Self-Reported Service Name as a Misclassified Proxy: An Artifact-Based Measurement Framework for Pre-Visit AI Consultation (AC-MDS), with a Proposed Feasibility Study in Psychiatric Outpatient Intake

This preprint proposes the AI Consultation Minimum Dataset (AC-MDS v0.2), an artifact-based measurement framework for capturing patients’ use of conversational AI before clinical visits. Existing studies primarily measure whether patients used AI, which service they report using, and their perceptions of the advice. AC-MDS instead treats the AI-generated response itself—the “artifact”—as the primary measurable exposure, because a self-reported service name may not identify the model tier or response characteristics that generated the advice. The framework consists of a nine-item intake instrument, a FHIR R4 reference implementation using Questionnaire, QuestionnaireResponse, DocumentReference, and Provenance resources, and a proposed validation procedure for LLM-derived content variables against dual human annotation. The manuscript also specifies a multicenter feasibility study in psychiatric outpatient intake in Japan. The proposed descriptive estimands include prevalence of pre-visit AI consultation, artifact content characteristics, tier-identifiability of self-reported service names, artifact yield relative to patient recall, and self-reported paid-tier use. No patient data have been collected for this manuscript. The proposed study is descriptive and does not support causal inference. This manuscript has not undergone peer review.