Limits...
Multiclass Support Vector Machine-Based Lesion Mapping Predicts Functional Outcome in Ischemic Stroke Patients.

Forkert ND, Verleger T, Cheng B, Thomalla G, Hilgetag CC, Fiehler J - PLoS ONE (2015)

Bottom Line: Furthermore, integration of the optional features led to improved mRS prediction results in all cases tested.Therefore, a graded SVM-based functional stroke outcome prediction using the problem-specific brain regions for lesion overlap quantification leads to promising results but needs to be further validated using an independent database to rule out a potential methodical bias and overfitting effects.The prediction of the graded mRS functional outcome could be a valuable tool if combined with voxel-wise tissue outcome predictions based on multi-parametric datasets acquired at the acute phase.

View Article: PubMed Central - PubMed

Affiliation: Department of Radiology and Hotchkiss Brain Institute, University of Calgary, 3330 Hospital Drive NW, Calgary, AB T2N 4N1, Canada.

ABSTRACT

Purpose: The aim of this study was to investigate if ischemic stroke final infarction volume and location can be used to predict the associated functional outcome using a multi-class support vector machine (SVM).

Material and methods: Sixty-eight follow-up MR FLAIR datasets of ischemic stroke patients with known modified Rankin Scale (mRS) functional outcome after 30 days were used. The infarct regions were segmented and used to calculate the percentage of lesioned voxels in the predefined MNI, Harvard-Oxford cortical and subcortical atlas regions as well as using four problem-specific VOIs, which were identified from the database using voxel-based lesion symptom mapping. An overall of 12 SVM classification models for predicting the corresponding mRS score were generated using the lesion overlap values from the different brain region definitions, stroke laterality information, and the optional parameters infarct volume, admission NIHSS, and patient age.

Results: Leave-one-out cross validations revealed that including information about the stroke location in terms of lesion overlap measurements led to improved mRS prediction results compared to classification models not utilizing the stroke location information. Furthermore, integration of the optional features led to improved mRS prediction results in all cases tested. The problem-specific brain regions and additional integration of the optional features led to the best mRS predictions with a precise multi-value mRS prediction accuracy of 56%, sliding window multi-value mRS prediction accuracy (mRS±1) of 82%, and binary mRS (0-2 vs. 3-5) prediction accuracy of 85%.

Conclusion: Therefore, a graded SVM-based functional stroke outcome prediction using the problem-specific brain regions for lesion overlap quantification leads to promising results but needs to be further validated using an independent database to rule out a potential methodical bias and overfitting effects. The prediction of the graded mRS functional outcome could be a valuable tool if combined with voxel-wise tissue outcome predictions based on multi-parametric datasets acquired at the acute phase.

No MeSH data available.


Related in: MedlinePlus

Illustration of the single processing steps used for generation of the problem-specific brain regions in three selected slices.From top to bottom: MNI reference atlas, infarct distribution map used to exclude voxels lesioned in less than five patients from statistical calculations, p-value map used to exclude voxel with a significance level p≥0.05 from the VOI generation, median mRS values of lesioned and non-lesioned voxels used to define the final VOIs based on the median mRS difference d (VOI1: d > 2, VOI2: 1 < d ≤ 2, VOI3: 0 < d ≤ 1, VOI4: remaining voxels).
© Copyright Policy
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4476759&req=5

pone.0129569.g001: Illustration of the single processing steps used for generation of the problem-specific brain regions in three selected slices.From top to bottom: MNI reference atlas, infarct distribution map used to exclude voxels lesioned in less than five patients from statistical calculations, p-value map used to exclude voxel with a significance level p≥0.05 from the VOI generation, median mRS values of lesioned and non-lesioned voxels used to define the final VOIs based on the median mRS difference d (VOI1: d > 2, VOI2: 1 < d ≤ 2, VOI3: 0 < d ≤ 1, VOI4: remaining voxels).

Mentions: For this purpose, the affine transformations determined by registering the atlas to each patient dataset were inverted and used to transform the individually segmented lesions into the MNI atlas space employing a nearest-neighbour interpolation. After transformation of all lesion segmentations into atlas space, patients were separated for each voxel within the brain tissue into a lesion group and non-lesion group, respectively. After group separation for each voxel, the corresponding mRS scores of the patients were used to calculate the voxel-wise significance level (p-value) as well as the t-score employing a two-sided t-test. Furthermore, the median mRS score was determined for both groups and used to calculate the corresponding median mRS score difference d between the two groups for each voxel (Fig 1). These calculations were performed separately for patients with left- and right-hemispheric infarcts. Since the number of patients in the intact and lesion groups varies for each voxel, a minimum group size of 5 patients was arbitrarily defined to be required for performing the voxel-wise statistics.


Multiclass Support Vector Machine-Based Lesion Mapping Predicts Functional Outcome in Ischemic Stroke Patients.

Forkert ND, Verleger T, Cheng B, Thomalla G, Hilgetag CC, Fiehler J - PLoS ONE (2015)

Illustration of the single processing steps used for generation of the problem-specific brain regions in three selected slices.From top to bottom: MNI reference atlas, infarct distribution map used to exclude voxels lesioned in less than five patients from statistical calculations, p-value map used to exclude voxel with a significance level p≥0.05 from the VOI generation, median mRS values of lesioned and non-lesioned voxels used to define the final VOIs based on the median mRS difference d (VOI1: d > 2, VOI2: 1 < d ≤ 2, VOI3: 0 < d ≤ 1, VOI4: remaining voxels).
© Copyright Policy
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4476759&req=5

pone.0129569.g001: Illustration of the single processing steps used for generation of the problem-specific brain regions in three selected slices.From top to bottom: MNI reference atlas, infarct distribution map used to exclude voxels lesioned in less than five patients from statistical calculations, p-value map used to exclude voxel with a significance level p≥0.05 from the VOI generation, median mRS values of lesioned and non-lesioned voxels used to define the final VOIs based on the median mRS difference d (VOI1: d > 2, VOI2: 1 < d ≤ 2, VOI3: 0 < d ≤ 1, VOI4: remaining voxels).
Mentions: For this purpose, the affine transformations determined by registering the atlas to each patient dataset were inverted and used to transform the individually segmented lesions into the MNI atlas space employing a nearest-neighbour interpolation. After transformation of all lesion segmentations into atlas space, patients were separated for each voxel within the brain tissue into a lesion group and non-lesion group, respectively. After group separation for each voxel, the corresponding mRS scores of the patients were used to calculate the voxel-wise significance level (p-value) as well as the t-score employing a two-sided t-test. Furthermore, the median mRS score was determined for both groups and used to calculate the corresponding median mRS score difference d between the two groups for each voxel (Fig 1). These calculations were performed separately for patients with left- and right-hemispheric infarcts. Since the number of patients in the intact and lesion groups varies for each voxel, a minimum group size of 5 patients was arbitrarily defined to be required for performing the voxel-wise statistics.

Bottom Line: Furthermore, integration of the optional features led to improved mRS prediction results in all cases tested.Therefore, a graded SVM-based functional stroke outcome prediction using the problem-specific brain regions for lesion overlap quantification leads to promising results but needs to be further validated using an independent database to rule out a potential methodical bias and overfitting effects.The prediction of the graded mRS functional outcome could be a valuable tool if combined with voxel-wise tissue outcome predictions based on multi-parametric datasets acquired at the acute phase.

View Article: PubMed Central - PubMed

Affiliation: Department of Radiology and Hotchkiss Brain Institute, University of Calgary, 3330 Hospital Drive NW, Calgary, AB T2N 4N1, Canada.

ABSTRACT

Purpose: The aim of this study was to investigate if ischemic stroke final infarction volume and location can be used to predict the associated functional outcome using a multi-class support vector machine (SVM).

Material and methods: Sixty-eight follow-up MR FLAIR datasets of ischemic stroke patients with known modified Rankin Scale (mRS) functional outcome after 30 days were used. The infarct regions were segmented and used to calculate the percentage of lesioned voxels in the predefined MNI, Harvard-Oxford cortical and subcortical atlas regions as well as using four problem-specific VOIs, which were identified from the database using voxel-based lesion symptom mapping. An overall of 12 SVM classification models for predicting the corresponding mRS score were generated using the lesion overlap values from the different brain region definitions, stroke laterality information, and the optional parameters infarct volume, admission NIHSS, and patient age.

Results: Leave-one-out cross validations revealed that including information about the stroke location in terms of lesion overlap measurements led to improved mRS prediction results compared to classification models not utilizing the stroke location information. Furthermore, integration of the optional features led to improved mRS prediction results in all cases tested. The problem-specific brain regions and additional integration of the optional features led to the best mRS predictions with a precise multi-value mRS prediction accuracy of 56%, sliding window multi-value mRS prediction accuracy (mRS±1) of 82%, and binary mRS (0-2 vs. 3-5) prediction accuracy of 85%.

Conclusion: Therefore, a graded SVM-based functional stroke outcome prediction using the problem-specific brain regions for lesion overlap quantification leads to promising results but needs to be further validated using an independent database to rule out a potential methodical bias and overfitting effects. The prediction of the graded mRS functional outcome could be a valuable tool if combined with voxel-wise tissue outcome predictions based on multi-parametric datasets acquired at the acute phase.

No MeSH data available.


Related in: MedlinePlus