Programa del congreso

Critically ill patients often experience pain, which can disrupt cardiorespiratory function, cause emotional distress, and lead to long-term complications. Although several analgesic and sedative medications are available, they may have profound, even life-threatening effects. Optimal strategies must therefore balance pain relief with patient safety. Reinforcement learning (RL) provides a framework for formalizing therapy as a sequence of decisions and identifying optimal policies. Small datasets, simplified actions, and omission of patient mortality have limited prior applications to sedation and analgesia. Our research addresses these gaps by jointly considering patient well-being and safety.

We modeled the task as a partially observable Markov decision process (POMDP) defined over observations of patient experience, cardiorespiratory function, metabolism, and one-year survival. The action space included four continuous dosing signals: opioids, propofol, benzodiazepines, and dexmedetomidine. We defined 34 reward functions and analyzed their asymptotic properties. We derived a cohort of 42591 intensive care stays from the MIMIC-IV database, and latent representations of these stays were learned to approximate the POMDP. Finally, the TD3-BC algorithm was applied to train our policies (data were split 80/20 for training and evaluation).

Both theoretical and empirical analyses highlighted a strategy that improved patient safety and comfort. The proposed actions were associated with a reduction of mortality by up to 59% and mean pain levels by up to 45% per stay, restricting opioids and propofol while relying on dexmedetomidine. Consequently, our results extend previous work through a richer model, a more comprehensive dataset, and policies that jointly minimize pain and one-year mortality.