TL;DR
This study leverages Natural Language Processing (NLP) to extract Social Determinants of Health (SDOH) from unstructured electronic health records (EHRs) of US Veterans, revealing a significant association between these factors and suicide risk. By integrating NLP with traditional structured data, the research provides a more comprehensive understanding of SDOH's impact, highlighting "legal problems" and "violence" as particularly strong predictors of suicide.
What Problem Does This Paper Address?
Social Determinants of Health (SDOH) – factors like socioeconomic status, housing, and education – are known to be critical predictors of health outcomes, including suicidal behaviors. However, traditional research has often relied on structured data within EHRs, which can be incomplete or lack the nuance needed to capture the full spectrum of SDOH. This gap limits our ability to accurately assess risk and develop effective interventions. This paper aims to bridge this gap by using NLP to extract SDOH from unstructured clinical notes, providing a richer dataset for analyzing the association between SDOH and suicide risk in a large veteran population.
Methods
The researchers employed a nested case-control study design using data from the US Veterans Health Administration (VHA) Corporate Data Warehouse. A key innovation was the development and application of a Natural Language Processing (NLP) system, based on the RoBERTa language model, to extract SDOH from unstructured EHR notes. This NLP system was trained on annotated data and designed to identify both the presence and period of SDOH factors. The study then compared the associations of SDOH extracted from structured data, NLP-extracted data, and a combination of both with suicide death. Statistical analysis involved conditional logistic regression to estimate adjusted odds ratios (aORs) and 95% confidence intervals (CIs).
Datasets
The study utilized the Electronic Health Record (EHR) database from the US Veterans Health Administration (VHA). This database contains demographic information, diagnoses, medications, procedures, clinical notes, and billing data for millions of US Veterans. For training the NLP model, a subset of 4,646 EHR notes from 1,393 patients was annotated by expert annotators. The main study cohort comprised 6,122,785 Veterans who received care in the VHA between October 1, 2010, and September 30, 2015.
Key Insights and Ideas
- NLP unlocks hidden SDOH: Unstructured clinical notes contain a wealth of SDOH information that is often missed by structured data alone. The NLP system was able to extract a significant amount of SDOH that would otherwise be unavailable.
- Comprehensive SDOH is crucial: Combining structured and NLP-extracted SDOH provides a more complete picture of a veteran's social context and its impact on suicide risk.
- "Legal problems" and "violence" are potent predictors: These specific SDOH factors, identified through both structured and NLP methods, showed the strongest associations with increased suicide risk among veterans.
- Structured data might underrepresent risk in healthier populations: The study hypothesizes that structured data might be less likely to capture SDOH for individuals who are less ill, potentially leading to an underestimation of risk in certain subgroups. NLP's ability to extract SDOH from clinical notes could help mitigate this "information bias."
Key Results
- The study analyzed a cohort of 6,122,785 Veterans, identifying 8,821 suicide deaths over a mean follow-up of 3.87 years.
- NLP-extracted SDOH covered, on average, 80.03% of all SDOH occurrences, significantly outperforming structured data's coverage (38.17%).
- All SDOH factors, whether from structured data, NLP, or combined, were significantly associated with an increased risk of suicide.
- The SDOH with the largest effect sizes were:
- Legal problems: aOR = 2.66 (95% CI = 2.46-2.89)
- Violence: aOR = 2.12 (95% CI = 1.98-2.27)
- When considering simultaneous exposure to two SDOH, combinations like "legal problems" and "violence" showed even higher odds ratios (e.g., aOR = 3.44 for NLP-extracted).
Limitations and Future Work
The authors acknowledge several limitations:
- Generalizability: The study population is limited to VA users, which may not represent the general US population.
- Residual Confounding: Despite adjusting for numerous factors, there's always a possibility of unmeasured confounders influencing the results.
- EHR Data Completeness: While the NLP approach improves SDOH capture, EHR data can still be incomplete or missing.
Future work suggested by the authors includes:
- Exploring the potential of intervening on SDOH to change suicide risk, potentially through experimental or quasi-experimental designs.
- Investigating the time-varying nature of SDOH, as these factors can change over time and influence suicide risk differently across short-term and long-term periods.
- Developing methods for seamlessly integrating NLP-derived SDOH information into existing healthcare systems to aid in risk assessment and suicide prevention.
Summary updated