Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Social Determinants of Health Prediction for ICD-9 Code with Reasoning Models

Sharim Khan, Paul Landes, Adam Cross, Jimeng Sun

TL;DR

This study investigates admission-level, multi-label prediction of eight ICD-9 V-codes for social determinants of health from unstructured clinical notes in MIMIC-III using reasoning language models. It compares frontier, open-source, and traditional LLMs under a few-shot prompting regime, including a novel amended-vs-non-amended evaluation to capture coding gaps between documentation and physician coding. GPT-5-mini achieves the best results on amended codes (72.2% exact, 89.1% micro-F1), with open-source models performing competitively and traditional models lagging, highlighting the value of reasoning capabilities for long, complex clinical notes. The work underscores a potential to bridge documentation and coding gaps to improve care and population health, while noting limitations in generalizability and annotation scope.

Abstract

Social Determinants of Health correlate with patient outcomes but are rarely captured in structured data. Recent attention has been given to automatically extracting these markers from clinical text to supplement diagnostic systems with knowledge of patients' social circumstances. Large language models demonstrate strong performance in identifying Social Determinants of Health labels from sentences. However, prediction in large admissions or longitudinal notes is challenging given long distance dependencies. In this paper, we explore hospital admission multi-label Social Determinants of Health ICD-9 code classification on the MIMIC-III dataset using reasoning models and traditional large language models. We exploit existing ICD-9 codes for prediction on admissions, which achieved an 89% F1. Our contributions include our findings, missing SDoH codes in 139 admissions, and code to reproduce the results.

Social Determinants of Health Prediction for ICD-9 Code with Reasoning Models

TL;DR

This study investigates admission-level, multi-label prediction of eight ICD-9 V-codes for social determinants of health from unstructured clinical notes in MIMIC-III using reasoning language models. It compares frontier, open-source, and traditional LLMs under a few-shot prompting regime, including a novel amended-vs-non-amended evaluation to capture coding gaps between documentation and physician coding. GPT-5-mini achieves the best results on amended codes (72.2% exact, 89.1% micro-F1), with open-source models performing competitively and traditional models lagging, highlighting the value of reasoning capabilities for long, complex clinical notes. The work underscores a potential to bridge documentation and coding gaps to improve care and population health, while noting limitations in generalizability and annotation scope.

Abstract

Social Determinants of Health correlate with patient outcomes but are rarely captured in structured data. Recent attention has been given to automatically extracting these markers from clinical text to supplement diagnostic systems with knowledge of patients' social circumstances. Large language models demonstrate strong performance in identifying Social Determinants of Health labels from sentences. However, prediction in large admissions or longitudinal notes is challenging given long distance dependencies. In this paper, we explore hospital admission multi-label Social Determinants of Health ICD-9 code classification on the MIMIC-III dataset using reasoning models and traditional large language models. We exploit existing ICD-9 codes for prediction on admissions, which achieved an 89% F1. Our contributions include our findings, missing SDoH codes in 139 admissions, and code to reproduce the results.
Paper Structure (19 sections, 3 figures, 5 tables)

This paper contains 19 sections, 3 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Dataset
  4. Experimental Setup
  5. Results
  6. Discussion
  7. Conclusion
  8. Limitations
  9. Acknowledgments
  10. Final Prompt for sdoh Coding
  11. Annotation Guide for sdoh ICD-9 V-codes
  12. V60.0: Homelessness
  13. V60.2: Inadequate Material Resources
  14. V60.4: No other household member able to render care
  15. V62.0: Unemployment
  16. ...and 4 more sections

Figures (3)

  • Figure 1: Free text clinical notes, and ICD-9 V-codes, are used as gold labels for prediction and to identify sdoh marked admissions.
  • Figure 2: Final optimized prompt used for admission-level sdoh ICD-9 V-code prediction. The prompt encodes hierarchy rules, precision-first criteria, negative examples, and explicit output formatting. This full text is provided to ensure reproducibility.
  • Figure 3: Annotation guide used to label sdoh circumstances in clinical text for ICD-9 V-code prediction. Each section defines the code, domain, inclusion and exclusion criteria, and examples. Annotators must use precision-first judgment and follow hierarchy rules as outlined.