Longitudinal analysis

[YoonhoH]

Dear experts including Joanne,

Hi,

It's inevitable that equipment will change over time in a long follow-up study.
We want to use data for all time points [e.g., baseline, 1st follow-up (scanner 1), 2nd follow-up (scanner 2), 3rd follow-up (scanner 3) ... ]. I already know that the time effect and scanner effect are mixed, then, already quantified together, making it impossible to completely separate them. Nevertheless, for longitudinal analysis, I would do the following process: 1. FreeSurfer-based longitudinal pipeline; 2. Longitudinal combat. I don't know if the time effect is removed with the numerical calculation when the scanner effect is removed due to longitudinal combat, but I want to see structural changes over time and make disease-specific comparisons. Could you have any advice for me?

Thank you

Sincerely,
Yoonho

[jcbeer]

Hi Yoonho,The time effect will be removed with scanner effect if there is perfect confounding.One possible study design for this scenario would have staggering of the time points, so for example maybe all subjects have time-point 1 on scanner 1. Then half of subjects have time point 2 on scanner 1, and half have time point 2 on scanner 2. Then half of subjects have time point 3 on scanner 2 and half have time point 3 on scanner 3, and so on. (Best design will depend on the length of time intervals and other details.)Another consideration is whether there is a control group. Even if their is confounding of scanner and time point, Combat will be able to preserve the differences between the groups at each time point. To preserve the longitudinal changes with a control group, it could be possible to scan some group for which minimal changes are expected between time points at the same time as other study subjects, and use this group to estimate scanner effects, and apply those scanner effect parameters to harmonize the rest of the study cohort.If you have already done the study, maybe you could use a control group approach. Otherwise I’m not sure if it’s possible to use Combat without removing the longitudinal trends.JoanneOn Jul 22, 2024, at 11:20 PM, YoonhoH ***@***.***> wrote: Dear experts including Joanne, Hi, It's inevitable that equipment will change over time in a long follow-up study. We want to use data for all time points [e.g., baseline, 1st follow-up (scanner 1), 2nd follow-up (scanner 2), 3rd follow-up (scanner 3) ... ]. I already know that the time effect and scanner effect are mixed, then, already quantified together, making it impossible to completely separate them. Nevertheless, for longitudinal analysis, I would do the following process: 1. FreeSurfer-based longitudinal pipeline; 2. Longitudinal combat. I don't know if the time effect is removed with the numerical calculation when the scanner effect is removed due to longitudinal combat, but I want to see structural changes over time and make disease-specific comparisons. Do you have any advice for me? Thank you Sincerely, Yoonho —Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[YoonhoH]

Hi Joanne,
I really appreciate your quick reply.
Our follow-up study is a population-based cohort for same time intervals and subjects (there are dropouts during follow-ups). We plan to harmonize the structural estimates (i.e., brain volume and cortical thickness) extracted from data for all time points with different MR scanner and acquisition parameters, and then create a database for analysis.

1. To use longitudinal combat, should the process with a longitudinal pipeline (e.g., FreeSurfer, ANTs, CAT12) be preceded?
2. What does "if there is perfect confounding" mean? We could extract the structural estimates after FreeSurfer-based longitudinal pipeline. The estimates (consisting of vertex) is represented as one value and could not that mean we already have a mix of scanner and time effects? Based on your regression model, it may be possible to separate and remove the scanner effect from the time effect. I was confused about the time effect disappearing along with the scanner effect ("the time effect will be removed with scanner effect).
3. On your first suggestion (study design), I didn't understand your first suggestion well. I think I couldn't apply that to our data, could you tell me about it in detail?
4. On second suggestion, I didn't quiet get it. Could you explain about it in detail?

Thank you very much

Sincerely,
Yoonho

[jcbeer]

Hi Yoonho, 1. To use longitudinal combat, should the process with a longitudinal pipeline (e.g., FreeSurfer, ANTs, CAT12) be preceded? Yes that seems reasonable. 1. *What does "if there is perfect confounding" mean?* Perfect confounding means that all of time point 1 is collected on scanner 1, all of time point 2 is collected on scanner 2, all of time point 3 is collected on scanner 3. In other words, each time point is collected on a different scanner. 1. We could extract the structural estimates after FreeSurfer-based longitudinal pipeline. The estimates (consisting of vertex) is represented as one value and *could not that mean we already have a mix of scanner and time effects?* I am not sure. 2. Based on your regression model, it may be possible to separate and remove the scanner effect from the time effect. I was confused about the time effect disappearing along with the scanner effect ("the time effect will be removed with scanner effect). If each time point is collected on a different scanner, then the harmonization model will have trouble determining whether changes are due to the scanner or due to the time trend. 3. On your first suggestion (study design), I didn't understand your first suggestion well. I think I couldn't apply that to our data, could you tell me about it in detail? The first suggestion is to distribute the timepoints among different scanners as much as that is possible. 4. On second suggestion, I didn't quiet get it. Could you explain about it in detail? The second suggestion is to scan a control group at each timepoint on each scanner. This group would not be expected to have any longitudinal changes in the brain features. This group could be used to estimate the scanner means and variances, and train a ComBat model, which could then be applied to the study group which is expected to show longitudinal changes. Hope that helps. Best of luck with your study! Joanne

…

On Tue, Jul 23, 2024 at 10:46 PM YoonhoH ***@***.***> wrote: Hi Joanne, I really appreciate your quick reply. Our follow-up study is a population-based cohort for same time intervals and subjects (there are dropouts during follow-ups). *We plan to harmonize the structural estimates (i.e., brain volume and cortical thickness) extracted from data for all time points with different MR scanner and acquisition parameters, and then create a database for analysis.* 1. To use longitudinal combat, should the process with a longitudinal pipeline (e.g., FreeSurfer, ANTs, CAT12) be preceded? 2. *What does "if there is perfect confounding" mean?* We could extract the structural estimates after FreeSurfer-based longitudinal pipeline. The estimates (consisting of vertex) is represented as one value and *could not that mean we already have a mix of scanner and time effects?* Based on your regression model, it may be possible to separate and remove the scanner effect from the time effect. I was confused about the time effect disappearing along with the scanner effect ("the time effect will be removed with scanner effect). 3. On your first suggestion (study design), I didn't understand your first suggestion well. I think I couldn't apply that to our data, could you tell me about it in detail? 4. On second suggestion, I didn't quiet get it. Could you explain about it in detail? Thank you very much Sincerely, Yoonho — Reply to this email directly, view it on GitHub <#8 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AA2EB2AONLIRWBSTMTO23Q3ZN45URAVCNFSM6AAAAABLJW7RUKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDENBWHE2TENBTHE> . You are receiving this because you commented.Message ID: ***@***.***>