Level 6, 99 Commercial Road

Author Of 1 Presentation

Neuropsychology and Cognition Poster Presentation

P0831 - Validation of longitudinal reaction time trajectories in multiple sclerosis using the MSReactor computerized battery and latent class analysis. (ID 1660)

Speakers
Presentation Number
P0831
Presentation Topic
Neuropsychology and Cognition

Abstract

Background

Longitudinal cognitive trajectories in MS are heterogenous and difficult to measure. Computerised tests designed to screen broad cognitive domains may be a reliable method to detect discrete trajectories of cognitive performance

Objectives

To validate a latent class model to identify trajectories of reaction time change in people with relapsing remitting MS (pwRRMS).

Methods

The MSReactor computerised cognitive battery is a self-administered, online set of reaction time tasks assessing psychomotor function (PsychoM), visual attention (VisAtt) and working memory (WorkingM). Participants completed both 6-monthly in-clinic testing and additional remote testing. Latent class analysis was used to model the longitudinal reaction times and identify discrete cognitive trajectories within the heterogeneous data. We applied a cross validation method to confirm the optimal models for all three reaction time tasks. Briefly, the cohort was split into training and test sets (50:50) and the mean root mean square error (RMSE) calculated for the difference between training and test trajectories calculated at each day of follow up, over 100 repetitions. To determine the minimum number of tests required to reliably classify an individual into a longitudinal trajectory the dataset for each task was reduced to 3, 4 and 5 tests per participant and classification compared to the complete dataset.

Results

We included 478 pwRRMS who had completed at least 3 MSReactor tests over a minimum of 30 days follow up. In total, 3846 individual tests were included in each model (median tests/participant=5 (range 3-332), mean follow up of 774 +- 359.5 days). Three latent classes were identified for each task, with the PsychoM task identifying a group of pwRRMS with a mean predicted trajectory of slowing reaction times. In validation, the PsychoM task was the most consistent with the smallest RMSE for each of the 3 classes (68 milliseconds (ms), 95%CI 59-77ms; 61ms, 95% CI 50-72ms and 16ms, 95% CI 14-18ms) followed by the VisAtt task (RMSE = 114ms, 95ms and 29ms) and WorkingM task (RMSE = 137ms, 138ms and 46ms). Reduced datasets of 3, 4 and 5 tests per participant in the PsychoM task were able to classify participants into the trajectories identified in the full dataset model, with 83%, 86% and 90% accuracy respectively.

Conclusions

Latent class modelling of longitudinal reaction times collected with MSReactor was able to detect discrete trajectories of cognitive performance. The PsychoM task latent class model identified a group of RRMS with a mean predicted slowing of reaction times and was the most consistent model in cross validation. Performing the modelling on just 3, 4 or 5 tests per participant was highly accurate in defining latent trajectories, giving the battery clinical utility where multiple years of follow up may not be realistic.

Collapse

Presenter Of 1 Presentation

Neuropsychology and Cognition Poster Presentation

P0831 - Validation of longitudinal reaction time trajectories in multiple sclerosis using the MSReactor computerized battery and latent class analysis. (ID 1660)

Speakers
Presentation Number
P0831
Presentation Topic
Neuropsychology and Cognition

Abstract

Background

Longitudinal cognitive trajectories in MS are heterogenous and difficult to measure. Computerised tests designed to screen broad cognitive domains may be a reliable method to detect discrete trajectories of cognitive performance

Objectives

To validate a latent class model to identify trajectories of reaction time change in people with relapsing remitting MS (pwRRMS).

Methods

The MSReactor computerised cognitive battery is a self-administered, online set of reaction time tasks assessing psychomotor function (PsychoM), visual attention (VisAtt) and working memory (WorkingM). Participants completed both 6-monthly in-clinic testing and additional remote testing. Latent class analysis was used to model the longitudinal reaction times and identify discrete cognitive trajectories within the heterogeneous data. We applied a cross validation method to confirm the optimal models for all three reaction time tasks. Briefly, the cohort was split into training and test sets (50:50) and the mean root mean square error (RMSE) calculated for the difference between training and test trajectories calculated at each day of follow up, over 100 repetitions. To determine the minimum number of tests required to reliably classify an individual into a longitudinal trajectory the dataset for each task was reduced to 3, 4 and 5 tests per participant and classification compared to the complete dataset.

Results

We included 478 pwRRMS who had completed at least 3 MSReactor tests over a minimum of 30 days follow up. In total, 3846 individual tests were included in each model (median tests/participant=5 (range 3-332), mean follow up of 774 +- 359.5 days). Three latent classes were identified for each task, with the PsychoM task identifying a group of pwRRMS with a mean predicted trajectory of slowing reaction times. In validation, the PsychoM task was the most consistent with the smallest RMSE for each of the 3 classes (68 milliseconds (ms), 95%CI 59-77ms; 61ms, 95% CI 50-72ms and 16ms, 95% CI 14-18ms) followed by the VisAtt task (RMSE = 114ms, 95ms and 29ms) and WorkingM task (RMSE = 137ms, 138ms and 46ms). Reduced datasets of 3, 4 and 5 tests per participant in the PsychoM task were able to classify participants into the trajectories identified in the full dataset model, with 83%, 86% and 90% accuracy respectively.

Conclusions

Latent class modelling of longitudinal reaction times collected with MSReactor was able to detect discrete trajectories of cognitive performance. The PsychoM task latent class model identified a group of RRMS with a mean predicted slowing of reaction times and was the most consistent model in cross validation. Performing the modelling on just 3, 4 or 5 tests per participant was highly accurate in defining latent trajectories, giving the battery clinical utility where multiple years of follow up may not be realistic.

Collapse