A female overactive bladder risk model developed by machine learning: based on 2007–2018 NHANES data

Abstract

Background

Overactive bladder (OAB) is a urinary system syndrome that has a serious impact on daily life. Currently, the methods for estimating the risk of OAB are relatively limited, mainly relying on the symptoms reported by patients themselves. There is an urgent need to develop new risk models for the OAB diagnosis. This study aims to assess the risk of OAB in the female population by training machine learning (ML) models.

Methods

Based on the National Health and Nutrition Examination Survey (NHANES) data from 2007 to 2018, a total of 10,807 female participants were included in the model. Support vector machine (SVM), logistic regression fitting, K-nearest neighbor (KNN), random forest (RF) algorithm, gradient boosting, decision tree (DT), extreme gradient boosting (XGBoost) were used to develop OAB risk models. Ten characteristic factors were used in the construction of the models.

Results

Among the seven ML algorithms, the RF model demonstrated the best performance with an area under the curve (AUC) value of 0.879. Among the 10 characteristic factors, hypertension was the most important influencing factor, and the impact of diabetes and sleep disorders on OAB risk cannot be ignored.

Conclusions

The results show that the female OAB risk model constructed by ML technology in this study has good diagnostic performance and interpretability, which is helpful to improve the diagnosis of OAB in the female population.

Highlight box.

Key findings

• A risk model for overactive bladder (OAB) in women was established using machine learning (ML) algorithms. Moreover, SHapley Additive exPlanations (SHAP) was used to interpret the contribution of each variable.

What is known and what is new?

• Despite the fact that existing literature has reported the associations between OAB and lifestyle habits, diet, physical examination findings, and blood biochemistry, no previous studies have integrated various types of variables to develop a diagnostic model for OAB.

• In our research, we not only constructed an ML model based on multiple types of variables but also explored the impact of each variable on the OAB diagnostic model using SHAP. This compares and clarifies the importance of different variables for OAB.

What is the implication, and what should change now?

• This study mainly focused on the diagnostic model of OAB in the female population. In the future, we plan to develop a model for males, as well as a model for all populations that includes gender as an incorporated variable.

Introduction

Background

Overactive bladder (OAB) is a syndrome characterized by urinary urgency. It is often accompanied by frequent urination and nocturia (1-3). Some patients also have urgent urinary incontinence. These symptoms bring great trouble to the patient ’s daily life, social activities and mental health. The prevalence of OAB increases in association with increasing age. Therefore, with the development of aging society, the prevalence of OAB will increase (4,5). The pathogenesis of OAB is complex, which may be related to detrusor overactivity, central nervous system abnormalities, pelvic floor dysfunction and other factors (6,7).

Rationale and knowledge gap

The diagnosis of OAB is an exclusionary clinical diagnosis, which requires excluding organic diseases through laboratory tests and imaging examinations, and then making a comprehensive diagnosis by combining the objective assessment of subjective symptoms (8). Urodynamic testing can improve diagnostic accuracy for atypical cases, those with combined lower urinary tract symptoms, or refractory cases, and its clinical utility has been widely recognized by both patients and physicians (9). However, the proficiency of technicians in operating instruments and equipment varies significantly across different hospitals, which may lead to deviations in test results. Physical examination, laboratory examination and imaging examination all lack specific diagnostic indicators (10).

Machine learning (ML) is an artificial intelligence strategy that predicts unknown data by learning from known data. Compared with classical statistical models (such as linear regression), ML is not limited by two-dimensional data, but also can deal with multidimensional. At the same time, when there are too many independent variables in the model, ML method can be used to reduce multidimensional data. The ML method can remove irrelevant variables and improve the generalization of the model. At the present, ML has gradually appeared in the medical field and is widely used in imaging-based auxiliary diagnosis and treatment, disease model construction, prognosis evaluation and drug efficacy prediction (11).

Objective

Based on the National Health and Nutrition Examination Survey (NHANES) data from 2007 to 2018, this study used ML technology to construct an OAB risk model, aiming to improve the efficiency and accuracy of female OAB diagnosis. We present this article in accordance with the TRIPOD reporting checklist (available at https://tau.amegroups.com/article/view/10.21037/tau-2025-282/rc).

Methods

Research design and interviewees

NHANES is a nationally representative health and nutrition survey project. It aims to collect information on the health and nutritional status of American residents by investigating people of different ages, races, and socioeconomic backgrounds. It provides data support for monitoring national health trends and formulating public health policies and interventions. This study selected female respondents who participated in the NHANES survey from 2007 to 2018 as the research object, and collected a total of 20,902 samples. All data were extracted before December 2024. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

OAB diagnosis

In this study, the Kidney Conditions-Urology questionnaire in NHANES was collected, and the results of questionnaires related to urgency incontinence and nocturia were collected to evaluate OAB (12). Specific questions include: ‘During the past 12 months, have you leaked or lost control of even a small amount of urine with an urge or pressure to urinate and you couldn’t get to the toilet fast enough? How frequently does this occur? During the past 30 days, how many times per night did you most typically get up to urinate, from the time you went to bed at night until the time you got up in the morning?’. By quantifying the symptom score of OAB, respondents with a total score of more than 3 points were judged as OAB.

Covariate evaluation

According to the description of OAB risk factors in the existing literature (13), the covariates included in this study comprised demographic information, physical activity level, sleep disorders, female reproductive health status, depression status, caffeine intake, smoking and drinking habits, diabetes history, hypertension history, blood routine examination results and common cardiovascular and cerebrovascular diseases.

Considering that ML requires a high amount of data, in order to ensure that the number of samples used for model training is sufficient, the study first counted the missing values in the covariates, and eliminated the covariates with a missing value ratio of more than 70%. Finally, 206 variables were excluded from the 233 original covariates, and 27 covariates were retained. Those covariates included age, race, education level, poverty-to-income ratio (PIR), white blood cell count (WBC), lymphocyte count, neutrophil count (NEU), high-density lipoprotein (HDL), high-density lipoprotein cholesterol index (HDLSI), platelet (PLT), total cholesterol (TC), total cholesterol index (TCSI), body mass index (BMI), glycosylated hemoglobin (glycohemoglobin), hypertension, diabetes, vigorous work activity, moderate work activity, walking or cycling, vigorous recreational activity, moderate recreational activity, minutes sedentary activity, trouble sleeping, caffeine intake (Caffein), ‘Age when first menstrual period occurred’, ‘Had regular periods in past 12 months’, ‘Ever taken birth control pills?’.

Data pre-analysis

In this study, Spearman correlation coefficient was used to test the correlation between covariates, and then BORUTA algorithm (14) was used to further screen covariates and obtained the importance of each covariate. Finally, the variables included in the training set and the test set were determined.

ML model construction and training

Due to the imbalance of binary variables in the initial data, the performance of the ML model may be affected. In this study, the synthetic minority oversampling technique (SMOTE) was used to process the data to improve the performance of the model (15). After determining the standard scale of the included features, the preprocessed data were randomly divided into training set and test set according to the ratio of 3:1. Seven ML algorithms, including LR, support vector machine (SVM), random forest (RF), gradient boosting, K-nearest neighbor (KNN), decision tree (DT) and extreme gradient boosting (XGBoost) (16), were used to construct a model for estimating the risk of OAB in the training set, and the model was applied to the test set for verification. All the above operations were based on Python 3.12.0.

Model performance evaluation

The accuracy, precision, recall, F1 score and area under the curve (AUC) were used to evaluate the performance of the seven ML models (17). By drawing the receiver operating characteristic (ROC) curve and confusion matrix, the performance evaluation results of the model were visually displayed.

The interpretation method of the constructed ML model

Different from the intuitive interpretation of the coefficients in the simple linear model, the constructed ML model has the characteristics of ‘black box’ and requires specific methods to explain (18). SHapley Additive exPlanations (SHAP) is a predictive interpretation method of ML model based on game theory (19). It is proposed by Lundberg and Lee, which provides a unified framework for explaining various black box ML models. Each feature in each sample has a corresponding SHAP value. The positive SHAP value indicates that the feature has a positive impact on the model, and the negative SHAP value indicates that it has a negative impact on the model. By calculating the weighted average of SHAP values of each feature, the overall impact of the feature on the risk model can be evaluated. The relationship between features and risk models was revealed by using macro feature importance map and macro bee colony map (20).

Statistical analysis

All statistical analysis and mapping work in this study were completed with the help of R 4.3.1 and Python 3.12.0. A two-tailed P value of less than 0.05 was considered statistically significant.

Results

Baseline characteristics of the study population

Figure 1 shows the dataset construction and ML training process. The study excluded 10,095 female participants with missing data and finally included 10,807 female participants. According to whether the OAB score is greater than 3, the participants in the study were divided into the OAB group and the non-OAB group, of which the OAB group had 2,541 people.

Figure 1.

Research flow chart. KNN, K-nearest neighbor; NHANES, National Health and Nutrition Examination Survey; OAB, overactive bladder; SVM, support vector machine; XGBoost, extreme gradient boosting.

Grouping analysis revealed significant differences between the OAB group and non-OAB group in terms of AGE, RACE, EDUCATION, PIR, and multiple laboratory indices including WBC, NEU, HDL, HDLSI, TC, TCSI and Glycohemoglobin. Anthropometric parameters such as BMI and comorbidities like hypertension, diabetes were also distinct between groups. In the female reproductive health project, there were also differences between the two groups in terms of “having regular periods in the past 12 months” and “ever having taken birth control pills”. Additionally, significant disparities were observed in lifestyle factors including moderate work activities, walking/cycling, vigorous recreational activities, moderate recreational activities, sleep disturbances, and caffeine intake. Detailed results are presented in Table 1. Compared with the non-OAB group, the participants in the OAB group were older and had the lower education level and poverty income ratio. There was no significant difference between the two groups in terms of lymphocyte, age when first menstrual period occurred, vigorous work activity, PLT, as well as minutes sedentary activity.

Table 1. Characteristics of the research participants.

Characteristics	Non-OAB, N=8,266	OAB, N=2,541	P value
Age (years)	46 [17]	57 [17]	<0.001
Race			<0.001
Mexican American	1,234 (15%)	366 (14%)
Other Hispanic	882 (11%)	309 (12%)
Non-Hispanic White	3,774 (46%)	1,034 (41%)
Non-Hispanic Black	1,485 (18%)	686 (27%)
Other race (including multi-racial)	891 (11%)	146 (6%)
Education			<0.001
Less than 9th grade	567 (6.9%)	365 (14%)
9–11th grade	957 (12%)	490 (19%)
High school graduate	1,752 (21%)	584 (23%)
Some college or AA degree	2,771 (34%)	771 (30%)
College graduate or above	2,219 (27%)	331 (13%)
PIR	2.58 [1.65]	2.03 [1.47]	<0.001
WBC (1,000 cells/μL)	7.28 [4.85]	7.49 [2.34]	<0.001
Lymphocyte (1,000 cells/μL)	2.25 [4.03]	2.22 [0.78]	0.90
NEU (1,000 cells/μL)	4.27 [1.71]	4.46 [1.88]	<0.001
PLT (1,000 cells/μL)	259 [64]	258 [74]	0.052
HDL (mg/dL)	58 [16]	56 [17]	<0.001
HDLSI (mmol/L)	1.49 [0.42]	1.46 [0.43]	<0.001
TC (mg/dL)	195 [41]	199 [42]	<0.001
TCSI (mmol/L)	5.05 [1.05]	5.15 [1.09]	<0.001
BMI (kg/m2)	29 [7]	32 [8]	<0.001
Glycohemoglobin (%)	5.62 [0.91]	6.06 [1.33]	<0.001
Age when first menstrual period occurred (years)	13 [2]	13 [2]	0.50
Had regular periods in past 12 months			<0.001
Yes	4,501 (54%)	708 (28%)
No	3,765 (46%)	1,833 (72%)
Ever taken birth control pills?			<0.001
Yes	5,795 (70%)	1,602 (63%)
No	2,471 (30%)	939 (37%)
Hypertension			<0.001
Yes	2,494 (30%)	1,397 (55%)	<0.001
No	5,772 (70%)	1,144 (45%)	<0.001
Diabetes			<0.001
Yes	736 (8.9%)	593 (23%)
No	7,530 (91%)	1,948 (77%)
Trouble sleeping			<0.001
Yes	2,220 (27%)	1,032 (41%)
No	6,046 (73%)	1,509 (59%)
Caffeine (mg)	133 [155]	124 [172]	<0.001
Vigorous work activity			0.20
Yes	1,012 (12%)	288 (11%)
No	7,254 (88%)	2,253 (89%)
Moderate work activity			<0.001
Yes	2,831 (34%)	746 (29%)
No	5,435 (66%)	1,795 (71%)
Walk or bicycle			<0.001
Yes	1,948 (24%)	514 (20%)
No	6,318 (76%)	2,027 (80%)
Vigorous recreational activities			<0.001
Yes	1,721 (21%)	199 (7.8%)
No	6,545 (79%)	2,342 (92%)
Moderate recreational activities			<0.001
Yes	3,659 (44%)	778 (31%)
No	4,607 (56%)	1,763 (69%)
Minutes sedentary activity	356 [203]	351 [203]	0.20

Data are presented as mean [SD] or n (%). BMI, body mass index; HDL, high-density lipoprotein; HDLSI, high-density lipoprotein cholesterol index; NEU, neutrophil; OAB, overactive bladder; PIR, poverty-to-income ratio; PLT, platelet; SD, standard deviation; TC, total cholesterol; TCSI, total cholesterol index; WBC, white blood cell.

Correlation analysis between variables and characteristic value screening

Figure 2 illustrates the correlation analysis of features. Hypertension was correlated with recent menstrual status (had regular periods in past 12 months) (Spearman’s r=0.4), age (Spearman’s r=0.47), and glycohemoglobin (Spearman’s r=0.36), demonstrating moderate positive associations. HDL exhibited a notable correlation with body mass index (BMI; Spearman’s r=0.35), while diabetes showed a moderate link with glycohemoglobin (Spearman’s r=0.47). Correlations among remaining variables were negligible or absent. Overall, inter-variable correlations were minimal, confirming the independence of included characteristic values.

Figure 2.

Spearman’s analysis examines the relationships among different variables. BMI, body mass index; HDL, high-density lipoprotein; HDLSI, high-density lipoprotein cholesterol index; NEU, neutrophil; PIR, poverty-to-income ratio; PLT, platelet; SD, standard deviation; TC, total cholesterol; TCSI, total cholesterol index; WBC, white blood cell.

Given the large number of variables remaining after initial screening—most with good independence—the BORUTA tool, based on the RF algorithm, was further employed to screen variables for ML. Ultimately, 10 optimal variables were identified for the training set: education level, recent menstrual status, hypertension, BMI, PIR, diabetes, vigorous recreational activities, moderate recreational activities, sleep disturbances, and age (Table 2).

Table 2. Perform variable selection using BORUTA.

Characteristic	Rank (importance of characteristic)	Selected
Education	1	True
Had regular periods in past 12 months	1	True
Hypertension	1	True
BMI	1	True
PIR	1	True
Diabetes	1	True
Vigorous recreational activities	1	True
Moderate recreational activities	1	True
Ever told doctor had trouble sleeping?	1	True
Age	1	True
Race	2	False
Glycohemoglobin	2	False
Vigorous work activity	2	False
Caffeine	2	False
TC	3	False
PLT	4	False
WBC	5	False
Ever taken birth control pills?	5	False
NEU	7	False
Walk or bicycle	8	False
HDL	9	False
Moderate work activity	10	False
Age when first menstrual period occurred	11	False
Minutes sedentary activity	12	False
Lymphocyte	12	False
HDLSI	14	False
TCSI	14	False

BMI, body mass index; HDL, high-density lipoprotein; HDLSI, high-density lipoprotein cholesterol index; NEU, neutrophil; PIR, poverty-to-income ratio; PLT, platelet; TC, total cholesterol; TCSI, total cholesterol index; WBC, white blood cell.

Model performance comparison

After SMOTE processing, 16,532 participants were divided into a training set (N=12,399) and a test set (N=4,133). Figure 3 shows the ROC curves and confusion matrices of the seven ML models. Among them, the ROC curve of the RF model is closest to the upper left corner, and the performance is the best (AUC =0.879, accuracy =0.795, precision =0.776, recall =0.830, F1 score =0.802). Table 3 lists the accuracy, AUC, precision, recall and F1 scores of each model. Although, in this disease risk model, the F1 score of KNN (F1 score =0.807) is slightly higher than that of RF, RF performs more excellently in terms of the AUC, accuracy, and precision. These metrics are of great significance for the accurate assessment of disease risks. The results show that the RF model performs best in the training set and the test set. Therefore, the RF model was selected for subsequent analysis.

Figure 3.

ROC curve and confusion matrix for seven ML models. (A) ROC curve and confusion matrix for logistic regression. (B) ROC curve and confusion matrix for SVM. (C) ROC curve and confusion matrix for KNN. (D) ROC curve and confusion matrix for random forest. (E) ROC curve and confusion matrix for gradient boosting. (F) ROC curve and confusion matrix for decision tree. (G) ROC curve and confusion matrix for XGBoost. (H) ROC curves for multiple classifiers. AUC, area under the curve; FPR, false positive rate; KNN, K-nearest neighbor; ML, machine learning; ROC, receiver operating characteristic; SVM, support vector machine; TPR, true positive rate; XGBoost, extreme gradient boosting.

Table 3. ML performance index.

Algorithm	AUC	Accuracy	Precision	Recall	F1-score
Logistic regression	0.790	0.717	0.721	0.709	0.715
SVM	0.771	0.698	0.690	0.722	0.706
KNN	0.863	0.788	0.743	0.883	0.807
Random forest	0.879	0.795	0.776	0.830	0.802
Gradient boosting	0.807	0.729	0.726	0.740	0.733
Decision tree	0.733	0.733	0.720	0.764	0.741
XGBoost	0.845	0.767	0.758	0.787	0.772

AUC, area under the curve; KNN, K-nearest neighbor; ML, machine learning; SVM, support vector machine; XGBoost, extreme gradient boosting.

Explainable RF algorithm using SHAP

Through the SHAP analysis, the SHAP values of the characteristic values of the training set are shown, reflecting the importance of each feature. The results show that hypertension contributes the most to the ML model, followed by age and “Ever told doctor had trouble sleeping?” (Figure 4A). From the perspective of the influence of characteristic values on the model output, most of the red points in hypertension, “Ever told doctor had trouble sleeping?” and diabetes features are distributed on the right side of the zero line, indicating that these features are positively correlated with OAB and have a positive impact on the overall output (Figure 4B). The distribution of age feature points is relatively scattered, with both positive and negative effects, indicating that the influence of age on model output is more complicated (Figure 4B). The distribution of education feature points is also scattered, but the overall high characteristic values (red dots) have a positive impact on the model output (Figure 4B). The impact by BMI, PIR, moderate recreational activities, “Had regular periods in past 12 months”, vigorous recreational activities is small (Figure 4B).

Figure 4.

SHAP for OAB risk model. (A) The SHAP summary plot of included variables. (B) The SHAP beeswarm plot of included variables. BMI, body mass index; OAB, overactive bladder; PIR, poverty-to-income ratio; SHAP, SHapley Additive exPlanations.

Discussion

In this study, a female OAB risk model was constructed, and seven ML models were used for training. The RF algorithm was finally selected by combining AUC, accuracy, precision, recall rate and F1 score. RF can be used to construct a model with multiple characteristic values and prevent overfitting (21). The model contains 10 characteristic parameters and has excellent performance (AUC =0.879). SHAP value analysis showed that blood pressure contributed the most to the model and had a positive effect, which provided an important basis for the diagnosis of OAB.

Before performing BORUTA analysis, we conducted a correlation analysis on the included variables to determine their suitability for BORUTA analysis. Spearman correlation analysis showed high correlations existed between hypertension and recent menstrual status, age, glycohemoglobin, as well as between diabetes and glycohemoglobin. After variable screening using BORUTA, among the variables included in the model, only hypertension and recent menstrual status showed a high correlation, which suggests that hypertension and recent menstrual status may have partially shared mechanisms in influencing OAB. Finally, after SHAP analysis, the contribution of hypertension was higher than that of recent menstrual status, indicating that priority should be given to managing hypertension in the prevention of OAB.

Hypertension is more common in the middle-aged and older adult population (22). Previous studies have found that OAB can induce pre-voiding hypertension in men and affect blood pressure management in men, but no effect of OAB on hypertension was observed in women (23,24). However, the participants included in this study were all women, and hypertension played a significant role in the RF model. Compared with traditional regression models, ML models are more sensitive to complex associations (25). This suggests that hypertension may be a potentially important feature affecting the prevalence of OAB in women, and compared with previous studies, our research can capture the nonlinear relationship between the included variables and OAB through the construction of ML models.

Sleep disorder is a common problem in women (26). Studies have confirmed that sleep disorders are positively correlated with the occurrence of OAB (27,28). OAB patients often have symptoms of poor sleep and frequent nights, which in turn affects the mental state during the day. In the model of this study, sleep problems have a positive effect on the model output, and the SHAP value is higher, which further confirms the close relationship between sleep quality and OAB. For women with sleep disorders, early prediction of the risk of OAB and timely implementation of intervention measures are of vital importance, as sleep disorders may exacerbate bladder dysfunction through circadian rhythm disorders or autonomic nervous system dysfunction.

This study also found that diabetes is an important risk factor for female OAB. SHAP analysis showed that diabetes had a strong positive contribution to the risk model. Clinical studies have shown that the prevalence of OAB in diabetic patients is significantly higher than that in non-diabetic patients, up to 20–40%, and with the prolongation of the course of diabetes, poor blood glucose control and complications, the risk of OAB further increases (29,30). Diabetes and OAB interact with each other. Diabetes increases the risk of OAB through a variety of mechanisms (31), OAB not only significantly affects the quality of life of patients with diabetes, but also has a synergistic effect with diabetes, in which diabetes promotes the development of OAB (32). Therefore, for patients with diabetes, attention should be paid to the screening and evaluation of OAB symptoms, so as to facilitate early detection and treatment, improve the quality of life of patients, and reduce complications.

However, there are some limitations in this study. Limited by the database, the exposure duration and dose of risk factors related to living habits were not considered. The inability to perform exclusion diagnosis is also one of the limitations. These factors need to be further explored in future studies.

Conclusions

In this study, seven ML models were constructed to explore the association between female OAB risk and living habits, common chronic diseases and demographic data. Compared with the other six algorithms, the RF model performed better. Among the 10 characteristics, hypertension was the most important influencing factor in the model, and the effects of sleep disorders and diabetes on the risk of OAB can not be ignored. Future research can develop more advanced algorithms to further verify the results of this study, and optimize relevant parameters to improve the accuracy of the model.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.