A predictive model for risk stratification of men in west China with benign prostatic hyperplasia

doi:10.21203/rs.3.rs-4672871/v1

Download PDF

Research Article

A predictive model for risk stratification of men in west China with benign prostatic hyperplasia

https://doi.org/10.21203/rs.3.rs-4672871/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Purpose: we sought to develop and validate a model based on machine learning to predict high-risk BPH in Chinese males in physical examination.

Materials and Methods: We retrospectively included 1099 Chinese southwest males with or without BPH in West China Hospital and its related medical consortium from November 2017 to November 2022 if they were determined by their physician to be at sufficient risk to warrant the urological ultrasound examination. Pearson correlation analysis was conducted to determine the predictive factors of BPH. And gradient boosting classifier-based (GBC) algorithm based on machine learning was used to design a model to predict individual risk of BPH.

Results: A training cohort (n = 659) and validation cohort (n = 220) were randomly selected to build and validate the BPH-risk predictive model, respectively. The model was then tested on an independent cohort (n = 220), identifying men at low-risk of BPH (n = 102) for whom the urological ultrasound examination would be necessary. The highest normalized mean scores across three complementary quality indicators showed that ten factors: age, smoking status (never, former, current), drinking status (none, occasional, regular), surgical history (yes, no), comorbidities (yes, no), red blood cell count, white blood cell count, platelet count, low-density lipoprotein, glycated hemoglobin, free prostate-specific antigen and packed cell volume were significant for predicting high-risk BPH. The best final model based on GBC reached out accuracy of 97.3% with the ten predictive factors.

Conclusion: While further external test in an intended-use cohort is needed, our proposed model based on laboratory and physical examination allows us to predict the risk stratification of BPH in Chinese southwest men. And the predictive model offers a promising tool for identifying individuals at high-risk of BPH who are being considered for urological ultrasound examination.

Benign prostatic hyperplasia

Risk prediction

Machine learning

Approximately one quarter of men worldwide experience symptoms of pain, tenderness, and burning sensation during urination in the lower urinary tract, which in many cases are caused by benign prostatic hyperplasia (BPH) [1]. Convenient decision support tools for diagnosing BPH are especially important because it is one of the most frequent causes of urinary obstruction in elderly men. Conventional diagnosis requires not only laboratory and physical examinations but also ultrasonography of the prostate. However, there may not easily be ultrasound equipment in some clinics, especially in resource-limited settings. Reports linking clinicodemographic factors and examination findings to risk of BPH [2–5]. Therefore, obviating the need for imaging to make diagnosis easier and more accessible to more individuals is a possible solution.

Previous studies have developed some intelligent systems to predict BPH with a small number of samples. For example, 44 participants were conducted to develop fuzzy intelligent systems [6], and 12 samples were used to train a computer vision model for detecting glandular component hyperplasia of the prostate [7]. In addition, a study [8] built a model to classify prostate cancer and BPH from patients with Luts, rather than investigating the risk factors and diagnosis of BPH. To the best of our knowledge, no BPH predictive models used raw data from laboratory and physical examination results collected for routine care, such as electronic medical records, to identify occult high-risk BPH. Recently, machine learning, which has already proven effective at diagnosing other urological disease [9–10], could be used to predict BPH purely on the basis of laboratory and physical examinations.

In this study, we sought to develop and validate a model based on machine learning to predict high-risk of BPH in Chinese men through physical examinations. We then tested the predictive model on populations at low or high risk of BPH with the goal of identifying men who should not undergo urological ultrasound examination. This model used machine learning algorithms to risk-stratify Chinese men with BPH who could be candidates for the urological ultrasound examination.

Study Design

This retrospective and observational study that included 1099 individuals who underwent examinations at West China Hospital and its related medical consortium (Chengdu, China) between November 2017 and November 2022. We divided the samples into 3 sets randomly: 659 samples for the training of a predictive model, 220 samples for a validation set and 220 samples for a test set. The study was approved by the Biomedical Ethics Review Committee of West China hospital (approval #2022 − 1871).

Identification of Risk Factors

We initially extracted data for 22 variables that we considered relevant to risk of BPH based on the previous literature [2–5] and our surgeons’ experience. But we decided not to retain the following five variables because values were missing > 20% in the individuals: urinary glucose, albumin-to-creatinine ratio, 2-h postprandial blood glucose, fasting blood glucose, and urinary protein. We explored pairwise correlations of the remaining 17 variables and the presence of BPH using Pearson correlation analysis. Furthermore, to assess the strength of associations between a given variable and the outcome of interest, we focused on the subset of the 17 variables that were associated with the average normalized scores across three indicators (Gini index, F score, and mutual information score) [11] from the presence of benign prostatic hyperplasia (609 individuals diagnosed with BPH). From this subset, we identified and selected the top 10 variables and data missing for the selected variables were imputed using the mean value (Fig. 1). Formulas to compute the three indicators are shown in Supplementary Material 1.

Modeling

Firstly, seven machine learning algorithms were selected to model: Random Forest, Logistic Regression, Gaussian Naive Bayes, Support Vector Machine, Multilayer Perceptron, XGBoost, and Gradient Boosting Classifier. And as an eighth model, we averaged these seven models together using the mean method to create an ensemble classification model. Parameters used in the various models are shown in Supplementary Material 2. Predictive performance of the eight models was compared in terms of precision, recall, accuracy and specificity (shown in Supplementary Material 3. Finally, based on the performance, the gradient boosting classifier-based (GBC) algorithm was chosen to design a model to predict BPH-risk for a given individuals.

Based on the demographic, laboratory and physical data of the 1,099 individuals in the study (Tables 1–2), Pearson related heatmap indicated that three laboratory factors, glycosylated hemoglobin (GHb_A1c), free prostate-specific antigen (FPSA) and glucose (GLC) with the other two baseline factors (surgery history, age) were potentially related with BPH (Fig. 2a). A model was built based on these five factors. GBC performed best among the machine learning algorithms, reaching the highest accuracy (95.0%) and AUC (0.99). To assess the strength of associations between a given variable and the outcome of interest, we chose 10 variables that achieved the highest normalized mean scores across three complementary quality indicators: age, smoking status (never, former, current), drinking status (none, occasional, regular), surgical history (yes, no), comorbidities (yes, no), red blood cell count, white blood cell count, platelet count, low-density lipoprotein, glycated hemoglobin, free prostate-specific antigen and packed cell volume (Fig. 2b).

Table 1

Demographic data and findings from laboratory and physical examinations used to develop and validate models to predict benign prostatic hyperplasia
Variable		Set			P value			95%CI
Variable		Training	Validation	Test	Training	Validation	Test	Training	Validation	Test
N		659	220	220
Age (SD)		57.17 (18.15)	60.48 (17.08)	57.14 (17.05)	< 0.001	< 0.001	< 0.001	(0.344,0.389)	(0.299,0.379)	(0.302,0.377)
Male		659 (100%)	220 (100%)	220 (100%)
Drinking status					< 0.001	< 0.001	< 0.001
	None	369 (55.99)	137 (62.27)	129 (58.64)
	Occasional	225 (34.14)	68 (30.91)	69 (31.36)
	Regular	65 (9.87)	15 (6.82)	22 (10.00)
Smoking status					< 0.001	< 0.001	< 0.001
	Never	436 (66.16)	148 (67.27)	127 (57.73)
	Former	79 (11.99)	25 (11.36)	31 (14.09)
	Current	144 (21.85)	47 (21.36)	62 (28.18)
Previous surgery					< 0.001	< 0.001	< 0.001
	Yes	268 (40.67)	113 (51.36)	89 (40.45)
	No	391 (59.33)	107 (48.64)	131 (59.55)
White Blood Cell Count, mean (SD), /uL		56.76 (408.78)	30.70 (123.01)	80.22 (677.69)	0.003	0.022	0.131	(0.004,0.017)	(0.002,0.007)	(-0.003, 0.034)
Packed Cell Volume, mean (SD), 10^9/L		0.50 (1.59)	0.44 (0.058)	0.45 (0.054)	0.636	< 0.001	< 0.001	(-0.004,0.007)	(-0.002, -0.001)	(-0.001, -0.0008)
Cholesterol, mean (SD), mmol/L		4.63 (0.99)	4.63 (0.91)	4.63 (0.87)	< 0.001	< 0.001	< 0.001	(-0.073, -0.045)	(-0.078, -0.030)	(-0.069, -0.025)
Red Blood, Cells, mean (SD), 10^12/L		4.86 (0.59)	4.93 (0.69)	4.89 (0.56)	< 0.001	< 0.001	< 0.001	(-0.096, -0.073)	(-0.089, -0.041)	(-0.094, -0.054)
Glycated Hemoglobin A1c, mean (SD), %		5.91 (0.84)	5.89 (0.71)	5.91 (0.78)	< 0.001	< 0.001	< 0.001	(0.052,0.081)	(0.017, 0.062)	(0.035,0.084)
Free Prostate-Specific Antigen, mean (SD), ng/m		0.81 (2.33)	0.82 (1.97)	0.60 (0.64)	< 0.001	0.002	< 0.001	(0.019, 0.039)	(0.013,0.038)	(0.013,0.038)
Low-Density Lipoprotein, mean (SD), mmol/L		2.75 (0.83)	2.78 (0.88)	2.78 (0.81)	< 0.001	< 0.001	< 0.001	(-0.112, -0.071)	(-0.139, -0.064)	(-0.130, -0.062)
Benign prostatic hyperplasia		353 (53.57)	138 (62.73)	118 (53.64)

Table 2

Demographic data and findings from laboratory and physical examinations used to develop and validate models to predict benign prostatic hyperplasia, stratified by the presence or absence of the condition
Variable	Unit	Median	Variance
White Blood Cell Count	/uL	56.24	464.82
Platelet Count	10^9/L	196.27	62.30
Hematocrit	L/L	0.48	1.28
Red Blood Cell Count	10^12/L	4.88	0.68
High-Density Lipoprotein	mmol/L	1.23	0.35
Uric Acid	umol/L	373.94	93.47
Low-Density Lipoprotein	mmol/L	2.76	0.90
Glucose	mmol/L	5.65	1.80
Cholesterol	mmol/L	4.63	1.07
Triglycerides	mmol/L	1.76	1.76
Free Prostate-Specific Antigen	ng/m	0.76	2.12
Glycated Hemoglobin A1c	%	5.71	0.96
Fasting Blood Sugar	mmol/L	6.38	1.97
2-Hour Postprandial Blood Sugar	mmol/L	12.35	4.81
Quantitative Urinary Protein	g/L	0.095	0.29

Comparing the models built on the top five BPH-high-related factors with the top ten factors, the result showed that the model based on the heatmap had lower accuracy, which were 92.3% on the validation set and 95.0% on the test set. Models built on the top six factors and the top ten factors of the best algorithm (Gradient Boosting Classifier) performed closely. On the validation set, the two models reached out the accuracy of 96.8% and 97.2% respectively, however, they showed the same accuracy of 97.3% on the test set.

Therefore, we used the top 10 factors to build our model. Of the eight predictive models, the gradient boosting classifier showed the highest accuracy (97.3%) and AUC (0.996) against the test dataset from 220 individuals (Fig. 3). Thus, this boosting-based algorithms performed best against our validation data. The Gaussian naive Bayes model, conversely, performed the worst in terms of the various indicators. The other models showed intermediate diagnostic performance, including the ensemble model.

This study demonstrates the feasibility of predicting BPH using machine learning based on laboratory and physical data accessible to clinicians even in resource-limited settings, which may obviate the need for expensive ultrasonography of the prostate. Of 22 demographic, laboratory and clinical variables that we screened for inclusion in our models, we identified 10 as quite strongly linked to risk of the condition, and these 10 included drinking. Our results highlight the need to advise men, especially older men, to reduce these behaviors in order to minimize risk of benign prostatic hyperplasia.

The median age of our entire sample of 1,099 individuals was 58 years, and 91% of those older than the median had BPH, compared to only 18% of those younger than the median. These results highlight age as a major risk factor for the condition, consistent with the literature [12]. Further study should explore whether urological ultrasound for BPH should be recommended for all men beyond a certain age.

Among the variables that we found to be tightly associated with BPH were levels of glycated hemoglobin and free prostate-specific antigen as well as history of surgery. All these associations have previously been reported [13–15], suggesting that our modeling focused on appropriate variables. Based on these associations, future work should explore whether BPH shares pathogenic pathways with diabetes, another prevalent disease involving elevated levels of glycated hemoglobin, or with prostate cancer, which is also associated with elevated levels of free prostate-specific antigen [16]. Future work should also explore how surgeries such as appendectomy, cystolithotomy and cholecystectomy influence risk of subsequent BPH.

We surveyed a range of algorithms and combined them into an ensemble model, and we found that boosting-based algorithms were particularly effective at detecting the condition, which may reflect their abilities to adapt and learn from their mistakes [17], as well as to handle complex, non-linear relationships [19]. These models also have the advantage of being interpretable, so clinicians can understand and therefore trust their predictions [18].

Differences were small between models built from the top 6 or top 10 factors based on three complementary quality indicators, indicating that adapting the 6-factor model is effective to predict the risk of BPH when available patient data are incomplete. The model built from the top 10 factors based on three quality indicators performed better than the model built from the top 5 factors selected from the Pearson heatmap, which indicates the feasibility of using these three machine learning indicators to analyze disease-related indicators.

It may be possible to improve the performance of our predictive models by incorporating scores on the International Prognostic Scoring System, which includes evaluation of intermittent urination and narrowing of the urine stream. We did not incorporate these data because they were not routinely recorded at our hospital during the enrollment period. Future work should also aim to reduce the risk of overfitting in our models by defining the minimal set of variables needed for accurate diagnosis. Ultimately our models should be validated in larger, preferably multicenter samples with a prospective design in order to assess the ability of machine learning to predict subsequent BPH.

Despite its limitations, the present study strongly demonstrates the potential for reliable detection of BPH using machine learning-based analysis of routine laboratory and physical examinations. This may improve diagnosis of BPH, particularly in resource-limited settings, and it may facilitate large-scale urological ultrasound examination of older men.

Our study presents valuable findings and provides essential guidance for future research endeavors. Further external testing in an intended-use cohort is needed. Our proposed model based on laboratory and physical examination allows us to predict BPH risk among men in west China, which may make it a promising tool for identifying individuals at high risk of BPH who are being considered for urological ultrasound examination.

ACR

Albumin-to-creatinine ratio

AUC

Area under the curve

BPH

Benign prostatic hyperplasia

CHOL

Cholesterol

FBG

Fasting blood glucose

FPSA

Free prostate-specific antigen

GBC

Gradient boosting classifier

GHb_A1c

Glycosylated hemoglobin

Glc

Glucose

GLU

Urinary glucose

GNB

Gaussian naive bayes

HDL

High-density lipoprotein

LDL

Low-density lipoprotein

Logistic regression

Microalbuminuria

MLP

Multilayer perceptron

PBG

2-Hour postprandial blood glucose

Pca

Prostate cancer

PCV

Cell volume

PLT

Platelet count

PRO

Urine protein quantification

RBCs

Red blood cells

Random forest

Standard division

SVM

Support vector machine

TAG

Triglycerides

Uric acid

WBC

White blood cell count

XGBoost

Extreme gradient boosting

Ethics approval and consent to participate

This retrospective use of patient data was approved by the Biomedical Ethics Review Committee of our hospital (approval 2022-1871).

Consent for publication

Not applicable.

Availability of data and materials

The datasets analysed during the current study are not publicly available due to the strict regulation of the hospital but are available from the corresponding author on reasonable request.

The version of the machine learning algorithm was sklearn 1.2.0. Program codes are available at github.com/hwj20/BPH_prediction.

Competing Interests

All authors disclosed no relevant relationships

Funding

None

Authors' contributions

Qiaosen Dong and Wanjing Huang wrote the main manuscript text and prepared figures and tables. Yalan Ye and Deyi Luo reviewed the manuscript.

Acknowledgments

This work was supported by the College Student Innovation and Entrepreneurship Training Program (grants 20231170L and C2024129719) and the Health Department Project "Study on risk prediction model of prostate hyperplasia in Chengdu area based on artificial intelligence" (grant GBKT23030).

Conflicts of Interest

None declared

Lee SWH, Chan EMC, Lai YK. The global burden of lower urinary tract symptoms suggestive of benign prostatic hyperplasia: a systematic review and meta-analysis, Scientific reports, vol. 7, p. 7984, 2017.
Loeb S, Kettermann A, Carter HB, Ferrucci L, Metter EJ, Walsh PC. Prostate volume changes over time: results from the Baltimore Longitudinal Study of Aging. J Urol. 2009;182:1458–62.
Platz EA, Joshu CE, Mondul AM, Peskoe SB, Willett WC, Giovannucci E. Incidence and progression of lower urinary tract symptoms in a large prospective cohort of United States men. J Urol. 2012;188:496–501.
Hammarsten J, Högstedt B, Holthuis N, Mellström D. Components of the metabolic syndrome—risk factors for the development of benign prostatic hyperplasia. Prostate Cancer Prostatic Dis. 1998;1:157–62.
Parsons JK, Carter HB, Partin AW, Windham BG, Metter EJ, Ferrucci L, Landis P, Platz EA. Metabolic factors associated with benign prostatic hyperplasia. J Clin Endocrinol Metabolism. 2006;91:2562–8.
Torshizi AD, Zarandi MHF, Torshizi GD, Eghbali K. A hybrid fuzzy-ontology based intelligent system to determine level of severity and treatment recommendation for Benign Prostatic Hyperplasia, Computer methods and programs in biomedicine, 113, p. 301–13, 2014.
Khalid SU, Syed A, Shah SSH. Machine learning approaches for the histopathological diagnosis of prostatic hyperplasia. Ann Clin Anal Med. 2020;11:425–8.
Zhang Y, Li W, Zhang Z, Xue Y, Liu Y-L, Nie K, Su M-Y, Ye Q. Differential diagnosis of prostate cancer and benign prostatic hyperplasia based on DCE-MRI using bi-directional CLSTM deep learning and radiomics. Volume 61. Medical & Biological Engineering & Computing; 2023. pp. 757–71.
de la Cruz Martín B, Adot Zurbano JM, Gutiérrez-Mínguez E, Gómez Sánchez E, Calvo S. Tamayo Gómez E. Development of a Predictive Model for the Diagnosis of Lower Urinary Tract Obstruction in Men. J Urol. 2022;208(3):668–75.
Van Neste L, Henao R, Wojno KJ, Signes J, DeHart J, Busta A, Marriott E, Willing M, Argentini A, Hurley PM, Korman H, Hafron J, Putzi M, Pieczonka CM, Karsh LI, Morris DS, Kassis AI, Kantoff PW. Development and Optimization of a Subtraction-Normalized Immunocyte Profiling Signature for Prostate Cancer Active Surveillance Risk Stratification. J Urol. 2024;211(3):415–25.
Witten IH, Frank E, Hall MA et al. Practical machine learning tools and techniques[J]. Data Mining. Fourth Edition, Elsevier Publishers, 2017.
Bouhadana D, Lu XH, Luo JW, Assad A, Deyirmendjian C, Guennoun A, Nguyen D-D, Kwong JCC, Chughtai B. Elterman and others, Clinical Applications of Machine Learning for Urolithiasis and Benign Prostatic Hyperplasia: A Systematic Review. J Endourol. 2023;37:474–94.
Hu M, Shu X, Yu G, Wu X, Välimäki M, Feng H. A risk prediction model based on machine learning for cognitive impairment among Chinese community-dwelling elderly people with normal cognition: development and validation study. J Med Internet Res. 2021;23:e20298.
Wei X, Niu X, Zhang X, Li Y. Deep Pneumonia: Attention-Based Contrastive Learning for Class-Imbalanced Pneumonia Lesion Recognition in Chest X-rays, in 2022 IEEE International Conference on Big Data (Big Data), 2022.
Shah M, Naik N, Hameed BZ, Paul R, Shetty DK, Ibrahim S, Rai BP, Chlosta P, Rice P, Somani BK. Current Applications of Artificial Intelligence in Benign Prostatic Hyperplasia. Turkish J Urol. 2022;48:262.
Duffy MJ. Biomarkers for prostate cancer: prostate-specific antigen and beyond. Clin Chem Lab Med vol. 2020;58:326–39.
Ferreira AJ, Figueiredo MAT. Boosting algorithms: A review of methods, theory, and applications[J]. Ensemble machine learning: Methods and applications, 2012: 35–85.
Stiglic G, Kocbek P, Fijacko N, et al. Interpretability of machine learning-based prediction models in healthcare[J]. Wiley Interdisciplinary Reviews: Data Min Knowl Discovery. 2020;10(5):e1379.

No competing interests reported.

SUPPLEMENTARYMATERIAL.docx

Download PDF

Reviewers invited by journal
26 Aug, 2024
Editor invited by journal
24 Jul, 2024
Editor assigned by journal
24 Jul, 2024
Submission checks completed at journal
24 Jul, 2024
First submitted to journal
02 Jul, 2024

You are reading this latest preprint version

A predictive model for risk stratification of men in west China with benign prostatic hyperplasia

Status:

Version 1

Abstract

Figures

Introduction

Materials and methods

Study Design

Identification of Risk Factors

Modeling

Results

Discussion

Conclusion

Abbreviations

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1