Identification of Subclinical Language Deficit Using Machine Learning Classification Based on Poststroke Functional...

Area of Science:

• Neuroscience
• Radiology
• Machine Learning

Background:

• Post-stroke language deficits often go undetected without specialized testing.
• Early identification of subclinical language deficits (SLD) is crucial for timely intervention.
• Resting-state functional MRI (rs-fMRI) offers a non-invasive method for brain assessment.

Purpose of the Study:

• To evaluate the efficacy of rs-fMRI functional connectivity (FC) in identifying stroke subjects at risk for SLD.
• To compare the discriminative ability of different low-frequency bands (slow-5, slow-4, LFO) for SLD detection.
• To develop a machine learning model for expedited classification of stroke-related language impairments.

Main Methods:

• Sixty participants (20 early-stage stroke with language deficit (LD+), 20 early-stage stroke without (LD-), 20 healthy controls (HC)) underwent rs-fMRI.
• Functional connectivity (FC) was analyzed within the language network across slow-5, slow-4, and low-frequency oscillations (LFO) bands.
• A multiclass support vector machine (SVM) with nested leave-one-out cross-validation was employed for classification.

Main Results:

• The slow-4 frequency band achieved the highest classification accuracy (70%), outperforming LFO (65%) and slow-5 (50%).
• Subgroup-specific accuracies were high: LD+ (81.6%) and LD- (78.3%) using slow-4, and HC (80%) using slow-4/LFO.
• Frontal FC distinguished stroke from healthy controls, while occipital FC differentiated between stroke subgroups.

Conclusions:

• rs-fMRI, particularly using the slow-4 frequency band, can effectively and rapidly identify stroke subjects at risk for SLD.
• This approach offers a promising, expedited alternative to traditional time-intensive neuropsychological evaluations.
• FC patterns in specific brain regions (frontal, occipital) provide valuable diagnostic information for stroke-related language deficits.