Distributional Transformation Improves Decoding Accuracy When Predicting Chronological Age From Structural MRI.

Area of Science:

• Neuroimaging
• Machine Learning
• Biostatistics

Background:

• Predicting subject-level variables from biological data, like age from structural MRI, can be challenging.
• Decoding algorithms may not preserve the original distribution of the variable being predicted.
• Distributional transformation (DT) offers a potential solution by aligning predicted values with the training data's distribution.

Purpose of the Study:

• To evaluate the effectiveness of distributional transformation (DT) in improving age prediction accuracy from structural MRI data.
• To assess DT's performance across different machine learning models in a low-dimensional setting.
• To test DT within the context of the 2019 Predictive Analytics Competition (PAC).

Main Methods:

• Applied multiple linear regression, support vector regression, and deep neural networks to predict chronological age from structural MRI data.
• Utilized a low-dimensional setting (fewer features than observations).
• Incorporated distributional transformation (DT) to adjust predicted age values to match the training data distribution.

Main Results:

• In a low-dimensional setting, no tested method significantly outperformed linear regression.
• Distributional transformation (DT) consistently improved decoding performance across most methods, reducing the mean absolute error (MAE) by approximately half a year.
• Deep regression models were an exception, showing no improvement with DT.

Conclusions:

• Distributional transformation (DT) is a valuable technique for enhancing the accuracy of predicting variables with inherent distributions, such as age.
• DT can be particularly advantageous when dealing with non-controlled variables in both healthy and diseased populations.
• The findings suggest DT is a beneficial post-processing step for decoding algorithms in neuroimaging and other fields.