Continuous-Time Decision Transformer for Healthcare Applications.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Healthcare Informatics

Background:

• Offline reinforcement learning (RL) shows promise for medical applications like treatment policy learning and clinical decision support.
• Existing Decision Transformer models are limited to discrete-time scenarios, hindering their use in continuous-time medical decision-making.

Purpose of the Study:

• To generalize Decision Transformer for continuous-time medical decision-making.
• To develop a model that learns optimal treatment plans and visit schedules from historical healthcare data.

Main Methods:

• Generalized Decision Transformer to Continuous-Time Decision Transformer (CTDT).
• Incorporated past clinical measurements, treatments, and visit timings into the model.
• Trained CTDT to predict future visit timings and treatment plans.

Main Results:

• CTDT outperformed existing methods in experiments on synthetic and simulated medical datasets.
• Demonstrated clinical utility in improving patient health outcomes and survival rates.
• Successfully learned high-performance policies from logged data of varying quality.

Conclusions:

• Continuous-Time Decision Transformer is a viable approach for offline RL in continuous-time medical domains.
• CTDT offers significant potential for enhancing clinical decision support and patient care.
• The model effectively learns from historical healthcare data to optimize treatment strategies.