An automated machine learning approach to predict brain age from cortical anatomical measures

Area of Science:

• Neuroscience
• Machine Learning
• Computational Biology

Background:

• Machine learning (ML) adoption is rising in neuroscience, yet selecting optimal algorithms and hyperparameters remains challenging.
• Automated machine learning (autoML) offers a solution by systematically searching for the best ML models and configurations.

Purpose of the Study:

• To evaluate the efficacy of an autoML library, TPOT (Tree-based Pipeline Optimisation Tool), for brain age prediction using neuroimaging data.
• To determine if autoML can identify superior models and hyperparameters compared to existing methods for predicting subject age from brain structure.

Main Methods:

• Applied TPOT, an autoML tool utilizing genetic programming, to optimize ML pipelines for brain age prediction.
• Utilized Freesurfer-derived anatomical data (thickness and volume) as input features.
• Compared TPOT's performance against a relevance vector regression model.

Main Results:

• TPOT achieved higher accuracy in brain age prediction (Mean Absolute Error [MAE]: 4.612 ± 0.124 years) than relevance vector regression (MAE: 5.474 ± 0.140 years).
• TPOT identified novel model combinations that generalize well to unseen data, outperforming established models.
• The autoML approach successfully navigated the model space without prior assumptions, yielding state-of-the-art results.

Conclusions:

• AutoML, specifically TPOT, demonstrates significant promise as a data-driven method for discovering optimal neuroimaging analysis models.
• This approach automates the complex process of model selection and hyperparameter tuning in neuroimaging research.
• AutoML facilitates the development of more accurate and generalizable predictive models for neuroscience applications.