Anonpsy: A Graph-Based Framework for Structure-Preserving De-identification of Psychiatric Narratives

TL;DR

Anonpsy reframes psychiatric narrative de-identification as a structure-preserving generation problem. By converting narratives into a temporal-event semantic graph, applying graph-constrained perturbations, and regenerating text conditioned on the edited graph, it preserves diagnostic structure while substantially reducing re-identification risk. Across human expert and GPT-5 evaluations, Anonpsy maintains diagnostic usability comparable to original and LLM-only baselines while yielding notably lower semantic similarity and identifiability than unconstrained rewriting. The results suggest that explicit structural representations coupled with constrained generation can enhance privacy for narrative-rich clinical data and may generalize to other domains where identity is encoded in event content rather than surface identifiers.

Abstract

Psychiatric narratives encode patient identity not only through explicit identifiers but also through idiosyncratic life events embedded in their clinical structure. Existing de-identification approaches, including PHI masking and LLM-based synthetic rewriting, operate at the text level and offer limited control over which semantic elements are preserved or altered. We introduce Anonpsy, a de-identification framework that reformulates the task as graph-guided semantic rewriting. Anonpsy (1) converts each narrative into a semantic graph encoding clinical entities, temporal anchors, and typed relations; (2) applies graph-constrained perturbations that modify identifying context while preserving clinically essential structure; and (3) regenerates text via graph-conditioned LLM generation. Evaluated on 90 clinician-authored psychiatric case narratives, Anonpsy preserves diagnostic fidelity while achieving consistently low re-identification risk under expert, semantic, and GPT-5-based evaluations. Compared with a strong LLM-only rewriting baseline, Anonpsy yields substantially lower semantic similarity and identifiability. These results demonstrate that explicit structural representations combined with constrained generation provide an effective approach to de-identification for psychiatric narratives.

Figures (3)

• Figure 1: Schematic illustration of the recallability--structure trade-off in psychiatric narrative de-identification. The horizontal axis represents preservation of clinical structure (approximated by diagnostic soft-F1), and the vertical axis represents protection against narrative recallability (inversely related to cosine similarity to the original narrative). Example excerpts illustrate how PHI masking preserves surface structure while retaining narrative identity, unconstrained SDC reduces recallability at the cost of semantic drift, and Anonpsy achieves a more balanced trade-off via structure-aware perturbation. (PHI: Protected Health Information; SDC: Synthetic Data Creation; $\rho$: Cosine similarity)
• Figure 2: Overview of the Anonpsy pipeline for de-identification of psychiatric narratives via structure-preserving generation. The conversion operator transforms a free-text psychiatric narrative into a schema-constrained semantic graph; the perturbation operator selectively edits identifiable contextual elements while preserving temporal structure and typed relations; and the generation operator produces the final de-identified narrative.
• Figure 3: Violin plot of human-rated re-identification risk scores (1--5) for paired Anonpsy and LLM-only narratives. The horizontal line at score 3 indicates the predefined "at-risk" threshold.