Problem
Prioritizing actions across multiple patients combines individual decisions with shared resource coordination. This project explores that relationship using reinforcement learning.
Implementation
I trained a Deep Q-Network for individual patient decisions and a supervisor agent for ward coordination. Using simulated patient vital signs, I explored monitoring, treatment, escalation, and emergency response decisions.
Technical decisions
- Separated individual patient decisions from ward resource coordination.
- Trained policies using reinforcement learning rather than hardcoding each decision.
- Integrated Claude to translate model decisions into readable explanations.
- Built dashboards to expose patient states, agent decisions, and system behavior.
Results
Dashboards show patient states, agent decisions, and Claude explanations so the learned policies can be reviewed.