Abstract
Background and objective:
To describe, via latent variable mixture modeling, distinct post-radical prostatectomy (RP) patient-reported outcome (PRO) recovery profiles, which are positioned to complement currently disseminated statistical averages for shared decision-making.
Patients and Methods:
Patients undergoing RP and completing the 26-item Expanded Prostate Cancer Index Composite 12 months after surgery were identified from the Michigan Urological Surgery Improvement. Hierarchical cluster analysis and latent variable mixture modelling was applied to urinary incontinence (UI) and sexual function (SF) recovery scores, and final models chosen based on optimal performance.
Results:
3,956 patients comprised the study cohort. Three distinct UI profiles were identified with prevalence of 49%, 37% and 14% from best to worst recovery, respectively. Four distinct SF profiles were identified with prevalence of 14%, 24%, 42%, and 20%, from best to worst recovery, respectively. The last two SF profiles had similar function scores but differed based on perception of function being bothersome. Limitations include incomplete PRO capture, which may introduce bias.
Conclusions:
We identify distinct UI and SF recovery profiles and their prevalence from a large, prospectively maintained registry, potentially improving interpretability of PRO data for decision making.
Keywords: localized prostate cancer, radical prostatectomy, patient reported outcomes, post-prostatectomy stress urinary incontinence, post-prostatectomy sexual dysfunction, functional outcomes, functional recovery
Introduction:
Treatment of prostate cancer has the potential to negatively impact patients’ urinary, sexual, bowel, hormonal function, as well as global quality of life1. Understanding functional recovery following treatment for prostate cancer is important for patient decision making given that most men survive at least 15 years following treatment2. To that end, large prospective series of patient reported outcomes (PROs) have been published, and their results have been applied for patient counseling 1,3,4. These studies have undoubtedly provided information that can improve the quality of decision making for localized prostate cancer. However, such works report recovery as a population mean; therefore, it can be challenging to set expectations for patients whose baseline or first set of post-operative functional outcomes are significantly better or worse than the reported average. This shortcoming is compounded by the fact that guidelines on post-operative functional recovery are not explicit about how to use existing data to counsel patients5. While it cannot be assumed that most patients precisely follow the population mean, limited data exist on how to characterize differences in functional recovery or how to translate this notion into actionable advice.
In the present study, we sought to better characterize urinary function and sexual function (SF) recovery for men undergoing radical prostatectomy (RP) in a large, diverse, cohort with prospectively collected patient reported outcomes (PROs) utilizing hierarchical cluster analysis and latent variable mixture modeling. These profiles have the potential to better inform patients about the range of possible recovery, not just at individual points in time, but longitudinally over their entire course.
Patients and methods:
Patient population
The Michigan Urological Surgery Improvement Collaborative (MUSIC) is a statewide quality improvement collaborative consisting of more than 250 urologists and 46 diverse practices. MUSIC PRO is a subprogram within MUSIC, in which practices can choose if they wish to participate and enroll their patients. Eighty-nine urologists representing 32 sites participate in PRO collection. Each practice obtained exemption or approval from the Institutional Review Board (IRB) prior to participation, and this analysis was deemed exempt from review by the University of Michigan IRB. The study cohort included men in the MUSIC prostate cancer registry who underwent RP from 2016 to 2021 with available PRO data.
Within MUSIC, PROs are captured before surgery and at 3, 6, 12, and 24 months post-operatively. Patients completed surveys measuring domain scores on the 26-item Expanded Prostate Cancer Index Composite (EPIC-26), a validated instrument for disease-specific function6. Distinct domains within EPIC-26 include urinary incontinence (UI), sexual function (SF), urinary irritative symptoms, bowel function, and hormonal symptoms; each domain is then scaled to a value of 0–100, with higher scores representing better function and quality of life. Patients received the EPIC-26 questionnaire via mail or e-mail, with reminders sent to non-responders; patients may “opt out” from PRO collection at any point. Those not responding to two consecutive questionnaires were not sent further correspondence. Inclusion criteria for this study entailed patients with who completed a PRO at 12 months, plus or minus 60 days, following surgery. Patients with missing PROs at any other timepoint were included. Completing a PRO was defined as answering all four UI questions and five of six SF questions from EPIC-26.
Study endpoints
The primary study endpoints were pre-operative and post-operative PROs. Our analysis focused on UI and SF as these are most relevant for pre-RP patient counseling. UI and SF related questions are listed in Supplement.
Statistical Analysis
Demographic and clinicopathologic factors, as well as baseline and post-operative PROs were summarized as counts and proportions or medians with interquartile range (IQR). Two patient-centered (rather than variable-centered) analytic techniques, one descriptive (hierarchical clustering) and one inferential (latent variable mixture modeling) were applied to the data. Supplementary Figures 1–3 provide a conceptual overview of the analytical approaches used in our analysis.
Hierarchical cluster analysis
Heatmaps were created to visualize the breadth of possible post-operative PRO responses by question. Combined with hierarchical clustering, the heatmap visualization was used to examine the dimensionality of the UI and SF domains. An uncentered correlation distance metric and an average linkage were used in the cluster analysis. Data were standardized prior to clustering but visualized in the heatmap in the unstandardized scale. ComplexHeatmap in R was used to fit and visualize two separate heatmaps for UI and SF 7. Each heatmap treated analyzed items associated with UI and SF separately across the five time points noted previously.
Latent profile analysis
One challenge in using clustering as a method for identifying distinctive profiles is that there are few widely accepted metrics for assessing the quality of the clusters (i.e., number and fit) and every case is assigned a single cluster regardless of fit 8. Latent mixture models, on the other hand, provide a probability of a case being a part of any cluster and there are multiple, widely accepted metrics for identifying an adequate number of clusters.
We therefore performed a longitudinal latent variable mixture model based on insights gained from the hierarchical cluster analysis. Longitudinal latent variable mixture models can be used to identify latent subgroups based on continuous indicator variables. There are multiple approaches to identifying latent subgroups using longitudinal data (e.g., latent transient analysis, longitudinal latent class growth curve analysis, growth mixture modeling, and hierarchical latent longitudinal growth curve 9). We used a longitudinal, repeated measures approach to allow for change across multiple variables at the same time point in a flexible and parsimonious way. PRO data from questions in UI and SF domains at all measurements (including preoperative) were used to identify profiles. Latent profiles can be formed based on different constraints applied to variances across variables between profiles the degree to which variables are allowed to covary within profiles 10 11. Four distinct combinations of variances and covariances were estimated, with each combination run for models with 1 to 6 possible profiles/classes, using TidyLPA, MClust, TidySEM, and OpenMX R packages10 12. Full Information Maximum Likelihood was used to account for missing PROs across pre-operative, 3-month, 6-month, and 24-month time points. Standard latent variable mixture modeling evaluation metrics were used to identify the final number of profiles per domain: Bayesian information criteria (BIC), entropy, and mean minimum class probability.
For SF and based on patterns observed in the cluster analysis, we modeled two variables for each timepoint: one that represented the mean of five items that addressed functional outcomes and the other for whether sexual function was perceived as problematic by patients. For UI, we modeled the mean across all items for each timepoint.
Results:
Of 6,843 patients contributing any PRO data during the study period, 3,956 had a PRO questionnaire available at 12 months post-surgery. Of these patients, 3,159 had a pre-operative questionnaire and 3,228 had a 3-month, 3,194 had a 6-month, and 2,462 had a 24-month questionnaire. Baseline characteristics of the study cohort are shown in Table 1.
Table 1.
Baseline demographic and disease characteristics study cohort (N = 3,956).
| Median (IQR) or N (%) | |
|---|---|
| Age, years | 65 (60, 69) |
| BMI, kg/m 2 | 28.7 (26.1, 32.2) |
| Prostate volume, mL | 37 (28, 49) |
| PSA, ng/mL | 6.4 (4.8, 9.3) |
| Prostatectomy Grade Group | |
| 1 | 508 (15%) |
| 2 | 1,555 (45%) |
| 3 | 786 (23%) |
| 4 | 386 (11%) |
| 5 | 245 (7%) |
| Nerve Sparing | |
| Bilateral | 2,366 (68%) |
| Unilateral | 477 (14%) |
| None | 651 (19%) |
| Race | |
| African American | 291 (8%) |
| Caucasian | 3,112 (90%) |
| Other | 65 (2%) |
| Relationship Status | |
| Married | 3,260 (87%) |
| Other | 488 (13%) |
Data not available for BMI (n = 808), prostate volume (n = 1324), PSA level (n= 845), RP Grade Group (n = 476), nerve-sparing status (n = 462), race (n = 488), and relationship status (n = 208).
Hierarchical cluster analysis of scores on questions related to UI and SF both demonstrated significant variability in post-operative recovery. Narrative descriptions and heatmaps (Supplementary Figure 5A and B) can be found in the Supplement. In exploring the data, we noticed an uncoupling of SF function and bother scores amongst a subset of patients, prompting us to pursue profile modelling that treated bother as a separate item (rather than averaged with other SF domain items).
Final latent profile models were informed by the cluster analysis whereby UI was characterized by the overall score across five time points but the SF domain was characterized using the mean of the first five questions and Q68 (separate from the mean). The final number of profiles were selected based on BIC, entropy, and min probability (Supplementary Table 1). Three UI profiles were identified, with individual prevalence of 49% (UI Profile 1—good recovery), 37% (UI Profile 2—moderate recovery), and 14% (UI Profile 3—poor recovery) in order from best to worst, respectively (Figure 1). The results highlight several observations around recovery in the UI domain that may not be represented in population mean profiles. For example, there is a larger group of patients (49%) with normal baseline scores, who worsen only marginally after surgery and subsequently recover well (Figure 1, Profile 1). In addition to good and poor recover patients is a sizeable cohort (37%) with normal baseline scores that decline substantially after surgery, and recover significantly but not completely (Figure 1, Profile 2).
Figure 1. Modelled UI PRO profiles.
Results of repeated measures latent profile models for urinary incontinence domain score over time. Each line represents a different recovery profile as suggested by model output, with individual prevalence of 48% (UI Profile 1—good recovery), 37% (UI Profile 2—moderate recovery), and 14% (UI Profile 1—poor recovery) from best to worst, respectively. Plus signs (+) represent overall population average.
Four SF recovery profiles were identified, with individual prevalence of 14% (SF Profile 4—“good recovery”), 24% (SF Profile 2—“moderate recovery”), 42% (SF Profile 1—“poor recovery, bothered”), and 20% (SF Profile 3—“poor recovery, not bothered”), in order from best to worst recovery (Figure 2). Two SF profiles with the worst long-term recovery (1 and 3) are distinguished by different SF bother scores, which is consistent with heatmap observations showing different clusters with poor SF and either high or low bother. While the population-wide decline in SF is more dramatic than in UI, there exists a small cadre of patients (14%) with excellent baseline function that is mostly preserved post-operatively (Figure 2, SF Profile 4). A larger subset of patients (24%) with mid-range pre-operative SF experiences a significant post-operative decline followed by variable recovery (Figure 2, SF Profile 2). Finally, there are two large groups of patients with relatively poor baseline function whose scores remain low after surgery.
Figure 2. Modelled sexual function PRO profiles.
Results of repeated measures latent profile models for sexual function domain score and patient “bother with sexual function” item over time. Each line represents a different recover profile as suggested by model output, with individual prevalence of 14% (SF Profile 4—“good recovery”), 24% (SF Profile 2—“moderate recovery”), 42% (SF Profile 1—“poor recovery, bothered), and 20% (SF Profile 3—“poor recovery, not bothered), from best to worst recovery. Plus signs (+) represent overall population average.
Tables 2 and 3 compare clinicopathologic variables across UI and SF profiles, respectively, based on patients’ most probable profile. Probability of class membership is shown in Supplementary Figure 4. Small differences were seen with generally more favorable disease characteristics, younger age, and more complete nerve sparing aligned with better recovery profiles.
Table 2.
Most likely urinary incontinence profile and clinicopathologic variables
| Profile 1 | Profile 2 | Profile 3 | |
|---|---|---|---|
| Age, years | 63.9 (59.0, 68.3) | 65.5 (60.8, 69.7) | 66.7 (61.5, 70.4) |
| PSA, ng/dL | 6.3 (4.7, 9.0) | 6.4 (4.8, 9.4) | 6.6 (5.0, 10.3) |
| Prostatectomy Grade Group | |||
| 1 | 17% | 13% | 10% |
| 2 | 48% | 41% | 43% |
| 3 | 21% | 26% | 21% |
| 4 | 9% | 13% | 15% |
| 5 | 6% | 7% | 11% |
| Nerve Sparing | |||
| Bilateral | 71% | 65% | 63% |
| Unilateral | 12% | 15% | 15% |
| None | 17% | 20% | 22% |
| Relationship Status | |||
| Married | 89% | 86% | 84% |
| Other | 11% | 14% | 16% |
Table 3.
Most likely sexual function profile and clinicopathologic variables
| Profile 1 | Profile 2 | Profile 3 | Profile 4 | |
|---|---|---|---|---|
| Age, years | 65.4 (60.8, 69.7) | 63.4 (58.3, 67.5) | 68.4 (64.1, 71.6) | 60.7 (56.3, 65.6) |
| PSA, ng/dL | 6.51 (4.9, 9.6) | 6.1 (4.6, 9.1) | 6.4 (4.8, 9.7) | 6.2 (4.7, 8.8) |
| Prostatectomy Grade Group | ||||
| 1 | 13% | 17% | 12% | 21% |
| 2 | 42% | 50% | 38% | 53% |
| 3 | 23% | 23% | 25% | 18% |
| 4 | 13% | 7% | 15% | 5% |
| 5 | 9% | 4% | 10% | 2% |
| Nerve Sparing | ||||
| Bilateral | 62% | 77% | 55% | 86% |
| Unilateral | 15% | 12% | 16% | 9% |
| None | 23% | 11% | 29% | 5% |
| Relationship Status | ||||
| Married | 86% | 87% | 88% | 89% |
| Other | 14% | 13% | 12% | 11% |
Discussion:
Recovery of urinary and sexual function after RP is highly variable from patient to patient. Therefore, reporting outcomes as a population mean can be sub-optimal in setting individual patient expectations. Our results demonstrate a novel set of post-RP functional recovery profiles using one of the largest reported PRO series to date. Identifying distinct recovery profiles can complement existing pre-operative counseling tools by moving beyond conceiving of recovery in relation to a single average but instead as multiple, possible profiles and their respective frequency. While our approach surfaced multiple distinct recovery profiles across UI and SF domains, it is important to note that they still describe a population and cannot be used as an individual patient prediction model, unlike other tools13.
Our results are consistent with and elaborate upon prior studies of similar patient populations. For example, the ProtecT (Prostate Testing for Cancer and Treatment) and CESAR (Comparative Effectiveness Analysis of Surgery and Radiation) studies report post-treatment PROs over time in the context of a randomized trial and prospective cohort, respectively, and overall suggest similar functional recovery to that observed in our registry3,4.
To illustrate the importance of our distinct analytic approach, consider a patient who deviates from the average recovery pattern beginning three months after surgery. For example, a patient with low UI scores at three months may be encouraged by the significant number of patients who approach baseline function after an early post-operative decline. Our analysis identified that approximately 14% of patients who had low UI scores at three months did not return to baseline continence. This is consistent with guidance around the consideration of surgery for incontinence at as early as six months post-RP, which recognizes that further recovery after this time point is marginal5. Therefore, patients falling into “UI Profile 3 —poor recovery” could be counseled on the low likelihood of recovery, recognizing that continued observation rather than timely referral may prolong negative outcomes unnecessarily. Similarly, patients in “SF Profile 1—poor recovery, bothered” could be counseled towards escalation of erectile function aids far earlier than the conventional timeframe of 1–2 years. The two SF profiles with the worst long-term outcomes were separated by a questionnaire item around degree of “bother.” This suggests that patients with poor function pre-RP who are ambivalent about sexual function tend to remain so after surgery (Figure 2, “SF Profile 3—poor recovery, not bothered”). This contrasts with a group of men with slightly higher baseline function who were indeed bothered by worsening SF scores (Figure 2A, “SF Profile 1—poor recovery, bothered”). Emphasizing the pre-operative identification of men with high likelihood of SF decline that would be bothered by a significant decline could improve patient consultation and treatment related decision-making. Conversely, patients in “SF Profile 4—poor recovery, not bothered” may receive more streamlined counseling.
Finally, our work builds upon previously reported novel approaches to PRO data for shared decision-making purposes. Martin, et al., used single institution PROs and a Bayesian hierarchical approach to demonstrate “best-case/worst-case” trajectories after RP14. While these are important considerations, our analysis augments this work and represents a step forward in the interpretability of information delivered to patients. Our analysis shows that indeed many patients fit into neither a best-case nor a worst-case trajectory, nor the population mean. Marzorati, et al., used a latent class growth analysis approach, describing recovery profiles, but represent a significantly smaller sample from a single European center of excellence and a smaller number of surgeons with presumably more homogenous experience and case volume15. In addition, Marzorati et al., did not include preoperative PROs in identifying their classes nor did they explore the potential for distinct functional and bother responses in either the UI or SF domains.
Our work should be considered within the context of its limitations. First, our work is retrospective and observational, and therefore subject to known (e.g., survey completion as a function of satisfaction or dissatisfaction with functional outcomes) and unknown biases. The modelling approaches used herein are not designed to make individualized predictions for a given patient, and it is important in using such a tool for counselling that the profiles, viewed together, describe a population. Individualized trajectory prediction is beyond the scope of the present work. Additionally, the identified profiles were not externally validated and may represent practice patterns specific to MUSIC; that said, the number of different providers and diversity of practice types suggest some degree of generalizability of our findings. Relatedly, the modelling approach will generate profiles regardless of “goodness of fit;” we have mitigated this by comparing several different models and selecting the best performing, though again, these should be externally validated 16. Additionally, some PRO capture is incomplete, which may alter prevalence estimates of phenotypes. Data on symptomatic treatment (e.g., phosphodiesterase inhibitors, intra-corporal injection, etc.) are coarse and subject to limitations of patient reporting and were thus not integrated; untreated profiles may be lower globally or have higher a prevalence of the “poor recovery” trajectory.
Despite these limitations, our results hold the potential to complement existing pre-operative patient counseling and expectation setting tools by demonstrating realistic recovery profiles and their relative prevalence in a large patient population. Further work should externally validate these phenotypes and explore the extent to which this type of information improves real-world patient and provider perception of the quality of pre- and post-RP counseling.
Supplementary Material
Acknowledgements
PL is supported by the NCI (T32 CA180984). Support for MUSIC is provided by Blue Cross Blue Shield of Michigan as part of the BCBSM Value Partnerships program. The opinions, beliefs and viewpoints expressed by the authors do not necessarily reflect those of BCBSM or any of its employees. The authors acknowledge the significant contributions of the clinical champions, urologists, administrators, and data abstractors in each participating MUSIC practice (details can be found at www.musicurology.com), as well as members of the MUSIC Coordinating Center at the University of Michigan.
Footnotes
conflicts of interest: No conflicts of interest.
References
- 1.Donovan JL, Hamdy FC, Lane JA, et al. Patient-Reported Outcomes after Monitoring, Surgery, or Radiotherapy for Prostate Cancer. N Engl J Med 2016;375(15):1425–1437. doi: 10.1056/NEJMoa1606221 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Hamdy FC, Donovan JL, Lane JA, et al. Fifteen-Year Outcomes after Monitoring, Surgery, or Radiotherapy for Prostate Cancer. N Engl J Med 2023;388(17):1547–1558. doi: 10.1056/NEJMoa2214122 [DOI] [PubMed] [Google Scholar]
- 3.Al Hussein Al Awamlh B, Wallis CJD, Penson DF, et al. Functional Outcomes After Localized Prostate Cancer Treatment. JAMA 2024;331(4):302–317. doi: 10.1001/jama.2023.26491 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Hamdy FC, Donovan JL. Patient-Reported Outcomes Following Treatment for Localized Prostate Cancer: Helping Decision Making for Patients and Their Physicians. JAMA 2017;317(11):1121–1123. doi: 10.1001/jama.2017.1703 [DOI] [PubMed] [Google Scholar]
- 5.Sandhu JS, Breyer B, Comiter C, et al. Incontinence after Prostate Treatment: AUA/SUFU Guideline. J Urol 2019;202(2):369–378. doi: 10.1097/JU.0000000000000314 [DOI] [PubMed] [Google Scholar]
- 6.Szymanski KM, Wei JT, Dunn RL, Sanda MG. Development and Validation of an Abbreviated Version of the Expanded Prostate Cancer Index Composite Instrument for Measuring Health-related Quality of Life Among Prostate Cancer Survivors. Urology. 2010;76(5):1245–1250. doi: 10.1016/j.urology.2010.01.027 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Gu Z, Eils R, Schlesner M. Complex heatmaps reveal patterns and correlations in multidimensional genomic data. Bioinformatics. 2016;32(18):2847–2849. doi: 10.1093/bioinformatics/btw313 [DOI] [PubMed] [Google Scholar]
- 8.Hennig C, Meila M, Murtagh F, Rocci R, eds. Handbook of Cluster Analysis. Chapman and Hall/CRC; 2015. doi: 10.1201/b19706 [DOI] [Google Scholar]
- 9.Herle M, Micali N, Abdulkadir M, et al. Identifying typical trajectories in longitudinal data: modelling strategies and interpretations. Eur J Epidemiol 2020;35(3):205–222. doi: 10.1007/s10654-020-00615-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Pastor DA, Barron KE, Miller BJ, Davis SL. A latent profile analysis of college students’ achievement goal orientation. Contemporary Educational Psychology. 2007;32(1):8–47. doi: 10.1016/j.cedpsych.2006.10.003 [DOI] [Google Scholar]
- 11.Weller BE, Bowen NK, Faubert SJ. Latent Class Analysis: A Guide to Best Practice. Journal of Black Psychology. 2020;46(4):287–311. doi: 10.1177/0095798420930932 [DOI] [Google Scholar]
- 12.Van Lissa CJ, Garnier-Villarreal M, Anadria D. Recommended Practices in Latent Class Analysis Using the Open-Source R-Package tidySEM. Structural Equation Modeling: A Multidisciplinary Journal. 2024;31(3):526–534. doi: 10.1080/10705511.2023.2250920 [DOI] [Google Scholar]
- 13.Agochukwu-Mmonu N, Murali A, Wittmann D, et al. Development and Validation of Dynamic Multivariate Prediction Models of Sexual Function Recovery in Patients with Prostate Cancer Undergoing Radical Prostatectomy: Results from the MUSIC Statewide Collaborative. European Urology Open Science. 2022;40:1–8. doi: 10.1016/j.euros.2022.03.009 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Martin C, Haaland B, Tward AE, et al. Describing the Spectrum of Patient Reported Outcomes after Radical Prostatectomy: Providing Information to Improve Patient Counseling and Shared Decision Making. Journal of Urology. 2019;201(4):751–758. doi: 10.1016/j.juro.2018.10.014 [DOI] [PubMed] [Google Scholar]
- 15.Marzorati C, Monzani D, Mazzocco K, et al. Predicting trajectories of recovery in prostate cancer patients undergone Robot-Assisted Radical Prostatectomy (RARP). PLoS One. 2019;14(4):e0214682. doi: 10.1371/journal.pone.0214682 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.van Smeden M, Naaktgeboren CA, Reitsma JB, Moons KGM, de Groot JAH. Latent class models in diagnostic studies when there is no reference standard--a systematic review. Am J Epidemiol 2014;179(4):423–431. doi: 10.1093/aje/kwt286 [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.