Predictive modeling of significance thresholding in activation likelihood estimation meta-analysis

Area of Science:

• Neuroimaging analysis
• Statistical methods in neuroscience
• Machine learning applications

Background:

• Activation Likelihood Estimation (ALE) meta-analyses use computationally intensive Monte-Carlo simulations for statistical thresholding (vFWE, cFWE, TFCE).
• These simulations are necessary to control for false positives in neuroimaging studies but can take many hours to complete.
• Current methods require substantial computational resources and time, limiting the scope of analyses.

Purpose of the Study:

• To develop and validate a machine learning approach to replace time-consuming Monte-Carlo simulations for ALE statistical thresholding.
• To create an efficient prediction model for determining significance thresholds (vFWE, cFWE, TFCE) in ALE meta-analyses.
• To reduce computational burden and energy consumption in neuroimaging meta-analysis.

Main Methods:

• Simulated 68,100 datasets with varying numbers of experiments, subjects, and foci to compute vFWE, cFWE, and TFCE thresholds.
• Trained XGBoost regression models on simulated data features (number of experiments, subjects, foci) for each thresholding technique.
• Validated models using 11 independent real-life ALE meta-analysis datasets (21 contrasts).

Main Results:

• The vFWE prediction model achieved near-perfect accuracy (R² = 0.996).
• TFCE and cFWE models demonstrated high prediction accuracies (R² = 0.951 and R² = 0.938, respectively).
• Predicted thresholds closely matched standard Monte-Carlo based thresholds, with average differences for cFWE less than two voxels.

Conclusions:

• The proposed machine learning approach accurately predicts ALE significance thresholds, offering a viable and efficient alternative to Monte-Carlo simulations.
• This method significantly reduces computation time and energy usage, enabling more complex analyses like leave-one-out sensitivity or subsampling.
• Adoption of this predictive model can streamline neuroimaging meta-analysis workflows and facilitate broader research applications.