Towards an Interpretable Functional Image-Based Classifier: Dimensionality Reduction of High-Density Diffuse Optical Tomography Data

High-density diffuse optical tomography (HD-DOT) is a wearable neuroimaging method that demonstrates high temporal and spatial resolution. While this data contains far richer information as a result, the high dimensionality and presence of complicated interconnections between data points requires the use of dimensionality reduction techniques to simplify the predictive modelling task without eliminating meaningful data features. To interrogate the possibility of designing a physiologically relevant HD-DOT feature set, cortical parcellations were applied to reconstructed images of brain activity to reduce the data dimensionality. A preliminary assessment of the predictive power of these parcel features was conducted on two binary tasks, with reasonable accuracies being achieved using standard classification models. Our results also demonstrated high spatial signal reproducibility across participants, which is promising for the application of image-based classification models that rely on spatial similarities to define separable class boundaries. These results provide insight into how the increased spatial resolution of HD-DOT can be leveraged to perform more accurate classification of neural data.