From Narrative to Knowledge: Developing an Ontology-Based Sdoh Extraction Pipeline from Pediatric Psychosocial Assessments

Background and Purpose
Social determinants of health (SDoH) are central to social work research and practice, yet much of this information remains embedded in unstructured clinical text narratives, limiting its use in empirical research and decision-making. In organ transplant social work, psychosocial assessments such as the Pediatric Psychosocial Assessment Tool (PPAT) capture multidimensional information on family context, mental health, financial stability, and adherence risk. However, these narrative summaries are rarely leveraged as structured data despite their potential to inform outcomes, intervention planning, and research. Prior work has demonstrated that the PPAT provides a valid and reliable framework for assessing psychosocial risk across domains, but scalable methods for extracting and standardizing this information are lacking. This study aimed to develop and evaluate an ontology-driven SDoH extraction pipeline using PPAT narrative summaries, demonstrating how social work-generated data can be transformed into structured, analysis-ready variables.

Methods
We developed a multi-stage ontology pipeline using a subsample of de-identified PPAT narrative summaries drawn from a larger cohort of pediatric heart transplant evaluations (N=189). A subset (n=33) was manually annotated by trained coders using a custom guideline informed by practitioner expertise and existing SDoH frameworks. Annotation followed a two-level structure: Level 1 identified SDoH domains and Level 2 contextual labels capturing attributes such as experiencer (i.e., parent/caregiver or pediatric patient) and severity. Six annotators worked in pairs with expert review to ensure consistency. We operationalized 16 SDoH categories and developed an ontology to standardize relationships among concepts. For extraction, we benchmarked transformer-based models for level 1 SDoH domain identification and applied instruction-tuned large language models for capturing level 2 contextual information. A structured function-calling schema constrained level 2 outputs to predefined ontology classes, and rule-based processing converted outputs into structured datasets. Model performance was evaluated using sentence-level multi-label classification metrics and agreement with annotated labels.

Results
The ontology pipeline successfully extracted SDoH concepts and contextual attributes from narrative summaries with moderate-to-strong performance across categories. Inter-rater agreement demonstrated acceptable reliability (76% agreement, k=0.53), supporting the feasibility of standardized coding of complex psychosocial narratives. Transformer-based and LLM approaches performed well in identifying domain-level SDoH (Level 1), while the structured LLM pipeline improved contextual classification (Level 2), particularly for attributes such as experiencer and severity. The pipeline generated a structured dataset capturing both the presence and context of SDoH factors, expanding the granularity of psychosocial variables available for analysis.

Conclusions and Implications
Social work clinical text can serve as high-value data sources for SDoH research using ontology-driven extraction methods. By transforming psychosocial assessment text into structured variables, this pipeline enables opportunities for predictive modeling, clinical decision support, and AI-enhanced assessment tools. Importantly, social workers can not only provide care but generate critical data infrastructure for health research. Although developed within pediatric transplantation, the framework is broadly applicable across social work domains where narrative assessment is central. Ontology-based SDoH extraction pipelines offer a scalable pathway to standardize psychosocial knowledge and integrate social work data into interdisciplinary research and practice.