Disconnection-Based Prediction of Poststroke Dysphagia

Kyung Jae Yoon; Chul-Hyun Park; Myung-Ho Rho; Minchul Kim

doi:10.3174/ajnr.A8074

Article Figures & Data

Figures

Tables

Download figure
Open in new tab
Download powerpoint
FIG 1.
Flow chart of participants with poststroke dysphagia recruited for disconnectome analysis.
Download figure
Open in new tab
Download powerpoint
FIG 2.
Schematic of using a CPM-SVM to predict poststroke dysphagia from the structural disconnectome. A, Processing scheme of a patient with internal capsule posterior limb infarction (red arrow). We segmented the infarct core using ADC maps and DWI. To normalize the lesion mask to the MNI space, we used warping results derived from the patient's FA map normalized to the FA template. In the end, we get the segmented lesion normalized in MNI space (pink in the image on the right). B, The lesion Quantification Toolkit uses the lesion segmentation in A to estimate the structural disconnection using the HCP-842 population-averaged streamline tractography atlas. The blue tractography on the left is the tracts disconnected by the internal capsule posterior limb infarction, and the brain plotted on the right shows the distribution of direct and indirect disconnection due to infarction. The patient had poststroke dysphagia probably owing to disruption in the motor network. C, Using the disconnection derived from B as prediction features, we adopted the CPM-SVM model and LOOCV to predict poststroke dysphagia.
Download figure
Open in new tab
Download powerpoint
FIG 3.
Leave-one-out classification result. A, Classification results in which, by using both the disconnection and SSPL, increased matrix of the motor network (red arrow) presented the highest accuracy. B, Confusion matrix of the motor network–based model, where the vertical axis is the true label of all patients. Blue boxes represent individuals correctly identified by the model. Orange boxes represent incorrect identification. Percentages in each box correspond to the proportions of subjects in the subgroup relative to the total subjects. C and D, Classification result when combined with demographic variables, indicating that the model, by using both disconnection and demographic features, shows the best prediction accuracy of 72.1%.
Download figure
Open in new tab
Download powerpoint
FIG 4.
Spatial network anatomy of the motor network-based model. A, Spatial extent of the “consensus edges” from the disconnection network. In the circular plot, regions are organized according to their anatomic locations, with more anterior regions at the top and more ventral and posterior regions displayed toward the bottom. Node size in the brain represents the degree of the node (ie, the number of connections with the node). B, Summarization of “consensus edges” from the disconnection network. The nodes were filtered to have at least 20 edges for visualization purposes.
Download figure
Open in new tab
Download powerpoint
FIG 5.
Representative case of a 62-year-old man with left middle cerebellar peduncle ischemic infarction. He did not have poststroke dysphagia, and the model also predicted negative results. A, Substantial infarction volume in DWI. B, Disconnected tract attenuation map (red) due to infarction. Nonetheless, the disconnection is mostly distributed in the cerebellar network, with no disconnection within the motor network (C, red square; D), which may be preventing the patient from poststroke dysphagia.
Download figure
Open in new tab
Download powerpoint
FIG 6.
Representative case of a 39-year-old woman with left corona radiata ischemic infarction. The patient did not have poststroke dysphagia. She had disconnection in the motor network and not predicted correctly only by disconnection matrix (C, red square). However, the model incorporating both demographic variable and disconnection predicted correctly, maybe owing to the patient's relatively young age of 39. A, Left corona radiata infarction in DWI. B, Disconnected tract attenuation map (red) due to infarction. Substantial amount of the disconnection distributed in the motor network (C, red square); however, the prediction model incorporating both disconnection and age, sex could predict correctly.

Tables

Figures

Clinical characteristics of the study population

	Full Per-Oral Nutrition (n = 43, 17 Women)	Nonoral Nutrition Necessary (n = 43, 13 Women)	P Value
Age (yr)	68.83 ± 15.69	75.09 ± 11.38	t = 2.264, P = .029^{^a}
HTN	35	34	χ² = 0.073, P = .786
DM	15	19	χ² = 0.778, P = .377
Onset to VFSS interval (days)	11.67 ± 7.50	12.24 ± 7.64	t = 0.327, P = .745
MRI scanner	Intera = 35 Ingenia = 5 Achieva = 3	Intera = 37 Ingenia = 4 Achieva = 2	χ² = 0.367, P = .832