Responsible AI Governance for Addressing Silent Risk in High-Risk Care Environments: Reimagining the VG-HITL Architecture and Decision Quality from an HRO Perspective

Digital efficiency in high-risk care settings can inadvertently silence life-critical signals embedded in unstructured narratives. We conceptualize Silent Risk as a governance failure in which clinically consequential cues (e.g., fall precursors, acute deterioration warnings, self-harm threats) are context-stripped and absorbed into routine administrative labels. Using 870 narrative records accumulated over 57 months in a long-term care facility, we triangulated iterative qualitative coding (NVivo) with cross-tabulation to trace concealment pathways within the legacy five-category scheme (Physical, Cognitive, Emotional, Behavioral, Others) and to operationalize concealment as the Risk Silence Rate (SR)—the proportion of records containing HRO-critical risk cues absorbed into non-critical categories. In an in-depth adjudication subset (N = 25) purposively sampled from screened potential risk cases, 12/25 contained Life_Safety risks dispersed across Physical, Emotional, and Cognitive categories; four recurrent Silent Risk types emerged: Life_Safety, Dignity, Depression, and Resistance (Meaningful Refusal). Grounded in the High Reliability Organization principle of preoccupation with failure, we propose a Responsibility-Embedded VG-HITL governance architecture that preserves context via a Value Graph, routes high-risk/uncertain/conflicting cases into mandatory expert adjudication, and feeds corrections back via RLHF to update detection rules and judgment logic, reframing decision quality toward minimizing fatal omissions and strengthening ethical accountability.