A normative modeling approach to quantify white matter changes and predict functional outcomes in stroke patients

Abstract

Objectives: The diverse nature of stroke necessitates individualized assessment, presenting challenges to case-control neuroimaging studies. The normative model, measuring deviations from a normal distribution, provides a solution. We aim to evaluate stroke-induced white matter microstructural abnormalities at group and individual levels and identify potential prognostic biomarkers.

Methods: Forty-six basal ganglia stroke patients and 46 healthy controls were recruited. Diffusion-weighted imaging and clinical assessment were performed within 7 days after stroke. We used automated fiber quantification to characterize intergroup alterations of segmental diffusion properties along 20 fiber tracts. Then each patient was compared to normative reference (46 healthy participants) by Mahalanobis distance tractometry for 7 significant fiber tracts. Mahalanobis distance-based deviation loads (MaDDLs) and fused MaDDL_multi were extracted to quantify individual deviations. We also conducted correlation and logistic regression analyses to explore relationships between MaDDL metrics and functional outcomes.

Results: Disrupted microstructural integrity was observed across the left corticospinal tract, bilateral inferior fronto-occipital fasciculus, left inferior longitudinal fasciculus, bilateral thalamic radiation, and right uncinate fasciculus. The correlation coefficients between MaDDL metrics and initial functional impairment ranged from 0.364 to 0.618 (p < 0.05), with the highest being MaDDL_multi. Furthermore, MaDDL_multi demonstrated a significant enhancement in predictive efficacy compared to MaDDL (integrated discrimination improvement [IDI] = 9.62%, p = 0.005) and FA (IDI = 34.04%, p < 0.001) of the left corticospinal tract.

Conclusion: MaDDL_multi allows for assessing behavioral disorders and predicting prognosis, offering significant implications for personalized clinical decision-making and stroke recovery. Importantly, our method demonstrates prospects for widespread application in heterogeneous neurological diseases.

Keywords: Mahalanobis distance; normative modeling; prognosis; stroke; white matter microstructure.

Figure 1

The workflow of this study. (A) Stroke lesion map. Unilateral hemispheric stroke lesions overlapped in 46 patients. The colored bars indicated the number of patients with lesions in the voxels. (B) Twenty main white matter fiber tracts were analyzed. (C,D) As an example, the corticospinal tract has been shown here. L, left side; R, right side; AFQ, automated fiber quantification; FA, fractional anisotropy; MD, mean diffusivity; AD, axial diffusivity; RD, radial diffusivity; MaD, Mahalanobis distance.

Figure 2

Pointwise comparison of diffusion parameters between groups of 20 white matter fiber tracts (only those with differences were shown). The blue lines represented the HC group and the orange lines represented the BGS group (solid lines for means and shading for standard deviations). The pink lines showed significantly changed segments (p < 0.05/20, FDR correction). HC, healthy controls; BGS, basal ganglia stroke; CST_L, left corticospinal; TR_L, left thalamic radiation; TR_R, right thalamic radiation; IFOF_L, left inferior fronto-occipital fasciculus; IFOF_R, right fronto-occipital fasciculus; ILF_L, left inferior longitudinal fasciculus; UF_R, right uncinate fasciculus.

Figure 3

Detection of patient-specific microstructural abnormalities with MaD-Tract. (A) The example presented the 4 univariate diffusion parameter distributions of the left corticospinal tract for one patient and the healthy control group. Segments that exceeded the anomaly index threshold (gray shading in the MaD plot) were labeled as anomalous. The orange area of the MaD curve above the threshold was depicted as MaDDL. (B) Heat map of the distribution of structural abnormalities by MaD tract analysis in each patient across the seven abnormal fiber tracts. If at least one segment is marked as abnormal, it is highlighted. (C) Scatter plot of the MaDDL for each patient. MaD, Mahalanobis distance; MaDDL, Mahalanobis distance-based deviation load.

Figure 4

(A) Scatterplots and correlations for NIHSS and MaDDL metrics as well as FA_{CST_L} from Pearson correlation analyses. (B) Scatterplots and correlations for NIHSS and MaDDL metrics as well as FA_{CST_L} from partial correlation analyses with age and sex as covariates. (C) MaDDL_multi and FA_{CST_L} regressed univariately against 3-month mRS (Model 1). MaDDL_{CST_L} and MaDDL_multi regressed against 3-month mRS with age and sex as covariates (Model 2). In addition, MaDDL_multi demonstrated a significant enhancement in predictive efficacy compared to MaDDL_{CST_L} (IDI = 9.62%, p = 0.005) and FA_{CST_L} (IDI = 34.04%, p < 0.001).