Delphi: A Neuro-Symbolic Framework for Individualized, Safe and Interpretable Treatment Recommendation

Clinical reinforcement learning (RL) holds promise for treatment recommendation. However, its adoption is hindered by black box decision processes, limited safety guarantees, and a lack of individualized treatment. To address these issues, we introduce Delphi, the first trainable neuro symbolic causal RL framework for dynamic treatment planning, designed to answer three core clinical questions: Why for this patient? Why is it safe? Why this action? Specifically, Delphi constructs: 1) causality aware state modeling, using discretized physiological variables and subgroup specific causal graphs; 2) adaptive symbolic rule constraints, combining clinical guidelines and behavior-derived rules into the RL system; and 3) interpretable decision fusion, where actions are selected based on joint neural symbolic Q values and explained via structured LLM-based justifications. We evaluate Delphi on MIMIC-III sepsis cohort with more than 20,000 trajectories, and experiments show that our Delphi achieves leading performance among existing methods. Moreover, Delphi introduces the first blinded physician evaluation of an explainable RL system in healthcare. Results demonstrate that Delphi consistently outperforms historical physicians' treatments in six dimensions, including adoption rate (+5.75%), understandability (+8.9%), safety (+10.4%), satisfaction (+9.35%), trust (+8.78%), effectiveness (+8.87%). These results highlight Delphi's potential as an interpretable, safe, and patientspecific AI assistant for critical care medicine.