# Diagnostic Performance of Prostate Cancer Disease‐Specific Phenotypes Identified Using Real‐World Databases: A Systematic Review

## TL;DR

This paper reviews how well different prostate cancer phenotypes can be identified using real-world data, showing that some methods are accurate but more validation is needed.

## Contribution

The study systematically summarizes the diagnostic accuracy of prostate cancer phenotypes using real-world databases.

## Key findings

- Prediction models for metastasis and bone metastasis show high accuracy with AUC > 0.7 and specificity above 90%.
- Claims-based algorithms for biochemical recurrence have low sensitivity but high specificity.
- Further validation is needed for phenotypes as therapeutic options for prostate cancer evolve.

## Abstract

Research using real‐world databases (RWD) often requires the development of computable phenotypes based on clinical reasoning‐based algorithms or prediction models with validation through a reference standard such as chart review. While there are studies reporting different phenotypes for key prostate cancer (PC) disease or outcomes, these have not been summarized systematically.

To conduct a systematic review (SR) to summarize validation statistics on PC‐specific phenotypes, including metastasis, biochemical recurrence (BCR), castration‐resistant prostate cancer (CRPC), hormone‐sensitive prostate cancer (HSPC), progression‐free survival, and performance status.

We conducted a SR in accordance with the Preferred Reporting Items for Systematic Reviews and Meta‐Analysis of Diagnostic Test Accuracy Studies guidelines. We systematically searched PubMed/Medline and EMBASE for studies reporting algorithms and prediction models for PC phenotypes based on structured RWD published between 2012 and 2024. A summary of algorithms and prediction models, along with their respective estimates of diagnostic accuracy compared to reference standards and/or measures of uncertainty, was provided. An area under the curve (AUC) > 0.7 was considered an acceptable phenotype.

Out of 7427 retrieved citations, 29 unique retrospective studies (31 citations) were included. Both claims‐based codes and prediction model‐based classification for any metastasis and bone metastases had an acceptable performance with high AUC (0.88 and > 0.7, respectively) and high specificity (above 90%) with a few having moderate sensitivity (60% to 100%). The prediction model‐based BCR classification had acceptable performance (AUC > 0.7); however, claims‐based BCR had moderate performance statistics with sensitivity in the range of 3%–19% and specificity in the range of 83%–98%. One claims‐based algorithm for metastatic CRPC had high sensitivity (77%) and specificity (100%). Studies for mHSPC were based on clinical reasoning without assessing their diagnostic accuracy. Claims‐based algorithms for performance status had at least 75% sensitivity and relatively high specificity.

Our SR highlights the acceptable accuracy of computable phenotypes for PC, including (bone) metastasis, BCR, and performance status within RWD. Further validation studies are needed for RWD‐based phenotypes to account for changes in therapeutic options in PC.