2

I hope this message finds you well. I am currently working on a research project involving the comparison of Kaplan-Meier curves from clinical trials with different observation periods. The challenge I'm facing is that the trials have varying lengths, with some having Kaplan-Meier curves spanning 48 months and others only 24 months. I am unsure about the best approach for combining these curves, whether to include all available data or align them at a specific time point.

Additionally, I am using the following methodology for extrapolation: I plan to extract coordinates from Kaplan-Meier curves using WebPlotDigitizer and then reconstruct cumulative Kaplan-Meier curves using the IPDfromKM-Shiny method in R. Given this approach, I have the following questions:

1.When combining Kaplan-Meier curves from trials with different lengths and using an extrapolation method, is it more appropriate to include all available data (e.g., 48 months for some trials, 24 months for others) or align them at a specific time point (e.g., 24 months)?

2.What considerations should be taken into account when dealing with Kaplan-Meier curves of varying lengths and using extrapolation to ensure a valid and meaningful comparison?

I appreciate any insights or recommendations you can provide on this matter. If you need additional details or if there are relevant statistical methods to consider, please feel free to share.

Thank you for your time and expertise.

bitmarks
  • 41
  • 1
    Assuming the trials are held in a consistent way, I don't see a problem with combining the data and making a new K-M. Just make sure censoring is accounted for properly. Is there any way just to access the raw data, or do you have to do it from a plot? – Alex J Jan 19 '24 at 02:54
  • Unfortunately, I do not have actual data; instead, I rely on extrapolation from the plot. I have created new Kaplan-Meier (KM) curves, but the issue lies in deciding the appropriate time frame, especially when the trials have different follow-up durations. If one trial has a follow-up period of 50 months and the other 25 months, should I use 25 months of data from both trials or 25 months of data from the first trial and 50 months of data from the second trial when reconstructing the KM curves? – bitmarks Jan 20 '24 at 16:25
  • I feel like I'm misunderstanding something here but... if you have the plot, and you know the sample size, and there also is marking on the plot when censoring occurred, you should be able to reproduce the original data (approximately) of the KM curve. And I don't think there's any issue with combining 48 and 24 month surveys. That being said, I don't do clinical trials, I just do survival analysis in other contexts so I'm not sure if there's certain specific considerations there – Alex J Jan 22 '24 at 22:10

1 Answers1

1

Have you considered reporting combined results for 24-month data, and combined results for 48-month data separately? Recall, only failures contribute to the statistics of survival analysis - censored subjects who don't fail or exit the study early contribute no information. Therefore, if you combine both time lengths, the resulting KM plot will have a strange jump in it at or nearby 24-months, causing a reader to ask: "it looks like there's a major problem with the data near 24-months --> something happened with the data which looks erroneous."

Regarding your second question, you need to ask why some trials followed up subjects for 24 months, and other followed up for 48 months. If everything was the same in the trials (inclusion/exclusion criteria: patient histories, disease duration, duration of prior meds, prior therapies, repeating recurrences, etc.) then there would likely not be a second set of trials that e.g. decided to follow-up for another two years. Two years of additional follow-up is a large difference in follow-up time, so I would think the primary outcome of the 48-month trials would be viewed clinically as another wholly different disease (constellation of symptoms, etc.) These are the sorts of issues you'd face during review of a paper, which would dampen enthusiasm. Hence, why do you need to throw everything in one hopper (one KM plot) anyhow?

Personally, I think the real answer to your question(s) is clinical, and not analytical. Thus, approach a clinician(s) who diagnoses such outcomes and find out why some trials go on for 2 years ("apples"), while other are fours years ("oranges")? The original papers would also likely reveal why follow-up was 2-years vs 4-years.

wjktrs
  • 852