Machine learning-based ischemic stroke detection and categorization with non-invasive plantar pressure data

The reliable and cost-effective automatic detection of stroke is crucial for prompt treatment and rehabilitation progress. However, the current diagnosis still relies on expensive brain imagining techniques and manual assessments, resulting in a need for medical visits, experts, or human errors.

This study has focused on development and validation of a machine learning-based screening algorithm using cost-effective and non-invasive foot pressure signals. The signals reflect affected motor-control mechanisms and gait disturbance following stroke with high temporal resolution. A novel set of foot pressure biomarkers has been proposed based on Empirical Fourier Decomposition combining an improved Fourier spectrum segmentation and a zero-phase filter bank to characterize the affected pressure distribution pattern. Afterward, the proposed approach combines the selection power of the ReliefF ranking algorithm and the classification power of support vector machine and K-nearest neighbors in a new framework to choose the discriminative pressure components, features, plantar areas, and sensors for offering high detection performance.

This automatic method has been evaluated using 198-foot plantar sensors recorded during walking in a sample of 82 subjects (46 healthy controls and 36 patients afflicted with stroke). It has achieved approximately perfect detection performance with an average accuracy rate of 99.19% and robust performance against clinical factors.

The proposed detection method is unique in its ability to provide robust performance against stroke sides and blood pressure status as well as to successfully detect stroke with a small number of biomarkers extracted from the toe and finger regions.