Prediction of post-stroke motor recovery benefits from measures of sub-acute widespread network damages

Abstract

Following a stroke in regions of the brain responsible for motor activity, patients can lose their ability to control parts of their body. Over time, some patients recover almost completely, while others barely recover at all. It is known that lesion volume, initial motor impairment and cortico-spinal tract asymmetry significantly impact motor changes over time. Recent work suggested that disabilities arise not only from focal structural changes but also from widespread alterations in inter-regional connectivity. Models that consider damage to the entire network instead of only local structural alterations lead to a more accurate prediction of patients' recovery. However, assessing white matter connections in stroke patients is challenging and time-consuming. Here, we evaluated in a data set of 37 patients whether we could predict upper extremity motor recovery from brain connectivity measures obtained by using the patient's lesion mask to introduce virtual lesions in 60 healthy streamline tractography connectomes. This indirect estimation of the stroke impact on the whole brain connectome is more readily available than direct measures of structural connectivity obtained with magnetic resonance imaging. We added these measures to benchmark structural features, and we used a ridge regression regularization to predict motor recovery at 3 months post-injury. As hypothesized, accuracy in prediction significantly increased (R ² = 0.68) as compared to benchmark features (R ² = 0.38). This improved prediction of recovery could be beneficial to clinical care and might allow for a better choice of intervention.

Keywords: Stroke; brain connectivity; motor recovery; prediction.

Graphical Abstract

Figure 1

Prediction model based on brain connectivity measures. First step: a lesion mask is drawn for each patient from the structural MRI. Second step: for each patient, the lesion mask is intersected with each of the 60 healthy streamline tractography connectomes deleting all white matter tracts passing through the lesioned area leading to 60 virtually lesioned connectomes. Third step: from each of these virtually lesioned connectomes, we estimate brain connectivity measures and the latter are averaged over connectomes. Fourth step: these measures are used as input for a ridge regression model to predict motor improvement.

Figure 2

Collinearity between predictive measures. (A) Spearman correlation matrix of all included variables. Yellow indicates a positive and blue a negative correlation. (B) Hierarchical clustering based on the Spearman correlation matrix.

Figure 3

Scatter plot showing fitter (cross symbols ) and non-fitter (rhomboid symbols) patients according to the proportional recovery model. The separation into two recovery groups is visible. Patients that recovered more than 30% threshold of FMA recovery score coincided with the fitter patients, whereas the patients that recovered less than 30% threshold of FMA recovery score coincided with the non-fitter group of patients. The line represents a simple linear regression including the fitter patients.

Figure 4

Prediction accuracy. (A) scatter plot of observed percentage of FMA recovery score versus predicted percentage of FMA recovery score for the five features set tested. In red the misclassified patients and in dark blue the correctly classified patients. The lines represent simple linear regressions including all participants. (B) Lines represent 95% confidence intervals for the proportional recovery prediction models R² using bootstrapping for the five feature sets. Square indicates R². Feature sets are significantly different if 95% confidence intervals do not overlap.

Figure 5

Stability of global measures. (A)R² versus VIF for Set 4. The dash line indicates the R² for the benchmark features (Set 3). (B) Number of times a feature is selected per fold. (C) Weights fitted for each feature (average and standard deviation over folds).