EUVIMEDOpdag EUVIMED
Open access

AI-Driven Reconstruction of Online Health Narratives Using Agent-Action-Outcome Modeling: Algorithm Development and Validation

Ommo Clark, Karuna Pande Joshi

JMIR AI · 2026-11-13 · 5 · e83496-e83496

DOI: 10.2196/83496 ↗

At a glance

Journal article in JMIR AI, published online on 2026-11-13. Written by Ommo Clark and 1 co-authors. Institution: Information Systems Department, University of Maryland, Baltimore County, 1000 Hilltop Circle, Baltimore, MD, 21250, United States, 1 410 455 1000. 46 references. The full text is freely available in PubMed Central.

AbstractSubjectsBibliografiske oplysningerCiteRelated studies

Abstract (original language)

Background

People share health experiences online as stories, not isolated facts. Most computational systems reduce these narratives to keywords or claims, overlooking the agency, causality, and interpretation that determine whether health information is safe or potentially harmful to act on. Content can use medically accurate terminology while describing unsafe behavior, rendering it invisible to fact-checking and keyword approaches. This limitation, which we term narrative blindness , constrains the ability of computational systems to extract, represent, and reason over narratives as coherent wholes.

Objective

This study introduces the Narrative Builder, a computational framework that reconstructs unstructured online health stories into machine-readable agent-action-outcome (AAO) narrative graphs, preserving structural, causal, and interpretive dimensions. We formalize narrative blindness as a measurable property of natural language processing systems and evaluate whether theory-grounded reconstruction can address it.

Methods

We operationalized Labov and Waletzky’s sociolinguistic narrative theory within a computational AAO schema, enriched with interpretive dimensions from psychological and clinical communication frameworks. From 5386 Reddit health threads (spanning February 2019 to November 2024), 2 domain experts (a biomedical researcher and a physician) annotated a gold-standard dataset of 2000 segments from 704 threads, with interrater reliability (Krippendorff α) of 0.75 to 0.82 across all annotated tasks (0.78-0.81 for the 4 interpretive classification tasks). A dual-model architecture combining BioBERT (Bidirectional Encoder Representations From Transformers for Biomedical Text Mining) with a conditional random field (CRF) layer for AAO extraction and multitask DistilBERT for interpretive classification was evaluated using 5-fold grouped cross-validation, grouped on the annotation-level key to prevent data leakage, with a separate held-out test set and a McNemar test against term frequency–inverse document frequency (TF-IDF) and BioBERT-without-CRF baselines.

Results

The Narrative Builder achieved a span-level F 1 -score of 0.850 for AAO extraction on the held-out test set, with per-type F 1 of 0.90 (action), 0.83 (agent), and 0.77 (outcome), outperforming the TF-IDF baseline (0.781; P

Subjects

Author keywords

DistilBERTLabovian theoryRedditagent-action-outcome modelingcomputational linguisticshealth communicationhealth misinformationnarrative analysisnarrative blindnessnatural language processingsocial media

Authors and institutions

  • Information Systems Department, University of Maryland, Baltimore County, 1000 Hilltop Circle, Baltimore, MD, 21250, United States, 1 410 455 1000.

Cite

APA

Clark, O., & Joshi, K. P. (2026). AI-Driven Reconstruction of Online Health Narratives Using Agent-Action-Outcome Modeling: Algorithm Development and Validation. JMIR AI, 5, e83496-e83496. https://doi.org/10.2196/83496

Vancouver

Clark O, Joshi KP. AI-Driven Reconstruction of Online Health Narratives Using Agent-Action-Outcome Modeling: Algorithm Development and Validation. JMIR AI. 2026;5:e83496-e83496. doi:10.2196/83496

BibTeX
@article{Clark2026,
  author = {Clark, Ommo and Joshi, Karuna Pande},
  title = {AI-Driven Reconstruction of Online Health Narratives Using Agent-Action-Outcome Modeling: Algorithm Development and Validation},
  journal = {JMIR AI},
  year = {2026},
  volume = {5},
  pages = {e83496--e83496},
  doi = {10.2196/83496}
}
Start for free →