Testing the disconnectome symptom discoverer model on out-of-sample post-stroke language outcomes

Hope, T. M. H., Neville, D., Talozzi, L., Foulon, C., Forkel, S. J., Thiebaut de Schotten, M., & Price, C. J. (2024). Testing the disconnectome symptom discoverer model on out-of-sample post-stroke language outcomes. Brain, 147(2), e11-e13. doi:10.1093/brain/awad352.
Stroke is common, and its consequent brain damage can cause various cognitive impairments. Associations between where and how much brain lesion damage a patient has suffered, and the particular impairments that injury has caused (lesion-symptom associations) offer potentially compelling insights into how the brain implements cognition.1 A better understanding of those associations can also fill a gap in current stroke medicine by helping us to predict how individual patients might recover from post-stroke impairments.2 Most recent work in this area employs machine learning models trained with data from stroke patients whose mid-to-long-term outcomes are known.2-4 These machine learning models are tested by predicting new outcomes—typically scores on standardized tests of post-stroke impairment—for patients whose data were not used to train the model. Traditionally, these validation results have been shared in peer-reviewed publications describing the model and its training. But recently, and for the first time in this field (as far as we know), one of these pre-trained models has been made public—The Disconnectome Symptom Discoverer model (DSD) which draws its predictors from structural disconnection information inferred from stroke patients’ brain MRI.5

Here, we test the DSD model on wholly independent data, never seen by the model authors, before they published it. Specifically, we test whether its predictive performance is just as accurate as (i.e. not significantly worse than) that reported in the original (Washington University) dataset, when predicting new patients’ outcomes at a similar time post-stroke (∼1 year post-stroke) and also in another independent sample tested later (5+ years) post-stroke. A failure to generalize the DSD model occurs if it performs significantly better in the Washington data than in our data from patients tested at a similar time point (∼1 year post-stroke). In addition, a significant decrease in predictive performance for the more chronic sample would be evidence that lesion-symptom associations differ at ∼1 year post-stroke and >5 years post-stroke.
Publication type
Journal article
Publication date

Share this page