Modeling Power Asymmetry in AI-Mediated Clinical Empathy: A Dynamic Disentanglement Approach for Patient Self-Disclosure Recovery

When patients see a doctor, they often hold back—afraid to mention symptoms, downplay pain, or second-guess their own concerns. This self-censorship, rooted in doctor–patient power imbalances, undermines clinical empathy and creates diagnostic blind spots that current AI systems fail to address. Drawing on the “empathy by design” perspective, alongside Foucault’s analysis of medical authority, we propose the Power Dynamics Disentanglement Framework (PDDF). PDDF dynamically quantifies power asymmetry in clinical dialogue, recovers suppressed patient disclosures via a dual adversarial network, and calibrates empathetic responses through tripartite multi-agent reinforcement learning. Tested on 1,200 real clinical conversations from breast cancer outpatient clinics in a single Chinese tertiary hospital, PDDF reduces emotional recognition blind spots by 67% and improves early intervention rates for high‑risk patients by 41% (relative improvement).These results should be interpreted as preliminary proof-of-concept evidence: the study is based on a single Chinese tertiary hospital, a female-only breast cancer cohort, and same-institution clinician annotations, meaning broader generalisability remains untested. The 41% improvement derives from an uncontrolled observational deployment without a randomized control group; improvements may partly reflect clinician adaptation rather than the system itself.