Robust diagnostic classification via Q-learning.

Area of Science:

• Clinical informatics
• Artificial intelligence in medicine
• Computational neuroscience

Background:

• Machine learning (ML) models offer valuable tools for complex clinical diagnoses.
• Current ML models require interpretability and robustness, especially with noisy or missing data.
• Diagnostic classification is often a complex decision-making process.

Purpose of the Study:

• To formulate diagnostic classification as a decision-making process using Q-learning.
• To develop interpretable and robust ML classifiers for clinical use.
• To enhance diagnostic accuracy while minimizing data requirements.

Main Methods:

• Utilized Q-learning, a reinforcement learning algorithm, to build diagnostic classifiers.
• Formulated the diagnostic classification task as a sequential decision-making problem.
• Simulated the differentiation of Autism Spectrum Disorder (ASD) from Attention Deficit-Hyperactivity Disorder (ADHD) in children.

Main Results:

• Developed Q-learning based classifiers demonstrating interpretability and robustness.
• Achieved high diagnostic accuracy in the simulated ASD vs. ADHD classification task.
• Showcased the ability to train robust classifiers by optimizing for accuracy and data efficiency.

Conclusions:

• Q-learning provides a viable framework for creating interpretable and robust diagnostic ML models.
• Reinforcement learning can enhance classifier performance in complex clinical scenarios.
• This approach holds promise for improving diagnostic tools in child neurology and psychiatry.