The answer was in our signatures all along

[principles]

FrankTalk has always had a healthy skepticism about how well the literature reflects real implant outcomes. Many of us feel the published numbers don’t match what we read and live here on FrankTalk.

Over the years, our signatures have quietly become a unique dataset of real-world implant journeys. From 15,000 threads in FrankTalk’s Implant subforum, I systematically extracted 1,639 unique signatures and used survival analysis to turn our lived experiences into a data-driven picture of how implants really perform over time.

The result is now published as a paper:
Community-Driven Survival Analysis of Penile Implant Outcomes: A 15-Year Retrospective Study of Online Forum Data

Thanks to everyone whose signature data made this possible.

[indeed]

Very cool man thanks for putting in the effort to do this!

Two suggestions:

- If understand correctly, you extracted signatures only. It would clean up the data a lot if you'd also extract all the comments and then comparte them with the signatures. I've seen quite a few examples where people didn't note their surgeries/revisions in the signatures. Idk if thats possible with limited LLM ressources though.

- I think if we could also analyze all the comments, the sample size could possibly be increased by a lot.

Very usefull information though already! Great work

[principles]

Very cool man thanks for putting in the effort to do this!

Two suggestions:

- If understand correctly, you extracted signatures only. It would clean up the data a lot if you'd also extract all the comments and then comparte them with the signatures. I've seen quite a few examples where people didn't note their surgeries/revisions in the signatures. Idk if thats possible with limited LLM ressources though.

- I think if we could also analyze all the comments, the sample size could possibly be increased by a lot.

Very usefull information though already! Great work

Thanks for the feedback! I appreciate the suggestions, but extracting data from posts (what I thought you meant by “comments”) rather than signatures would increase the complexity exponentially. Signatures are structured medical timelines that users self-maintain, making them parseable at scale. Posts are unstructured conversational text where users might mention their implant date across dozens of comments over years, discuss other people’s implants, use relative time references (“3 months post-op”), and require disambiguation across ~15,000+ threads * ~10 posts each * ~200 tokens * dual-LLM verification = ~60M tokens. When you factor in API costs and man-hours due to the increasing complexity, with no guarantee of improved data quality or statistical validity, it becomes infeasible.
The N (sample size) from signatures is actually already one of the largest to date (most clinical studies have N=100–400), and the methodology was deliberately designed around what’s feasible for a community member with limited resources. That said, who knows, maybe in the future