INTRODUCTION
Lower urinary tract symptoms (LUTS) commonly affect a significant portion of the male population, especially with advancing age. These symptoms, which include a slow urinary stream, hesitancy, frequent urination, urgency, and nocturia, can severely impact quality of life, often causing considerable discomfort and inconvenience [1]. The causes of LUTS are multifactorial, with benign prostatic hyperplasia (BPH) being a major contributor. Although the initial approach to treatment usually involves medical management with alpha-blockers and 5-alpha reductase inhibitors, surgical intervention may become necessary if these conservative measures are unsuccessful [2].
Surgical treatment options, including transurethral resection of the prostate, holmium laser enucleation of the prostate (Ho-LEP), and newer minimally invasive techniques, are designed to relieve bladder outlet obstruction and improve urinary flow [3]. However, predicting the outcomes of these surgical interventions remains challenging, especially in patients with impaired bladder contractility [4]. Diminished bladder contractility, often referred to as detrusor underactivity (DU), complicates the clinical picture and makes it difficult to predict the level of symptomatic relief after surgery [5-9]. This uncertainty necessitates a more nuanced approach to patient selection and preoperative counseling.
Several technologies have been developed to efficiently and quantitatively analyze male LUTS/BPH. A range of wearable and nonwearable devices are available for this purpose; however, appropriate interpretation of the results requires both urological and technical knowledge. With the advancement of artificial intelligence (AI) technology, various deep learning (DL) and machine learning (ML) systems capable of interpreting quantitative analyses have been developed. In most studies, the predictions of recovery rates in LUTS and BPH were performed using DL and ML methods [10,11].
The AI model, which effectively categorizes multiple classes and improves the recovery rate, is based on a deep neural network (DNN) [12]. DNNs have demonstrated high performance in solving complex, nonlinear ML problems across various industries and scientific disciplines. Additionally, extreme gradient boosting (XGBoost) is a robust ML algorithm used for regression and classification tasks. The random forest classifier (RFC) is an ensemble learning algorithm that enhances prediction accuracy by integrating multiple decision trees. Similarly, the support vector machine (SVM) is a potent ML algorithm employed for both classification and regression analysis [13,14]. We utilized XGBoost, RFC, and SVM for a comparative analysis.
Our study aimed to develop an AI-based predictive model to assess the likelihood of symptomatic improvement following surgical treatment for male LUTS, focusing on patients with compromised bladder contractility. By harnessing the power of AI, we sought to provide a more reliable prognostic tool capable of guiding clinicians in making informed decisions, ultimately improving patient outcomes and optimizing the use of healthcare resources. This research represents a significant step towards personalized medicine in the management of LUTS, offering hope for better-targeted therapies and enhanced patient satisfaction.
MATERIALS AND METHODS
Study Subjects and Definitions
We reviewed 1,933 patients who underwent HoLEP from July 2008 to January 2024 at Samsung Medical Center. Prior to surgery, all patients were evaluated using their past medical history, the International Prostate Symptom Score (IPSS), the International Continence Society Male Short-Form Questionnaire, a 3-day voiding diary, transrectal ultrasonography (TRUS), and a urodynamic study (UDS). Postoperatively, patients returned to the clinic at 1 or 2 weeks and at 1 month, followed by every 3 months for assessments of uroflowmetry, postvoid residual (PVR), and IPSS. The primary outcomes of this study were the rate of change in maximum flow rate (MFR) and voiding efficiency (VE) at one month postoperatively. VE was defined as the percentage of volume voided compared to the prevoid bladder volume (voided volume/[voided volume+PVR]). This study received approval from the institutional review board at our institution.
Data Preprocessing
We included 1,142 patients with complete data, dividing them into 2 groups: 992 patients for model training from July 2008 to December 2022, and 150 patients for external validation from January 2023 to January 2024. The DNN employed was designed for multiclass classification, aiming to simultaneously predict changes in the MFR and VE 1 month postsurgery. These changes were categorized into 3 classes: for MFR, class I ranged from 0 to 15 mL/sec, class II from 15 to 20 mL/sec, and class III from 20 to 60 mL/sec; for VE, class I ranged from 0 to 0.5, class II from 0.5 to 0.75, and class III from 0.75 to 1.0. This model was used to predict recovery rates in a nuanced manner. The dataset exhibited class imbalance, which posed challenges during analysis. To mitigate this, we applied random over-sampling (ROS) to the training set, a widely used technique to address class imbalance by duplicating instances of the minority class to equalize the dataset [15]. This technique is known to enhance the model’s predictive accuracy on minority classes and improve overall performance. By implementing the ROS method, we increased the training set to 1,266 patients. The prediction model for both MFR and VE was trained using 33 variables with no missing data (Supplementary material 1).
In addition, the categorical data underwent one-hot encoding, a method that converts categorical data into a binary format. This approach prevents the implication of any ordinal relationship, reduces bias, and ensures compatibility with various DL and ML algorithms [16]. After applying one-hot encoding, the number of features increased from 33 to 44 variables. The numerical data was scaled using the Standard Scaler, a fundamental preprocessing step that improves the performance and efficiency of ML models. The Standard Scaler transforms each feature, x, in the dataset as follows:
where z is the standardized value, x is the original data value, μ is the mean of the data, and σ is the standard deviation of the data. The Standard Scaler is indispensable in the development of robust ML models by ensuring that the features contribute equally, improving convergence speed and making model interpretation easier, as shown in equation (1).
AI Models for Recovery-Rate Classification
To effectively analyze urology data 1-month postsurgery, a DNN was developed for data learning. This DNN utilized multiclassification, categorizing data into 3 classes based on the changes in MFR and VE, respectively. It accounted for the nonlinearity among variables and benefited from the ability to automatically learn data characteristics without manual feature setting. The probability value for each class was determined by applying the SoftMax function to the output layer of the DNN model. Additionally, we implemented multitask learning (MTL) within the DNN framework to increase learning efficiency and performance. MTL improves generalization and robustness by exploiting commonalities and differences across tasks, allowing the model to train on multiple related tasks simultaneously [17].
XGBoost is a high-performance ML algorithm used for regression and classification tasks, RFC is a type of ensemble learning algorithm that improves prediction performance by combining multiple decision trees, and SVM is a powerful ML algorithm used for classification and regression analysis. We conducted a comprehensive evaluation using our dataset to assess the performance of various AI models that employed XGBoost, RFC, and SVM for comparative analysis.
To optimize hyperparameters such as the number of layers, number of nodes, and batch size, we employed a grid search. This method systematically searches through a specified subset of hyperparameters to determine the optimal combination for the models. We integrated the least squares method (LSM) into the grid search process. The LSM calculates the standard deviation across all classes for each hyperparameter combination. This calculated standard deviation serves as the scoring metric, guiding the selection of optimal hyperparameters that balance performance across classes. The LSM is defined as:
where E is the total error, n is the number of observations, xi is the actual value for the i-th observation and x̂i is the predicted value for the i-th observation [18].
Evaluation of AI Models
To evaluate the performance of the AI models, the dataset of 1,266 patients was prepared for stratified 5-fold cross-validation, split in a 4:1 ratio. Additionally, another 150 patients from the dataset were set aside for independent external validation. The internal validation set was randomly divided into 5 parts, and the internal validation test was conducted 5 times. External validation was performed on the last 20% of the dataset using the fully developed AI models [19].
Feature importance is a technique that identifies which features in a dataset most significantly impact the predictive performance of a model. In this study, we utilized permutation importance to quantify a feature’s importance by measuring the change in model performance when the feature values are randomly shuffled [20]. Additionally, we used a feature importance metric that incorporates both the statistical randomness of data splitting from 5-fold cross-validation and the mechanical errors caused by systematic biases such as random numbers, iterations, scoring, and inaccuracies inherent in the modeling process.
RESULTS
Baseline Characteristics
The mean age of the patients was 70.4 years, and the mean prostate volume was 66.7 mL. The preoperative mean MFR was 7.9 mL/sec, and the mean PVR was 120 mL. In UDS, the mean bladder outlet obstruction index (BOOI) was 55.8, and the mean bladder contractility index (BCI) was 104.8. Patient characteristics are described in detail in Table 1.
Recovery-Rate Classification Using the AI Model
The DNN model demonstrated the best performance in terms of its microaverage area under the receiver operating characteristic curve (AUC) and the highest multilabel specificity, although its multilabel sensitivity was relatively low (microaverage AUC 0.884±0.006, multilabel sensitivity 0.783±0.020, multilabel specificity 0.891±0.010) in multiclassification for the amount of change in MFR. For VE, the DNN model also showed superior performance (microaverage AUC 0.817± 0.007, multilabel sensitivity 0.660±0.014, multilabel specificity 0.830±0.007). We evaluated the performance of 3 other ML models—namely, XGBoost, RFC, and SVM—to enable a relative comparison with the DNN models. Each model was trained and tested using a consistent dataset split, with 80% allocated for training and 20% for testing in the internal dataset. For the amount of change in MFR, the XGBoost model (microaverage AUC 0.842±0.017, multilabel sensitivity 0.790±0.022, multilabel specificity 0.895±0.011), the RFC model (microaverage AUC 0.841±0.013, multilabel sensitivity 0.788±0.018, multilabel specificity 0.894±0.009), and the SVM model (microaverage AUC 0.698±0.013, multilabel sensitivity 0.597± 0.017, multilabel specificity 0.7959±0.009) all presented lower performance than the DNN model. Similarly, for VE, the XGBoost model (microaverage AUC 0.776±0.013, multilabel sensitivity 0.701±0.018, multilabel specificity 0.851±0.009), the RFC model (microaverage AUC 0.751±0.027, multilabel sensitivity 0.668±0.037, multilabel specificity 0.834±0.018), and the SVM model (microaverage AUC 0.638±0.023, multilabel sensitivity 0.517±0.031, multilabel specificity 0.759±0.015) also showed lower performance than the DNN model. The statistical analysis indicated that the microspecificity of the DNN model and the XGBoost model was statistically similar, with no significant difference in their values within one standard deviation. Among the 4 types of multi-classifications, the DNN model exhibited the best performance for both the amount of change in MFR and VE. The detailed performance of each model with external validation is presented in Table 2. We identified meaningful features and ranked the top 5 features in order of significance as follows: for the amount of change in MFR, TRUS_prostate total volume (0.046±0.012 [stat]±0.003 [sys]), serum prostate-specific antigen (PSA) (0.044±0.012 [stat]± 0.003 [sys]), Uroflow_PVR (0.041±0.013 [stat]±0.002 [sys]), Uroflow_MFR (0.041±0.012 [stat]±0.003 [sys]), and age (0.031±0.013 [stat]±0.002 [sys]); for VE, TRUS_T zone volume (0.033±0.014 [stat]±0.003 [sys]), serum PSA (0.032± 0.012 [stat]±0.002 [sys]), age (0.022±0.012 [stat]±0.002 [sys]), IPSS_voiding subscale (0.021±0.012 [stat]±0.003 [sys]), and Uroflow_PVR (0.020±0.011 [stat]±0.002 [sys]) as shown in Fig. 1. Here, “stat” represents statistical error and “sys” denotes systematic error.
DISCUSSION
This study presents a significant advancement in the predictive modeling of surgical outcomes following HoLEP, specifically focusing on VE and MFR. A key innovation of our study is the use of BCI and BOOI as continuous variables, rather than relying on categorical thresholds, to more accurately reflect the individual characteristics of each patient.
While DU is commonly identified through urodynamic studies, it is primarily defined as “a contraction of reduced strength and/or duration, resulting in prolonged bladder emptying and/or failure to achieve complete bladder emptying within a normal time span” [21]. Although this definition is descriptive, many studies define male DU using a BCI below 100 as a convenient cutoff [7,8,22-24]. However, this method tends to oversimplify the complex interplay between bladder contractility and bladder outlet obstruction. While BCI and BOOI are effective tools for assessing bladder function, using these indices as categorical variables may lead to an oversimplification that fails to account for the unique combination of BCI and BOOI in each patient. Therefore, our study considered BCI and BOOI as continuous variables, allowing us to capture the full spectrum of these interactions and provide a more detailed and accurate prediction of postsurgical outcomes.
A meta-analysis that examined the outcomes of HoLEP in patients with DU primarily utilized a BCI threshold of less than 100 to categorize patients [23]. While this binary approach is useful for broad analyses, it does not consider the significant variability in BCI and BOOI values among individuals. The combination of bladder contractility and bladder outlet obstruction varies greatly from one patient to another, complicating the application of a universal threshold. For instance, 2 patients might both have a BCI below 100, yet one could have a considerably higher BOOI, resulting in differing surgical outcomes. By treating BCI and BOOI as continuous variables, we can more accurately capture this variability and offer more personalized predictions for each patient. This approach enables clinicians to evaluate not merely whether a patient’s BCI falls below or above a specific threshold, but also how the interaction between BCI and BOOI affects postsurgical outcomes. This is especially crucial for accurately predicting MFR and VE, which heavily depend on the dynamics between bladder contractility and outlet obstruction. Our model’s incorporation of these continuous variables renders it a more effective tool for predicting outcomes than traditional models that depend on fixed cutoffs.
Our DNN model outperformed other ML models, including random forests and logistic regression, in predicting surgical outcomes. The strength of the DNN model lies in its capacity to discern complex, nonlinear relationships among multiple preoperative variables, such as continuous BCI and BOOI values. This capability enabled our model to achieve higher accuracy, with AUC values of 0.884 for MFR and 0.817 for VE, surpassing traditional models like the random forest model developed by Ito et al. [25], which achieved an AUC of 94.7%. By incorporating a broader array of preoperative factors, including prostate volume, PSA levels, and PVR, our model provides a more comprehensive assessment of each patient’s unique condition. The inclusion of continuous BCI and BOOI values enhances the model’s predictive accuracy, allowing for a more personalized approach. This is especially beneficial for patients with DU or borderline bladder function. Such an approach equips clinicians with more precise tools for preoperative counseling and decision-making, especially in cases involving impaired bladder contractility.
The inclusion of VE as a key outcome in our predictive model is supported by a substantial body of prior research, which demonstrates that bladder function can recover following the surgical correction of bladder outlet obstruction, even in patients with DU. Several studies have indicated that patients with a low BCI can still experience significant improvements in voiding function following HoLEP. These findings challenge the traditional reluctance to recommend surgery for patients with impaired bladder contractility, offering new hope for those who might otherwise be considered poor surgical candidates. Our model builds on these findings by providing a more individualized prediction of VE, supported by the continuous measurement of BCI and BOOI. This allows clinicians to offer a more accurate prognosis for patients with DU, ensuring that they receive the best possible care. By predicting how VE will improve postoperatively, our model can assist in preoperative counseling, helping to reassure patients that their bladder function is likely to improve after surgery, even in cases of significant detrusor acontractility.
In conclusion, our findings demonstrate that AI models can effectively predict changes in both MFR and VE, offering valuable guidance for clinicians managing patients with LUTS and compromised bladder contractility. The model’s ability to integrate a broad range of preoperative factors, including individualized BCI values, positions it as a potentially transformative tool in urology. Future research should aim to expand the dataset, include longer-term outcomes, and investigate additional preoperative factors to enhance the model’s accuracy further. By building on these findings, AI-based models have the potential to revolutionize BPH management, enabling truly personalized surgical planning and improved patient outcomes.
PDF Links
PubReader
ePub Link
XML Download
Full text via DOI
Download Citation
Supplement
Print
