Hybrid statistical and machine learning modeling of cognitive neuroscience data

Area of Science:

• Neuroscience
• Statistics
• Machine Learning

Background:

• Nested data structures are common in cognitive neuroscience experiments with repeated measurements from multiple brain locations within subjects.
• Existing analysis methods often overlook the dependency structure inherent in these repeated measurements, potentially leading to inaccurate conclusions.
• There is a need for advanced analytical approaches that account for data dependencies in cognitive neuroscience.

Purpose of the Study:

• To compare statistical and machine learning methods for analyzing nested neuroscience data, considering the dependency structure of repeated measurements.
• To evaluate the performance of various algorithms, including hybrid forms, on novel neuroscience datasets.
• To assess algorithm fitting performance using contaminated datasets and cross-validation.

Main Methods:

• Exploration of statistical methods (e.g., Generalized Linear Mixed Models) and machine learning algorithms adapted for repeated measures.
• Implementation and comparison of distinct algorithms, including hybrid approaches.
• Evaluation of model performance using metrics like Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Error (MAE) on contaminated datasets via cross-validation.

Main Results:

• The Generalized Linear Mixed Model (GLMM) tree, incorporating random term indices for functional near-infrared spectroscopy (fNIRS) optode locations, exhibited the best predictive performance.
• The GLMM tree achieved the lowest MSE, RMSE, and MAE among the evaluated methods.
• A trade-off was observed between predictive accuracy and computational speed, with the GLMM tree requiring the most computational time.

Conclusions:

• The GLMM tree offers a powerful approach for analyzing nested cognitive neuroscience data by effectively handling dependencies.
• Accurate analysis of fNIRS data requires methods that account for the nested structure and measurement dependencies.
• Researchers must consider the balance between computational cost and predictive accuracy when selecting analytical methods for complex neuroscience datasets.