Neural correlates of imaging biomarkers for primary motor deficits after stroke - more than just the primary motor system?

Abstract

Hemiparesis is a common consequence of stroke to the primary motor system. Previous studies suggested that damage to additional brain areas might play a causal role in the occurrence and severity of hemiparesis and its recovery. Knowledge of these regions might be applied in the creation of imaging biomarkers for motor outcome prediction if lesion information from such areas carries predictive value. We assessed acute and chronic paresis of the upper limb in 102 patients with unilateral stroke. In a first experiment, the neural correlates of acute and chronic upper limb paresis were mapped by lesion behaviour mapping. Following the same approach, a lesion biomarker of corticospinal tract (CST) damage was mapped. This analysis served as an artificial control condition as the biomarker, by definition, is only causally related to damage of the CST. Mapping acute or chronic upper limb paresis implicated areas outside of the primary motor system. Likewise, mapping the CST lesion biomarker implicated several areas outside of the CST with high correspondence to areas associated with upper limb paresis. Damage to areas outside of the primary motor system thus might, to some degree, not play a causal role in hemiparesis. In a second experiment, we showed that lesion information from these areas outside of the primary motor system can be used to predict motor outcome. This was even the case for the CST lesion biomarker. Although the only causal source underlying the CST lesion biomarker was damage to the CST, lesion information that mainly included non-CST regions was able to predict the biomarker (non-significantly) better than information taken from only the CST itself. These findings suggest that simple theory-based biomarkers or qualitative rules to infer post-stroke outcome from imaging data might perform sub-optimally, as they do not consider the complexity of lesion data. Instead, high-dimensional models with data-driven feature selection strategies might be required.