Promises, Pitfalls, and Basic Guidelines for Applying Machine Learning Classifiers to Psychiatric Imaging Data, with...

Area of Science:

• Neuroscience
• Psychiatry
• Machine Learning

Background:

• Psychiatric disorders often involve subtle brain function alterations and significant individual variability.
• Current diagnostic methods rely on time-consuming behavioral assessments by expert teams.
• Machine learning (ML) offers potential for faster, simpler psychiatric disorder diagnosis using neuroimaging data, but methods are often inaccessible.

Purpose of the Study:

• To describe and evaluate ML classification pipelines for diagnosing autism spectrum disorder (ASD) using resting-state fMRI data.
• To provide a tutorial-style explanation of ML approaches, assumptions, and challenges for researchers and clinicians.
• To offer accessible MATLAB code and data for replicating the analyses.

Main Methods:

• Utilized resting-state fMRI data from the ABIDE multisite repository.
• Compared several popular ML classifiers: support vector machines, neural networks, and regression approaches.
• Developed and explained classification pipelines in a tutorial format.

Main Results:

• Out-of-the-box ML classifiers achieved classification accuracies of approximately 60-70% for ASD.
• Provided a clear explanation of the rationale, assumptions, and potential pitfalls of each classification method.
• Made data and MATLAB code publicly available.

Conclusions:

• ML classifiers show promise for improving the speed and efficiency of psychiatric disorder diagnosis.
• Further methodological development is needed to enhance classification accuracy and facilitate clinical integration.
• Accessible explanations and tools can bridge the gap between ML research and clinical practice.