Introduction: As most patients with prostate cancer die from a different cause, current guidelines recommend treatment only for patients expected to live more than 10 years. Subsequently, estimating life expectancy has become an important component of prostate cancer management. Existing tools, however, have been developed from limited datasets and offer only modest predictive ability. We investigated whether augmenting current tools with patient-reported or claims-based health measures improves predictive accuracy.
Methods: From the SEER-CAHPS database, which links cancer information, Medicare claims, and patient surveys, we selected men 65 years and older diagnosed with prostate cancer from 2004 to 2013. We then identified three existing tools that estimate life expectancy: two based on claims data (CHO/Hawken and Daskivich models) and the Hoffman model based on patient-reported data. We evaluated each model’s performance in this cohort and assessed the incremental value of adding different data (i.e., patient-reported data to claims-based models and vice versa) to predict other-cause mortality, using competing risk regression.
Results: Among 3,240 men diagnosed with prostate cancer and a median follow-up of 4.25 years, 246 (7.62%) died of prostate cancer and 631 (19.48%) died of other causes. As highlighted by the Table, the three tools performed similarly well, with 10-year time-dependent AUCs ranging from 0.738 to 0.783. Though modest, the addition of different data types improved model performance. Specifically, the addition of claims-based health measures to the Hoffman model improved the AUC from 0.738 to 0.744. The addition of patient-reported health measures to the claims-based Cho/Hawken and Daskivich models improved the AUCs from 0.749 to 0.753 and 0.783 to 0.796, respectively.
Conclusions: Current life expectancy models for prostate cancer performed at levels similar to their initial development studies, serving as further validation. Enriching these models with other data types resulted in incremental improvements, suggesting that prediction can improve modestly with increased data capture. Additional research is required to assess what combination of inputs provides the optimal balance of model performance, feasibility of data collection, and meaningfulness. Source of
Funding: Brooke Namboodri Spratte was supported by Summer Medical Student Fellowship Program sponsored by the American Urological Association through the support of the Herbert Brendler, MD, Research Fund.
Hung-Jui Tan, MD, MSHPM was supported by a Mentored Research Scholar Grant in Applied and Clinical Research, MRSG-18-193-01-CPPB, from the American Cancer Society as well as the NIH Loan Repayment Program.