Improving Deep Reinforcement Learning With Transitional Variational Autoencoders: A Healthcare Application.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Biology

Background:

• Reinforcement learning (RL) offers potential for personalized medicine but direct patient training is ethically challenging.
• Environment models, trained on retrospective data, can simulate patient trajectories for RL agent development.

Purpose of the Study:

• To introduce transitional variational autoencoders (tVAE), a novel generative neural network architecture.
• To enable the training of RL agents using simulated patient data derived from retrospective healthcare records.

Main Methods:

• Developed a tVAE model for direct mapping between clinical measurements at adjacent time points.
• Utilized retrospective patient data to train the generative model.
• Compared tVAE performance against state-of-the-art sequential decision-making and generative models.

Main Results:

• The tVAE model demonstrated the ability to generate more realistic patient trajectories compared to existing methods.
• The architecture requires minimal distributional assumptions and features identical training and testing setups.
• Successfully simulated patient data for training effective treatment policies.

Conclusions:

• Transitional variational autoencoders (tVAE) provide a robust method for creating environment models in healthcare.
• This approach facilitates the ethical development of reinforcement learning agents for personalized treatment strategies.
• tVAE advances the application of AI in clinical decision-making and treatment optimization.