An individual-level weighted artificial neural network method to improve the systematic bias in BrainAGE analysis

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• BrainAGE analysis is crucial for understanding brain development in neuropsychiatric disorders.
• A known "regression toward mean" bias affects BrainAGE predictions, underestimating older and overestimating younger individuals' brain ages.
• This bias can hinder accurate assessment of atypical brain development patterns.

Purpose of the Study:

• To introduce an individual-level weighted artificial neural network (ANN) method to mitigate the "regression toward mean" effect in BrainAGE analysis.
• To evaluate the proposed method's performance against traditional machine learning techniques using simulated and real datasets.
• To compare the efficacy of different ANN activation functions (sigmoid vs. ReLU) in this context.

Main Methods:

• Development and application of an individual-level weighted artificial neural network (ANN) model.
• Testing the model on a large simulation dataset (5000 subjects) and a real-world dataset (135 subjects).
• Comparative analysis against traditional machine learning methods for BrainAGE prediction.
• Evaluation of sigmoid versus ReLU activation functions within the ANN.

Main Results:

• The individual-level weighted ANN strategy significantly reduced the "regression toward mean" effect compared to traditional methods.
• Prediction performance remained comparable to existing machine learning approaches.
• The sigmoid activation function demonstrated superior performance over the ReLU function in the ANN model.
• The proposed method effectively addressed biases in BrainAGE estimations.

Conclusions:

• The novel individual-level weighted ANN method offers a viable solution to the "regression toward mean" bias in BrainAGE analysis.
• This strategy enhances the accuracy of assessing atypical brain development in neuropsychiatric disorders.
• The findings support the use of weighted ANNs and sigmoid activation functions for improved BrainAGE studies.