Predicting prostate cancer-specific mortality using SEER

Non-metastatic prostate cancer presents unique challenges in medical decision making because of the generally favourable oncological prognosis—more than 98% 10-year cancer-specific survival—counterbalanced by the morbidity of treatment,1American Cancer SocietyCancer facts and figures 2019.https://www.cancer.org/research/cancer-facts-statistics/all-cancer-facts-figures/cancer-facts-figures-2019.htmlDate: 2019Date accessed: January 17, 2021Google Scholar including urinary incontinence and sexual dysfunction. As a result, the accepted standard of care ranges from active surveillance (ie, careful monitoring) to multimodal therapy that can include a combination of surgery, radiation, and hormonal therapy. Patients with non-metastatic prostate cancer face the difficult task of weighing short-term and possibly long-term side-effects against the risk of metastases and death within 10–20 years. Thus, a shared decision making model between clinicians and patients is widely advocated,2Sanda MG Cadeddu JA Kirkby E et al.Clinically localized prostate cancer: AUA/ASTRO/SUO guideline. Part I: risk stratification, shared decision making, and care options.J Urol. 2018; 199: 683-690Crossref PubMed Scopus (377) Google Scholar and requires high quality prognostic tools to appropriately inform such complex decisions. Changhee Lee and colleagues3Lee C Light A Alaa A Thurtle D van der Schaar M Gnanapragasam VJ Application of a novel machine learning framework for predicting non-metastatic prostate cancer-specific mortality in men using the Surveillance, Epidemiology, and End Results (SEER) database.Lancet Digit Health. 2021; (published online Feb 3.)https://doi.org/10.1016/S2589-7500(20)30314-9Summary Full Text Full Text PDF PubMed Scopus (6) Google Scholar present an interesting application of the machine learning framework, Survival Quilts, to predict 10-year non-metastatic prostate cancer-specific mortality using the Surveillance, Epidemiology, and End Results (SEER) database. Survival Quilts combines multiple known survival models, such that each existing model is given greater weight at time intervals in which the model provides better incremental performance.4Lee C Zame WR Alaa AM van der Schaar M Temporal quilting for survival analysis.Proc Mach Learn Res. 2019; 89: 596-605Google Scholar Survival Quilts was used in this way to generate individual risk predictions in a variable agnostic manner. Lee and colleagues3Lee C Light A Alaa A Thurtle D van der Schaar M Gnanapragasam VJ Application of a novel machine learning framework for predicting non-metastatic prostate cancer-specific mortality in men using the Surveillance, Epidemiology, and End Results (SEER) database.Lancet Digit Health. 2021; (published online Feb 3.)https://doi.org/10.1016/S2589-7500(20)30314-9Summary Full Text Full Text PDF PubMed Scopus (6) Google Scholar showed good discrimination (c-index 0·829, 95% CI 0·820–0·838) for 10-year prostate cancer mortality, which was similar to the traditional multivariable models (PREDICT Prostate c-index 0·820, 0·811–0·829; Memorial Sloan Kettering Cancer Center [MSKCC] nomogram c-index 0·787, 0·776–0·798). Additionally, decision curve analysis showed that Survival Quilts provides a small incremental net benefit over the PREDICT and MSKCC models for treatment decision making at moderate (0·1 and 0·3) threshold probabilities of risk for 10-year cancer-specific mortality. The results of this analysis are thought provoking as urological oncologists strive to provide personalised information to guide treatment for patients facing complex clinical decisions. Lee and colleagues3Lee C Light A Alaa A Thurtle D van der Schaar M Gnanapragasam VJ Application of a novel machine learning framework for predicting non-metastatic prostate cancer-specific mortality in men using the Surveillance, Epidemiology, and End Results (SEER) database.Lancet Digit Health. 2021; (published online Feb 3.)https://doi.org/10.1016/S2589-7500(20)30314-9Summary Full Text Full Text PDF PubMed Scopus (6) Google Scholar suggest that a shift from tier-based groups to a more nuanced calculation of risk might be better for informing clinical decision making. They also note that a machine learning framework, such as Survival Quilts, can more easily integrate new input variables as soon as they become available. As the field dives deeper into genomics, advanced imaging, and novel treatment techniques, the amount of data to integrate when making a clinical assessment are likely to continue to increase. Yet, there are clear limitations to such a strategy. Any model is only as accurate as the data from which it is built. One possible reason for why Survival Quilts did not considerably outperform traditional clinical models in this analysis is that the input variables were mostly the same for both models (eg, age, prostate-specific antigen, Gleason grade, and clinical stage). Although Lee and colleagues3Lee C Light A Alaa A Thurtle D van der Schaar M Gnanapragasam VJ Application of a novel machine learning framework for predicting non-metastatic prostate cancer-specific mortality in men using the Surveillance, Epidemiology, and End Results (SEER) database.Lancet Digit Health. 2021; (published online Feb 3.)https://doi.org/10.1016/S2589-7500(20)30314-9Summary Full Text Full Text PDF PubMed Scopus (6) Google Scholar contend that machine learning frameworks can more quickly process new variables as they become available, clinical endpoints associated with these new variables are also required. For example, to include prostate MRI as a variable in Survival Quilts, a dataset of patients with prostate MRI information is needed and their corresponding 10–20 year survival outcomes. In these ways, machine learning is still bound by the same considerations as traditional clinical modelling. Additionally, specific limitations for this study3Lee C Light A Alaa A Thurtle D van der Schaar M Gnanapragasam VJ Application of a novel machine learning framework for predicting non-metastatic prostate cancer-specific mortality in men using the Surveillance, Epidemiology, and End Results (SEER) database.Lancet Digit Health. 2021; (published online Feb 3.)https://doi.org/10.1016/S2589-7500(20)30314-9Summary Full Text Full Text PDF PubMed Scopus (6) Google Scholar include the reliance on an observational dataset with a large amount of missing data and limited follow-up duration (10 years for non-metastatic prostate cancer survival). Agreement between population-based studies and randomised clinical trials has been questioned previously,5Soni PD Hartman HE Dess RT et al.Comparison of population-based observational studies with randomized trials in oncology.J Clin Oncol. 2019; 37: 1209-1216Crossref PubMed Scopus (81) Google Scholar, 6Nordon C Karcher H Groenwold RH et al.The “efficacy-effectiveness gap”: historical background and current conceptualization.Value Health. 2016; 19: 75-81Summary Full Text Full Text PDF PubMed Scopus (67) Google Scholar and results derived from large datasets such as SEER need to be appropriately contextualised. Advancing technology provides more data and risk estimates with increasingly granular detail, but how it impacts patients cannot be forgotten. Previous work7Nayak JG Scalzo N Chu A et al.The development and comparative effectiveness of a patient-centered prostate biopsy report: a prospective, randomized study.Prostate Cancer Prostatic Dis. 2020; 23: 144-150Crossref PubMed Scopus (4) Google Scholar has shown that patient preference for narrative description and risk classification when understanding a prostate cancer diagnosis, and moving from a tier-based risk grouping system (eg, low, intermediate, high) to a continuous measurement (ie, percentage risk) might not be beneficial to all patients. As this group has previously summarised,8Gnanapragasam VJ Informing informed decision-making in primary prostate cancer treatment selection.BJU Int. 2020; 125: 194-196Crossref PubMed Scopus (1) Google Scholar despite a plethora of currently available decision making tools there remains a high degree of uncertainty and most patients still rely on health-care providers to guide their decision making. Because probability and evaluating individual risk is notoriously challenging to understand,9Spiegelhalter D Gage J What can education learn from real-world communication of risk and uncertainty?.The Mathematics Enthusiast. 2015; 12: 4-10Google Scholar, 10Binder K Krauss S Bruckmaier G Marienhagen J Visualizing the Bayesian 2-test case: the effect of tree diagrams on medical decision making.PLoS One. 2018; 13e0195029Crossref PubMed Scopus (13) Google Scholar predictive models need to not only be accurate, but also easily interpretable by clinicians and patients to become functional. Survival Quilts and other continuous predictive tools should be paired with a user friendly interface and shown to be meaningful to prostate cancer patients before use in a clinical setting can be recommended. Although there is no right answer for how best to use the growing armamentarium of tools, the goal of research—to inform everyday discussions between patients and their health-care providers—needs to be kept in mind. We declare no competing interests. Application of a novel machine learning framework for predicting non-metastatic prostate cancer-specific mortality in men using the Surveillance, Epidemiology, and End Results (SEER) databaseA novel machine learning-based approach produced a prognostic model, Survival Quilts, with discrimination for 10-year prostate cancer-specific mortality similar to the top-ranked prognostic models, using only standard clinicopathological variables. Future integration of additional data will likely improve model performance and accuracy for personalised prognostics. Full-Text PDF Open Access