold models and LASSO, and evaluated 822 NMIBC patients followed up for over a decade. The genomic models yielded AUROC values ranging from 0.55 to 0.62, while clinic-pathological models performed slightly better, with AUROC values between 0.57 and 0.76. These results indicate that SNP profiles alone are poor predictors of NMIBC recurrence and progression, and their inclusion in clinico-pathological models adds limited value. The limitations of the study include a relatively small sample size in certain subgroup analysis and the limited predictive ability of common SNPs in NMIBC outcomes.

Chang et al. (83) utilized urinary proteomic profiling to develop a non-invasive method for diagnosing and monitoring bladder cancer recurrence. By analyzing urine samples from 279 patients, the authors used a multi-support vector machine-recursive feature elimination (mSVM-RFE) algorithm to identify 13 protein markers for diagnosis and 11 markers for recurrence monitoring. The diagnostic model achieved high sensitivity (90.9%) and specificity (73.3%), while the recurrence monitoring model reached 75% sensitivity and 81.8% specificity. This study’s strength lies in its non-invasive nature, providing a practical alternative to invasive procedures like cystoscopy. However, the reliance on high-resolution mass spectrometry and the single-center design may limit the generalizability of the findings. The model’s utility is promising, but larger, multi-center validation studies are required to ensure its efficacy in diverse clinical settings.

Frantzi et al.’s study (84) formulated and validated urine-based biomarker panels for primary and recurrent bladder cancer detection using capillary electrophoresis-mass spectrometry (CE-MS). Their methodology Combined statistical analysis and machine learning: initial statistical tests identified significant peptide biomarkers, which were further refined using SVMs via MosaCluster software to create and optimize high-dimensional biomarker panels. Case-control comparisons across multicenter cohorts identified the biomarkers, with the primary panel achieving an AUROC of 0.87 (91% sensitivity, 68% specificity) and the recurrent panel attaining an AUROC of 0.75 (88% sensitivity, 51% specificity) during independent validation. By incorporating all available biomarkers, optimized panels improved performance further, achieving AUROCs of 0.88 for primary and 0.76 for recurrent cancer. These findings highlight the potential for non-invasive urine-based tests to complement or reduce the need for invasive cystoscopy in bladder cancer diagnosis and monitoring. Despite these promising results, the study was limited by unadjusted confounding variables, such as tumor size and hematuria, and by its cross-sectional design. Further research is needed to confirm their clinical utility and performance in real-world settings.

This study by Krochmal et al. (85) analyzed CE-MS peptidomics data (detected types and amounts of peptides in a biological sample) from 98 bladder cancer patients to develop an ML model to predict recurrence. With a training set of 50 patients, Cox regression identified 36 peptides predictive of relapse which were then input into a Random Forest model. The training set produced an accuracy of 100%, while test set accuracy is not mentioned. The limited size of the training dataset (n=48), and the unreported test accuracy, calls into question the study’s generalizability. A comparison with established clinical tools would have highlighted any advancements over existing methods.

Mucaki et al. (61) applied biochemically-inspired machine learning (ML) models, specifically supervised SVMs, to predict responses to chemotherapy agents, including cisplatin, a critical drug in bladder cancer treatment. The models were described as “biochemically-inspired” because they incorporated genes with established biological relevance to the mechanisms of action and resistance of cisplatin, ensuring the selection of features grounded in prior biochemical knowledge rather than relying purely on data-driven approaches. These gene signatures were designed to reflect pathways involved in apoptosis, DNA repair, and drug transport, enhancing interpretability and biological validity. Using a dataset of 90 cancer patients, the cisplatin-specific gene signature achieved 71.2% accuracy in predicting bladder cancer recurrence, with exceptional performance in non-smokers (100%) and 79% accuracy in smokers. The study also utilized ensemble averaging across multiple thresholds to improve model robustness. However, a key limitation was the reliance on breast cancer cell line data to train the SVM models, which limits their applicability to bladder cancer-specific contexts, as no retraining was performed on bladder cancer datasets. While this work highlights the promise of AI-driven approaches for identifying recurrence risk factors, it also underscores the need for developing bladder cancer-specific models to improve clinical relevance.

Two other studies [Zhan et al. (86) and Gogalic et al. (87)] explored the application of protein panels in recurrence prediction. Zhan et al. utilized MALAT1, PCAT-1, and SPRY4-IT1 biomarkers and achieved an accuracy of 81.3% and a sensitivity and specificity of 0.625 and 0.850 respectively. Additionally, tumor stage showed a statistically significant correlation as a predictor, with PCAT-1 identified as an independent predictor. Gogalic et al. combined common clinicopathological markers with ECadh, IL8, MMP9, EN2, and VEGF biomarkers. Their model, incorporating these markers, yielded an AUROC of 0.84.

Ajili et al. (88) applied ANNs to predict bladder cancer recurrence after Bacillus Calmette-Guerin immunotherapy. Using a multilayer perceptron model in MATLAB, the researchers incorporated patient characteristics, tumor attributes, and treatment details, achieving high performance, with a mean square error of 0.02634. The model accurately classified 39 out of 40 cases, yielding sensitivity of 96.66%, specificity of 100%, and positive/negative predictive values of 100% and 90.9%, respectively. However, the small sample size and difficulties in determining the optimal network topology (e.g., hidden layer nodes) limited the study. These results showcase the potential of ANNs in bladder cancer prognosis but illuminate the need for larger datasets and further validation before clinical application.

Zhang and Ma (89) investigated the predictive value of two clinical biomarkers, CA50 and total bilirubin (TBIL), for bladder cancer recurrence using ML models. The study evaluated the individual and combined predictive performance of these biomarkers in a cohort of 345 bladder cancer patients. The results demonstrated that CA50 had an AUC of 0.602 (p = 0.038) and TBIL had an AUC of 0.585 (p = 0.014), indicating that both biomarkers were moderate predictors of recurrence. When combined, the AUC increased to 0.623 (p = 0.013), showing an improvement in predictive power, though still relatively modest. While the combination of these two biomarkers offered better performance than either one alone, the overall predictive accuracy remained limited, suggesting that additional features or biomarkers may be necessary to enhance the model’s ability to predict bladder cancer recurrence. The study provides valuable insights into the role of clinical biomarkers in cancer prognosis but highlights the need for further optimization to improve predictive robustness.

Schwarz et al. (90) examined the role of explainability in ML models for predicting bladder cancer recurrence, enhancing their clinical utility. In this study, the authors used three ML models—SVM, gradient boosting, and ANNs—and compared them to logistic regression for predicting 2-year recurrence in urothelial carcinoma patients. Gradient boosting performed best, with an F1-score of 83.89% and AUC of 70.82%. To address the black-box nature of these models, the authors employed permutation feature importance and feature importance ranking measure to explain the most influential features driving predictions, such as therapeutic measures. This approach enhances the transparency of ML models, potentially increasing their adoption in clinical settings. However, while they attempt to address the “black-box” issue using feature importance measures like permutation feature importance and feature importance ranking measure, the interpretability of these models remains limited. These feature importance methods often fail to provide clinically actionable insights, as they simply show which features contribute most to the prediction without explaining why.

Integrating diverse feature types—such as imaging, morphological, clinical, and genomic markers—has proven to significantly enhance the predictive performance of AI models for NMIBC recurrence. By leveraging multiple data sources, these models offer a more comprehensive understanding of tumor behavior and patient risk profiles. This section reviews key studies that have successfully combined these feature types, highlighting their methodologies, outcomes, and potential for clinical application.

One of the earliest studies in this area, conducted by Catto et al. (52), aimed to compare the predictive accuracies of neuro-fuzzy modeling (NFM), neural networks (NN), and traditional statistical methods for predicting bladder cancer recurrence. The study used data from 109 NMIBC patients who were treated with TURBT. The data was used to train and test the predictive models. The study found that both NFM and NN predicted the patients’ relapse with an accuracy ranging from 88% to 95%, which was superior to statistical methods (71-77%). The difference was statistically significant, indicated by the p-value of less than 0.0006. The low p-value suggests that these differences are not just due to random chance. NFM appeared better than NN at predicting the timing of relapse (p < 0.073). Importantly, NFM offered a transparency advantage over NN, allowing for easier clinical validation and manipulation of input variables for exploratory predictions. This early work highlighted the potential of combining different data types to boost predictive power while maintaining interpretability.

Lucas et al. (91) further advanced this approach by incorporating both histopathological image features and clinical data to predict recurrence-free survival in NMIBC patients. Their model, built using Convolutional Neural Networks (CNNs) and Bidirectional GRUs, achieved a 1-year AUC of 0.62 (n=359) and a 5-year AUC of 0.76 (n=281). Their prediction model consisted of a multi-step process that combined features extracted from the histopathological images with the clinical data to create an overall outcome. Their model’s strength lies in the autonomous analysis of histopathological data, reducing human bias and possibly capturing more nuanced features. However, the study’s relatively small, single-center dataset and extended inclusion period raised concerns about generalizability and consistency in clinical practices over time. Nevertheless, the integration of image and clinical data illustrated the power of multi-modal approaches in predicting bladder cancer recurrence.

In this study (92), a deep learning-based approach (DeepSurv) was used by Jobczyk et al. to recalibrate the prediction tools for the recurrence and progression of NMIBC in a cohort of 3,892 patients. The existing risk groups, EORTC and CUETO, showed moderate performance in predicting survival outcomes. The deep learning models displayed improved accuracy. In the training group of 3,570 patients, the c-indices were 0.650 for recurrence-free survival and 0.878 for progression-free survival. In the validation group of 322 patients, the c-indices stood at 0.651 for recurrence-free survival and 0.881 for progression-free survival. The models surpassed the performance of standard risk stratification tools and demonstrated no signs of overfitting. These findings highlight the potential of deep learning models in enhancing the prediction of recurrence and progression in NMIBC, offering a valuable tool for personalized patient care.

Despite the encouraging results presented in multiple studies, several significant hurdles exist before ML-based models can be successfully implemented in clinical settings. Below, we discuss key challenges identified in our review.

To fully leverage ML in NMIBC management, future efforts should prioritize developing multimodal models, improving interpretability, integrating ML into clinical workflows, and addressing ethical and regulatory barriers.