https://orcid.org/0000-0001-5262-7399
Figures
Abstract
While the 10-year survival rate for localized prostate cancer patients is very good (>98%), side effects of treatment may limit quality of life significantly. Erectile dysfunction (ED) is a common burden associated with increasing age as well as prostate cancer treatment. Although many studies have investigated the factors affecting erectile dysfunction (ED) after prostate cancer treatment, only limited studies have investigated whether ED can be predicted before the start of treatment. The advent of machine learning (ML) based prediction tools in oncology offers a promising approach to improve the accuracy of prediction and quality of care. Predicting ED may help aid shared decision-making by making the advantages and disadvantages of certain treatments clear, so that a tailored treatment for an individual patient can be chosen. This study aimed to predict ED at 1-year and 2-year post-diagnosis based on patient demographics, clinical data and patient-reported outcomes (PROMs) measured at diagnosis. We used a subset of the ProZIB dataset collected by the Netherlands Comprehensive Cancer Organization (Integraal Kankercentrum Nederland; IKNL) that contained information on 964 localized prostate cancer cases from 69 Dutch hospitals for model training and external validation. Two models were generated using a logistic regression algorithm coupled with Recursive Feature Elimination (RFE). The first predicted ED 1 year post-diagnosis and required 10 pre-treatment variables; the second predicted ED 2 years post-diagnosis with 9 pre-treatment variables. The validation AUCs were 0.84 and 0.81 for 1 year and 2 years post-diagnosis respectively. To immediately allow patients and clinicians to use these models in the clinical decision-making process, nomograms were generated. In conclusion, we successfully developed and validated two models that predicted ED in patients with localized prostate cancer. These models will allow physicians and patients alike to make informed evidence-based decisions about the most suitable treatment with quality of life in mind.
Citation: Hasannejadasl H, Roumen C, van der Poel H, Vanneste B, van Roermund J, Aben K, et al. (2023) Development and external validation of multivariate prediction models for erectile dysfunction in men with localized prostate cancer. PLoS ONE 18(3): e0276815. https://doi.org/10.1371/journal.pone.0276815
Editor: Alvaro Galli, CNR, ITALY
Received: January 21, 2022; Accepted: October 14, 2022; Published: March 3, 2023
Copyright: © 2023 Hasannejadasl et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: Due to patient privacy we are not allowed to share the original dataset publicly. Customized data sets from the Netherlands Cancer Registry (NCR) can be made available under strict terms and conditions. All applications for data will be reviewed for their compliance with national privacy legislation as well as IKNL objectives. The dataset involved in this study can be requested via IKNL.nl website (https://iknl.nl/forms/dataapplication) and should be referred to by its reference number: K19.117.
Funding: This study was funded by Zorg Instituut Nederland which officially granted to Esther J. Bloemen- van Gurp.
Competing interests: NO authors have competing interests
Introduction
Globally, prostate cancer is the second most common cancer among men [1, 2]. In the Netherlands, more than 12,000 men are diagnosed with prostate cancer every year, which makes it the most common cancer among men [3]. Diagnosis and treatment of prostate cancer not only affect the patient’s quality of life (QoL) in the short term. Post-treatment QoL studies also indicate worsening of the patient’s sexual, urinary and bowel problems in the long term [4]. As a result, patients newly diagnosed with localized prostate cancer and their doctors face the challenge of choosing the appropriate treatment option from the main options including radical prostatectomy (RP), brachytherapy (BT), external beam radiotherapy (EBRT), and active surveillance (AS) [5]. Despite similar survival rates, each treatment has a specific side effect profile, and the way they affect the QoL of patients differs. The most common side effect that affects QoL in these patients is ED, a condition in which the patient is unable to get or reliably hold an erection [6–8]. ED generally appearsafter treatment and affects many patients in the long term with over 80% reporting ED up to 42 months after diagnosis [8, 9].
The rate of ED largely depends on the treatment given, individual characteristics such as age, and pre-treatment sexual function that can be measured by patient-reported outcome measures (PROM). Accurate prediction of men at increased risk of ED development could lead to better decision-making and a reduction of complications from the condition. Shared decision-making (SDM) is the preferred model to inform patients in preference-sensitive situations such as localized prostate cancer. SDM sessions are more effective when patients are provided with evidence-based information that takes their values into account as well as their risks [10, 11].
Given the influence of individual characteristics of a patient on the development of post-treatment ED, personalized prediction models are required. However, the majority of current ED risk prediction models were only built for specific types of treatments or using predictors that are not available at baseline. For example, Oh et al. built a model to predict the risk of developing ED after radiotherapy [12]. However, the lack of routine access to genetic markers at the time of disease diagnosis challenges this model’s applicability. In addition, despite the importance of external validation to assess the reproducibility of the model in a new population, most studies are limited to internal validation [13]. The aim of this study is to solve these issues by developing and externally validating multivariate prediction models that predict ED after treatment in localized prostate cancer patients using treatment information, clinical data, and PROMs that are available at baseline.
Methods
Dataset
A nationwide prospective study aimed at the evaluation of prostate cancer care in the Netherlands was performed in previous years, resulting in the ProZIB dataset (ProstaatkankerZorg In Beeld or “Insight into Prostate Cancer Care”) [14]. This study was embedded in the Netherlands Cancer Registry which is hosted by the Netherlands Comprehensive Cancer Organisation. The ProZIB study was initiated as a collaboration between the Netherlands Comprehensive Cancer Organisation, all medical associations involved in prostate cancer care, and patient advocacy groups [14]. A subset of this ProZIB dataset was provided for the current study and contains clinical data and PROMs based on the EPIC26 questionnaire at diagnosis and at 12 and 24 months after diagnosis. All patients with localized or locally advanced prostate cancer, but without regional lymph node involvement or distant metastasis (cT1-3, N0, M0) were included. Further details about the larger dataset are described in the article by Vernooij et al. [14].
At diagnosis, the following data elements were available: (i) PROMs data from the EPIC26 questionnaire; (ii) tumor characteristics, such as tumor staging, PSA at diagnosis and ISUP (International Society of Urological Pathology) Gleason grade group; (iii) patient characteristics, such as age, height, weight, smoking status, comorbidities, and information regarding treatment. Next to the baseline PROMs, PROMs 12 and 24 months after diagnosis were available. In addition, the dataset also contained information about the hospital at which the diagnosis and/or treatment took place.
Treatment categories.
Various (combinations of) treatments were reported in the dataset. These were combined into 4 treatment categories that are generally chosen after diagnosis:
- RP with or without lymph node dissection and with/without hormone therapy.
- EBRT alone, with lymph node dissection, or with lymph node dissection and hormone therapy.
- BT alone or combined with hormone therapy.
- No active therapy (NAT) including AS as well as watchful waiting (WW).
Fifteen patients were excluded for not receiving any of the treatment options or receiving a combination of treatments. Hormone therapy was added as a separate variable alongside the selected treatment categories of interest.
This resulted in a dataset containing a total of 949 patients.
Outcome.
The outcome predicted in these prediction models is the frequency of erections, question 10 in the EPIC-26 questionnaire, “How would you describe the FREQUENCY of your erections during the last 4 weeks?” [15]. People who answered "never" were considered to be free of ED, while other answers indicated partial or full ED. To classify patients that could never have an erection and those that still could, we performed a binary transformation of the outcome. This binary transformation combined patients that answered number 1 (never) into one group and those that answered 2–5 in the other group.
Data preprocessing
Missing data.
Where possible, missing values were removed. First, missing values were removed on a variable level, where variables with high portions of missing values were removed. Then missing values were removed on a patient level, i.e. patients with more missing values than the 95th percentile of the distribution of missing values were removed. Since more values were missing in the EPIC-26 questionnaire data at the 2-year time point than the 1-year time point, two separate datasets were created to achieve the highest possible amount of data for both time points. Fig 1 shows the flow of the study population.
Each model’s variables and data are shown in gray.
Batch effects.
Principal Component Analysis (PCA) [16] was used to investigate potential batch effects between groups of patients. Batch effects are potentially confounding effects present in the data that are correlated to the outcome that is to be predicted, but not of biological or medical origin. In our analyses, PCA was used to test whether the following input variables caused any batch effects in our dataset: the hospital of diagnosis, age, tumor staging, PSA level, the treatments given, the presence of diabetes, the presence of cardiovascular disease, alcohol use, and smoking status.
Training and test set.
The original dataset contained data of 964 patients diagnosed and treated in 69 hospitals in the Netherlands. According to the TRIPOD statement [17], independent statistical validation is essential to guarantee reproducibility of prediction models in a wider population. It classifies external validation into two types; type 3 and 4. Type 3 TRIPOD validation refers to external validation with a dataset that is gathered at a different location or time. We created a type 3 validation dataset that differed based on location based on the following:
with the following protocol:75% of the total number of patients were selected for the training set and 25% for the test set.
AND
75% of all hospitals in the dataset were selected for the training set and 25% for the test set, i.e. all patients from one hospital were selected for either the training set or the test set, but no hospital occurred in both sets.
- Hospitals were sorted based on the number of patients diagnosed at that specific hospital.
- The hospital with the largest number of patients was selected for the training set and the second-largest hospital was selected for the test set.
- Each subsequently-largest hospital was then placed in either the training or test set based on the distribution of patients in the training set vs. test set (75% vs. 25% +- 5%). In other words, if the training set contained more than 80% of the selected patients, the next group of patients from one hospital was placed in the test set. If the test set exceeded 30% of the entire dataset, the next group of patients was added to the training set and so on.
By randomly allocating hospitals there was the risk of having a training set that is disproportionately small or large compared to the test set. Therefore, we opted for this method that allocates hospitals based on their patient size and would get us as close as possible to the desired division of 75/25 for training and testing.
Data analysis
Univariate analyses.
Before the statistical modeling, we investigated which demographic variables were statistically significant between patients that could no longer get an erection and those that still could based on a Student’s t-test in the case of a normal distribution or Wilcoxon rank test in the case of a non-normal distribution. In addition, a false discovery rate correction (FDR) [18] was applied using the R “fdrtool” package [19] to generate q-values and reduce the chances of reporting false positives. A similar approach was taken to investigate potential significant differences between the training and test set for the following variables: (i) age at diagnosis; (ii) clinical tumor stages; (iii) PSA at diagnosis; (iv) ISUP grade group.
Multivariate analyses.
To predict the ED outcome at 1 and 2 years post-diagnosis, we used a bootstrapped binomial logistic regression coupled with Recursive Feature Elimination from the “glmnet” [20] and “caret” [21] packages in the R programming language (version 3.6.3). After the development of the prediction models, Receiver Operating Curves (ROCs) were generated using the R “pROC” package [22] and sensitivity, specificity and overall accuracy were calculated. In addition, we created a nomogram of both logistic regression models with the “rms” package in R [23].
Calibration.
Despite the importance of discrimination in classifying patients, the degree of quantitative reliability of the model for the observed results is crucial in clinical decision-making. In well-calibrated models, the predicted results are consistent with the results of the study population. The results show that calibrated models are more helpful in practice even with lower AUC [24]. For the above-mentioned reason, we performed calibration of the original models for the 1-year and 2-year models. Specifically, we estimated and updated the intercept of the original model (“recalibration in the large”). For the calibration plots we used the proposed methodology of Van Calster et al. [25] and the calibration plots function of Gerts et al. [26] using the “ModelGood” package in R [27].
Open science
Due to patient privacy we are not allowed to share the original dataset publicly. However, in order to adhere as much as possible to the open science principles, we have made available the code to generate the results presented in this manuscript on GitHub (https://github.com/riannefijten/codeErectileDysfunctionArticle). Customized data sets from the Netherlands Cancer Registry (NCR) can be made available under strict terms and conditions. All applications for data will be reviewed for their compliance with national privacy legislation as well as IKNL objectives. The dataset involved in this study can be requested via IKNL.nl website (https://iknl.nl/forms/dataapplication) and should be referred to by its reference number: K19.117.
Results
Characteristics of the patient population in our cohort are displayed in Table 1. Our subcohort had a high prevalence of cardiovascular disease (CVD); 52% of patients had some form of cardiovascular disease. Diabetes was less prevalent, with 12% of patients suffering from it. Most patients in our subcohort received no active therapy (n = 351 and 270, at 1 and 2 years respectively), and 279 & 225 patients received radical prostatectomy, 166 & 133 received EBRT and 52 & 42 patients received brachytherapy.
Preprocessing resulted in two datasets; one for each outcome time point. The dataset for the 1-year outcome consisted of 848 patients and the 2-year dataset contained 670 patients due to a higher level of non-response on the EPIC-26 questionnaire at this later time point. These datasets were also checked for non-outcome related batches, but no batch effects were found.
The distribution of the ED outcome of each time point in the dataset is reported in Fig 2. Despite only 19% reporting complete ED at baseline, this increased to 46% and 47% at 1 and 2 years, respectively. In contrast, 79% of men were able to get an erection at diagnosis, which decreased to 54% and 53% within the first and second years (Fig 3).
The ROC curves of the independent test set for the model predicting the reliability 1 year (A) and 2 years (B) post-diagnosis.
Subsequently, the dataset was split up into a training and validation set and tested for differences. In the 1-year dataset, no significant differences were found between the training- and test set for age (Wilcoxon p = 0.72), tumor T stage (Wilcoxon p = 0.95), tumor N stage (Wilcoxon p = 0.09), PSA at diagnosis (Wilcoxon p = 0.46) and ISUP grade group (Wilcoxon p = 0.07). In the 2-year dataset, significant differences were found between the training and test set for tumor T and N stage (Wilcoxon p = 0.02 for both), but not for age (Wilcoxon p = 0.12), PSA at diagnosis (Wilcoxon p = 0.12), and ISUP grade group (Wilcoxon p = 0.80). The distribution of T stages between the training and test sets differs slightly, mainly due to a larger percentage of patients in the test set having stage 1 and a greater percentage of patients in the training set having an N0 stage.
Model training
Two models were created. The first model predicted the 1-year ED outcome based on 10 variables with a training accuracy of 77.7% with a sensitivity and specificity of 89.9% and 66.8% respectively. The selected variables included both patient-related factors such as alcohol use and clinical variables such as the tumor T stage and the ISUP grade group. The second model predicted the 2-year outcome based on 9 variables and had a training accuracy of 73.9% with a sensitivity of 86.4% and a specificity of 62.3%. Several variables overlapped between the two models, but several were different. For instance, cardiovascular disease was important to the 1-year outcome, but not the 2-year outcome.
In order to encourage reproducibility, Tables 2 and 3 contain the logistic regression coefficients required to compute the model. Calibration-in-the-large was performed but did not elicit a substantial in the coefficient provided in these tables [for 1-year the updated intercept is 0.05485955 and for the 2-year is 0.1988651]. In these tables the univariate statistical significance of each variable independently is displayed. These results show that each of the selected variables is statistically significant except for alcohol use for the 1-year outcome and abdominal/pelvic/rectal pain for the 2-year outcome. Other variables that were not selected by the algorithm were also significant between the patients that could not get an erection and those that still could, such as the PSA at diagnosis and sexual functioning. These are shown in detail S2 and S3 Tables.
In addition, the univariate FDR-corrected q-value is reported for each variable.
In addition, the univariate FDR-corrected q-value is reported for each variable.
Furthermore, statistical testing between training and test data for important variables including demographics, predictors, and ED outcome was performed for both time points. For each predictor, we tested for significant differences between the training and validation sets. In the 1-year dataset, significant differences were found between the training and test sets for alcohol use. In the 2-year dataset, significant differences were found between the training and test set for tumor T and diabetes. The distribution of T stages between the training and test sets differs slightly, mainly due to a larger percentage of patients in the test set having stage 1 and a greater percentage of patients in the training set having an N0 stage. Details of the statistical results are shown in S4 and S5 Tables.
Model validation
After model training, the (independent) validation set was used to validate the model and thus identify overfitting. The accuracy, sensitivity and specificity of the models on the test set are depicted in Table 4. The overall accuracy of the 1-year model was 3.1% higher than that of the 2-year model and had a better specificity.
The associated validation set ROC curves are shown in Fig 3 and report AUCs of 0.84 and 0.81 for 1 year and 2 years post-diagnosis respectively.
Calibration
We generated the calibration plots for the original and the calibrated model for the 1-year and 2-year models that displayed in Figs 4 and 5. The curves of the calibrated models did not show a significant difference in terms of overestimating or underestimating the risk of ED compared to the original models.
The solid black line represents an ideal model where the predicted probabilities correspond with the actual probabilities. The gray band represents the standard error of the mean (SEM) while the solid blue line shows the performance of the model.
The solid black line represents an ideal model where the predicted probabilities correspond with the actual probabilities. The gray band represents the standard error of the mean (SEM) while the solid blue line shows the performance of the model.
Nomograms
In addition to reporting the coefficients and accuracy of the models, we also translated both models to nomograms for easy use. The nomogram for the 1-year model is displayed in Fig 6 and the one for the 2-year model is displayed in Fig 7.
Each predictor was given a score on a point scale representing its level of impact on ED development. The highest total score was 100 points for the treatment group, highlighting the importance of this variable. Meanwhile, equal scores for predictors like diabetes and hormone therapy in the 1-year nomogram, indicate the same influence of them on ED. The low score assigned to alcohol use after one year and hormone treatment after two years indicate a relatively small impact of these variables on outcome.
The outcome probability can be calculated by adding the total score and placing it on the total points scale. For example, a man with a very small problem of abdominal pain (20 points), presence of diabetes (0 points), Charlson comorbidity index of 1 (20 points), ISUP grade group of 1 (40 points) tumor stage of T2 (10 points), frequency of erection 50% times (35 points), quality of erection of not at all (0 points), having hormone therapy (0 points), choosing EBRT as treatment (35 points) had a total score of 160 points, indicating a 20% chance of not having ED after 2 years.
Discussion
In this study, we developed and validated a logistic regression model (TRIPOD type 3) that predicts the frequency of ED 1 and 2 years post-diagnosis based on data available at the moment of diagnosis. These models are 75.3% and 73.7% accurate in the external validation set for the 1-year and 2-year outcomes respectively. For the purpose of external validation in this study, we divided the data by hospital, so that the data set from one hospital was either used to train or validate. A comparison between training and test data for important variables including demographics, predictors, and ED outcome was performed (item 13c in TRIPOD guideline). Despite the significant differences between the training and test set for alcohol use in the 1-year dataset, and tumor T and diabetes in the 2- year dataset, the external validation was successful. It is crucial for a model to be generalizable across a variety of populations with different distributions, which is referred to as model generalizability. Nomograms were also created to predict the ability of erection based on the analysis results for two-time points. In addition, both models calibrated and calibration curves showed agreements between predicted and observed probabilities.
In spite of the fact that ED has five levels of severity, we used a binary method for several reasons. Considering medical professionals’ opinions and the fact that even mild ED impacts the personal lives of patients significantly [28], a binary model that predicts the risk of developing ED is more applicable. Further, because of the uneven distribution of five levels of ED within the population, predicting the rare level(s) is challenging and consequently prone to error. As a result, binomial logistic regression, was used in our study.
The variables to predict the ED outcome were related to the tumor as well as patient-related factors; for example the tumor T stage and the quality of their erections at diagnosis. Since ED after treatment was related to pre-treatment potency status, in this study we used PROMs data that patients filled in at three-time points: before treatment, 1-year, and 2-year post-treatment to capture a person’s perception of their own health. The two included comorbidities were also predictive of the outcomes, with cardiovascular disease predictive of the 1-year outcome and diabetes as indicative of the outcome at both time points. This is consistent with previous literature that these two diseases influence ED in patients with prostate cancer [29, 30].
In both validations, the sensitivity (the accuracy for patients that could never have an erection) was substantially higher than the specificity (the accuracy for patients that could still get an erection). This is most likely a result of the heterogeneity introduced by combining patients with ED complaints ranging from mild to severe for our binary classification with logistic regression.
Despite ED being the most frequent side effect of prostate cancer treatment, there are not many suitable and clinically applicable models in the scientific literature that predict the ED chance at a certain time point with data available upon diagnosis. Furthermore, there are currently no AI-based prediction models that predict ED for patients that undergo active surveillance. One study developed a web-based prediction model to predict common prostate cancer side effects such as sexual function after treatment. The models were internally validated and they used bias-corrected R-squared values to report their models’ performance which cannot be compared to our results [31].
Oh et al. investigated the use of genetic markers to determine the chances of developing ED after radiotherapy in 236 patients at a single institute. The data was collected from patients with at least one year follow-up. They used random forest regression to build the model and internally validated the model [12]. The AUC of the model is 0.65, which is lower than our models. In addition, genetic markers are most likely not available at diagnosis which could hamper its clinical applicability.
In addition, Haskins et al. predicted the chance of ED two years post-treatment in 776 patients who underwent prostatectomy and radiotherapy based on pretreatment conditions. Multivariate logistic regression was used and resulted in an AUC of 0.71 for prostatectomy and 0.66 for radiotherapy, which is lower than the accuracy of the models in our study. The model was not externally validated, but internally validated by bootstrapping [32].
A study similar to ours predicted ED two years after treatment based on pretreatment demographics and clinical data of 1027 patients from nine centers in the USA. For each primary treatment option, including prostatectomy, BT, and EBRT, separate models were generated and validated externally [33]. AUC is the only performance metric that was reported in this study (0.77 for prostatectomy, 0.87 for EBRT, and 0.90 for BT) which shows that their models perform similarly well to ours but use different predictors, such as age for all groups and BMI for BT. In our models, age was not selected as a predictive variable for our outcome. We did not have sufficient data to include the BMI. In both studies, pretreatment erectile function and hormone therapy were reported as predictors. Furthermore, no information is provided about how their data was split into training and validation sets.
In addition to the statistical performance of a model, it is important that the predictors identified in our study are clinically explainable and relevant to patients [34]. Our prediction models were based on a combination of clinical information, such as tumor T stage and the presence of cardiovascular disease or diabetes, and patient characteristics, such as the presence of erectile dysfunction at baseline. The most common predictors mentioned in the literature are age, pretreatment potency, and neurovascular bundle injury during treatment [35, 36]. Preservation of the neurovascular bundle is however a variable that will not be known at diagnosis and is therefore not applicable to our use case. Alemozaffar et al. [33] showed that the nerve-sparing surgery technique is related to the chance of keeping an erection. Age was included in our analysis, but was not selected as predictive for ED. Pretreatment potency was defined by two variables in our dataset (the erectile frequency and quality) and was found as predictive of ED after 1 or 2 years. In addition, previous research has found a correlation between diabetes and treatment-related ED [35]. This is consistent with our findings that diabetes is predictive of ED at both time points, yet was not found by Alemozaffar et al. [33].
Considering the multiple treatment options, treatment decision-making can be challenging, especially for younger men who have a longer life expectancy [37]. With time, more men will show significant improvement in erection function if the nerve was not damaged. For instance, eighteen months post-surgery, ED rates decrease to 56% among those who undergo two-sided nerve-sparing procedures. However, patients in all groups experience worse erection, after two to three years. Our results are consistent with previous studies that found 78% of patients had ED fully or partially, one year and 81% two years after diagnosis. In addition, the odds of men developing ED are higher with RP compared to EBRT or BT (85% vs. 50%) [8]. As a result, clinicians will be able to provide more guidance to patients during their consultations.
However, it is not clear which combination of treatments and identified factors reduce the risk of ED. In other words, does a 70-year-old patient with a history of cardiovascular disease benefit more from surgery or radiotherapy? The importance of this topic is for guiding SDM sessions from population-based to individual-based, implementing personalized medicine and deciding on the optimal treatment.
Further studies are needed to identify subgroups of patients who have a lower chance of developing ED with the benefit of a certain treatment.
To the best of our knowledge, this is the first study that provided nomograms to predict the risk of ED for localized prostate cancer. A nomogram is an easy-to-use tool to estimate the probability of an outcome by adding scores of each predictor. Since the ED is predictable at the diagnosis, doctors can use the generated nomograms to identify patients at risk of ED development and reduce the burden of the disease.
Despite its strengths, this study has some limitations. First, we used regression logistics in this study, which investigates the occurrence ED in a binary way. However, ED is classified into different degrees in terms of quality and frequency. Further research is needed to explore the relationship between ED and predictors with classification models such as linear regression. Second, as all the data were collected from hospitals in the Netherlands, results might differ in other regions or countries [38]. For instance, different rates of ethnic diversity and different cancer detection and treatment execution in different regions could influence the applicability of our models to other populations. We would therefore highly encourage researchers from other countries to validate our model with their own data. In addition, the data shows a clear preference in the treatment decision-making for prostatectomy and watchful waiting, with much fewer patients receiving brachytherapy or EBRT. Our models could therefore potentially be less accurate for these patients. Furthermore, only two comorbidities (CVD and diabetes) were considered as risk factors for ED post-treatment due to the evidence for their role in ED development in scientific literature. Therefore, other comorbidities might play a role in ED development. Furthermore, we did not apply imputations to avoid the risk of bias in our results. Nevertheless, future studies should consider multiple imputations which can create accurate imputations for handling missing values. Finally, as a result of data limitations, this study followed patients for up to two years after treatment. According to our results, there was no significant difference in the percentage of men affected by ED at 1 year versus 2 years, suggesting that ED is a long-term side-effect. Future studies should investigate how many patients will be affected by this complication in the long term.
Conclusion
In conclusion, we developed and externally validated prediction models for ED at 1 and 2 year post-diagnosis based on data from Dutch prostate cancer patients. In this study we identified tumor T stage, frequency and quality of erections, lack of energy, ISUP group, presence of CVD, diabetes, hormone therapy, and treatment group as the most important predictors of posttreatment ED.
According to the results of our study, ED is not a transient, but instead, a long-term side effect, and men who undergo prostatectomy have a higher risk of developing ED than those who choose other treatment methods. Using this information, patients are able to make informed decisions about their own treatment preferences together with their clinicians by taking a closer look at the risks and benefits of each alternative. Our models will be incorporated into a patient decision aid (PDA) to aid SDM, enabling care providers and patients to make informed decisions about localized prostate cancer treatment focusing on QoL.
Supporting information
S1 Table. The variables used as input for the prediction models for both years.
https://doi.org/10.1371/journal.pone.0276815.s001
(PDF)
S2 Table. Results of the statistical testing between all possible predictors and the erectile dysfunction outcome at 1-year post-diagnosis.
https://doi.org/10.1371/journal.pone.0276815.s002
(PDF)
S3 Table. Results of the statistical testing between all possible predictors and the erectile dysfunction outcome at 2 years post-diagnosis.
https://doi.org/10.1371/journal.pone.0276815.s003
(PDF)
S4 Table. Results of the statistical testing between training and validation sets for important variables including demographics, predictors, and erectile dysfunction outcome in the 1-year dataset.
These results include the p-value for each potential variable.
https://doi.org/10.1371/journal.pone.0276815.s004
(DOCX)
S5 Table. Results of the statistical testing between training and validation sets for important variables including demographics, predictors, and erectile dysfunction outcome in the 2-year dataset.
These results include the p-value for each variable.
https://doi.org/10.1371/journal.pone.0276815.s005
(DOCX)
References
- 1. Rawla P. Epidemiology of Prostate Cancer. World J Oncol. 2019 Apr;10(2):63–89. pmid:31068988
- 2. Aladwani M, Lophatananon A, Ollier W, Muir K. Prediction models for prostate cancer to be used in the primary care setting: a systematic review. BMJ Open. 2020 Jul 1;10(7):e034661. pmid:32690501
- 3.
Cijfers over prostaatkanker [Internet]. [cited 2021 Apr 23]. Available from: https://www.zgt.nl/aandoening-en-behandeling/onze-specialismen/de-prostaatkliniek/nieuws/cijfers-over-prostaatkanker
- 4. Houédé N, Rébillard X, Bouvet S, Kabani S, Fabbro-Peray P, Trétarre B, et al. Impact on quality of life 3 years after diagnosis of prostate cancer patients below 75 at diagnosis: an observational case-control study. BMC Cancer. 2020 Aug 12;20(1):757.
- 5. Ankolekar A, Vanneste BGL, Bloemen-van Gurp E, van Roermund JG, van Limbergen EJ, van de Beek K, et al. Development and validation of a patient decision aid for prostate Cancer therapy: from paternalistic towards participative shared decision making. BMC Med Inform Decis Mak. 2019 Jul 11;19(1):130. pmid:31296199
- 6.
UpToDate [Internet]. [cited 2021 Feb 15]. Available from: https://www.uptodate.com/contents/image?imageKey=PI%2F106517
- 7. Dyer A, Kirby M, White ID, Cooper AM. Management of erectile dysfunction after prostate cancer treatment: cross-sectional surveys of the perceptions and experiences of patients and healthcare professionals in the UK. BMJ Open [Internet]. 2019 Oct 3 [cited 2021 Feb 15];9(10). Available from: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6797309/ pmid:31585974
- 8. Emanu JC, Avildsen IK, Nelson CJ. Erectile Dysfunction after Radical Prostatectomy: Prevalence, Medical Treatments, and Psychosocial Interventions. Curr Opin Support Palliat Care. 2016 Mar;10(1):102–7. pmid:26808052
- 9. Downing A, Wright P, Hounsome L, Selby P, Wilding S, Watson E, et al. Quality of life in men living with advanced and localised prostate cancer in the UK: a population-based study. Lancet Oncol. 2019 Mar;20(3):436–47. pmid:30713036
- 10. Field J, Holmes MM, Newell D. PROMs data: can it be used to make decisions for individual patients? A narrative review. Patient Relat Outcome Meas. 2019 Jul 29;10:233–41. pmid:31534379
- 11. Verma D, Bach K, Mork PJ. Application of Machine Learning Methods on Patient Reported Outcome Measurements for Predicting Outcomes: A Literature Review. Informatics. 2021 Sep;8(3):56.
- 12. Visweswaran S, Ferreira A, Ribeiro GA, Oliveira AC, Cooper GF. Personalized Modeling for Prediction with Decision-Path Models. PLOS ONE. 2015 Jun 22;10(6):e0131022. pmid:26098570
- 13. Nazha A, Komrokji R, Meggendorfer M, Jia X, Radakovich N, Shreve J, et al. Personalized Prediction Model to Risk Stratify Patients With Myelodysplastic Syndromes. J Clin Oncol. 2021 Nov 20;39(33):3737–46. pmid:34406850
- 14. Ng K, Sun J, Hu J, Wang F. Personalized Predictive Modeling and Risk Factor Identification using Patient Similarity. AMIA Summits Transl Sci Proc. 2015 Mar 25;2015:132–6. pmid:26306255
- 15. Vernooij RWM, Cremers RGHM, Jansen H, Somford DM, Kiemeney LA, van Andel G, et al. Urinary incontinence and erectile dysfunction in patients with localized or locally advanced prostate cancer: A nationwide observational study. Urol Oncol. 2020 Sep;38(9):735.e17–735.e25. pmid:32680820
- 16. Principal component analysis. Chemom Intell Lab Syst. 1987 Aug 1;2(1–3):37–52.
- 17. Collins GS, Reitsma JB, Altman DG, Moons KGM, TRIPOD Group. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement. The TRIPOD Group. Circulation. 2015 Jan 13;131(2):211–9. pmid:25561516
- 18. Benjamini Y, Hochberg Y. Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing. J R Stat Soc Ser B Methodol. 1995;57(1):289–300.
- 19. Strimmer BK and K. fdrtool: Estimation of (Local) False Discovery Rates and Higher Criticism [Internet]. 2021 [cited 2021 Jul 2]. Available from: https://CRAN.R-project.org/package=fdrtool
- 20. Friedman J, Hastie T, Tibshirani R, Narasimhan B, Tay K, Simon N, et al. glmnet: Lasso and Elastic-Net Regularized Generalized Linear Models [Internet]. 2021 [cited 2021 Jul 2]. Available from: https://CRAN.R-project.org/package=glmnet
- 21. Kuhn M, Wing J, Weston S, Williams A, Keefer C, Engelhardt A, et al. caret: Classification and Regression Training [Internet]. 2021 [cited 2021 Jul 2]. Available from: https://CRAN.R-project.org/package=caret
- 22. Robin X, Turck N, Hainard A, Tiberti N, Lisacek F, Sanchez J-C, et al. pROC: Display and Analyze ROC Curves [Internet]. 2021 [cited 2021 Jul 2]. Available from: https://CRAN.R-project.org/package=pROC
- 23. Jr FEH. rms: Regression Modeling Strategies [Internet]. 2021 [cited 2021 Jul 2]. Available from: https://CRAN.R-project.org/package=rms
- 24. On behalf of Topic Group ‘Evaluating diagnostic tests and prediction models’ of the STRATOS initiative, Van Calster B, McLernon DJ, van Smeden M, Wynants L, Steyerberg EW. Calibration: the Achilles heel of predictive analytics. BMC Med. 2019 Dec;17(1):230.
- 25. Van Calster B, Nieboer D, Vergouwe Y, De Cock B, Pencina MJ, Steyerberg EW. A calibration hierarchy for risk models was defined: from utopia to empirical data. J Clin Epidemiol. 2016 Jun;74:167–76. pmid:26772608
- 26. Gerds TA, Andersen PK, Kattan MW. Calibration plots for risk prediction models in the presence of competing risks. Stat Med. 2014 Aug 15;33(18):3191–203. pmid:24668611
- 27. Gerds TA. ModelGood: Validation of risk prediction models [Internet]. 2014 [cited 2021 Aug 5]. Available from: https://CRAN.R-project.org/package=ModelGood
- 28. Pinkawa M, Gagel B, Piroth MD, Fischedick K, Asadpour B, Kehl M, et al. Erectile dysfunction after external beam radiotherapy for prostate cancer. Eur Urol. 2009 Jan;55(1):227–34. pmid:18375048
- 29. van de Poll-Franse LV, Kwan L, Reiter RE, Lee SP, Litwin MS. The influence of cardiovascular disease on health related quality of life in men with prostate cancer: a 4-year followup study. J Urol. 2008 Apr;179(4):1362–7; discussion 1367. pmid:18289562
- 30. Development and Internal Validation of a Web-based Tool to Predict Sexual, Urinary, and Bowel Function Longitudinally After Radiation Therapy, Surgery, or Observation. Eur Urol. 2020 Aug 1;78(2):248–55. pmid:32098731
- 31. Oh JH, Kerns S, Ostrer H, Powell SN, Rosenstein B, Deasy JO. Computational methods using genome-wide association studies to predict radiotherapy complications and to identify correlative molecular processes. Sci Rep. 2017 Feb 24;7:43381. pmid:28233873
- 32. Haskins AE, Han PK, Lucas FL, Bristol I, Hansen M. Development of clinical models for predicting erectile function after localized prostate cancer treatment. Int J Urol. 2014;21(12):1227–33. pmid:25056284
- 33. Alemozaffar M, Regan MM, Cooperberg MR, Wei JT, Michalski JM, Sandler HM, et al. Prediction of Erectile Function Following Treatment for Prostate Cancer. JAMA J Am Med Assoc [Internet]. 2011 Sep 21 [cited 2021 Feb 15];306(11). Available from: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3831607/ pmid:21934053
- 34. Amann J, Blasimme A, Vayena E, Frey D, Madai VI, the Precise4Q consortium. Explainability for artificial intelligence in healthcare: a multidisciplinary perspective. BMC Med Inform Decis Mak. 2020 Nov 30;20(1):310. pmid:33256715
- 35. Merrick GS, Wallner KE, Butler WM. Management of sexual dysfunction after prostate brachytherapy. Oncol Williston Park N. 2003 Jan;17(1):52–62; discussion 62, 67–70, 73. pmid:12599931
- 36. Hyun JS. Prostate cancer and sexual function. World J Mens Health. 2012 Aug;30(2):99–107. pmid:23596596
- 37.
Erectile Dysfunction After Prostate Cancer [Internet]. 2019 [cited 2022 Oct 18]. Available from: https://www.hopkinsmedicine.org/health/conditions-and-diseases/prostate-cancer/erectile-dysfunction-after-prostate-cance
- 38. Taitt HE. Global Trends and Prostate Cancer: A Review of Incidence, Detection, and Mortality as Influenced by Race, Ethnicity, and Geographic Location. Am J Mens Health. 2018 Nov;12(6):1807–23. pmid:30203706