TL;DR
This paper introduces a novel computational approach that integrates social determinants of health (SDoH) from electronic health records (EHR) and scientific literature into a comprehensive knowledge graph (KG). By combining this SDoH-enriched KG with existing biological knowledge, the researchers significantly improve the prediction of Alzheimer's disease (AD)-related biological relationships, offering new insights into the complex interplay between social factors and disease progression.
What Problem Does This Paper Address?
Alzheimer's disease (AD) is a complex neurodegenerative disorder with a significant and growing global impact. While biological factors are well-studied, the influence of social determinants of health (SDoH) – conditions like socioeconomic status, education, and environment – on AD risk and progression is increasingly recognized but poorly understood. A major challenge in studying SDoH in AD is the fragmented nature of this information, scattered across diverse sources like EHRs (often unstructured) and scientific literature. Manually collecting and integrating this data is impractical. This paper aims to bridge this gap by developing a computational framework to systematically extract, integrate, and analyze SDoH information alongside biological knowledge to uncover novel insights into AD etiology.
Methods
The core of this research lies in constructing a novel knowledge graph, "SDoHenPKG," which integrates SDoH information with existing biological knowledge. The process involves several key steps:
- SDoH Entity Extraction: Large language models (LLMs), specifically GPT-40, are employed to extract SDoH entities from both literature abstracts and EHR discharge summaries. To address privacy concerns with EHR data, entities are first aggregated and then classified by the LLM into predefined SDoH categories.
- Knowledge Graph Construction:
- Literature KG: PubMed abstracts related to AD and SDoH are processed using Pubtator 3.0 for biomedical entity recognition and then analyzed with GPT-40 and co-occurrence analysis to extract SDoH-related triplets (entity-relation-entity).
- MIMIC KG: Patient-level SDoH information is extracted from MIMIC-III EHR data, including discharge summaries, diagnoses, and prescriptions.
- SDoH KG: The Literature KG and MIMIC KG are integrated to form a comprehensive SDoH KG.
- SDoHenPKG: This SDoH KG is then integrated with PrimeKG, a widely-used biological KG, to create the final SDoHenPKG. This integration enriches PrimeKG with SDoH context.
- Graph-Based Knowledge Discovery: Graph convolutional networks (GCNs) are used for link prediction tasks on both PrimeKG and SDoHenPKG. This allows for the inference of new relationships within the graphs. Three types of link prediction tasks are performed:
- Random Mask Task: Evaluates general link prediction performance.
- Targeted Genes Task: Assesses the recovery of known AD-related gene-gene relationships.
- Exploratory Prediction Task: Focuses on predicting novel SDoH-gene interactions.
Datasets
- Literature Data: PubMed abstracts (2019-2024) related to Alzheimer's Disease and SDoH.
- EHR Data: MIMIC-III database, specifically data from 686 patients diagnosed with AD or related conditions.
- Biological Data for Validation:
- Proteomics Data: From the AMP-AD Diverse Cohorts and ROSMAP datasets.
- Single-cell RNA-seq Data: From the Seattle Alzheimer's Disease Brain Cell Atlas (SEA-AD).
The paper does not introduce a new benchmark dataset but rather leverages existing, large-scale datasets to build and validate its novel knowledge graph.
Key Insights and Ideas
- SDoH as a Crucial Contextual Layer: The study strongly suggests that integrating SDoH information significantly enhances the predictive power of biomedical knowledge graphs. This implies that social factors are not merely correlational but play a vital role in the underlying biological mechanisms of AD.
- LLMs for SDoH Extraction are Powerful but Require Careful Handling: The use of LLMs like GPT-40 demonstrates their capability in extracting complex SDoH information from unstructured text. However, the paper highlights the importance of privacy-preserving techniques when dealing with sensitive EHR data.
- Heterogeneous Data Integration is Key: Combining information from diverse sources like scientific literature and clinical EHRs, and then further integrating it with biological knowledge, provides a more holistic and robust understanding of complex diseases like AD.
- Novel SDoH-Gene Interactions Uncovered: The exploratory prediction task successfully identified potential novel interactions between SDoH and genes implicated in AD, which were then partially validated through biological data. This opens new avenues for research into the mechanisms linking social environments to AD pathology.
Key Results
- Improved Link Prediction: The SDoHenPKG consistently outperformed the baseline PrimeKG across various link prediction tasks. For the random mask task, SDoHenPKG showed statistically significant improvements in MRR (Mean Reciprocal Rank) for 4 out of 7 relations.
- Enhanced Recovery of AD Genes: In the targeted gene mask task, SDoHenPKG achieved higher MRR values for all 23 selected AD-associated genes, with 6 genes showing statistically significant improvements after multiple testing corrections.
- Identification of Potential SDoH-Gene Links: The exploratory prediction task identified 758 genes potentially influenced by SDoH and related to AD.
- Biological Validation: Proteomics analysis revealed differential protein expression in 59 (AMP-AD) and 42 (ROSMAP) proteins, with 20 showing consensus differential expression. Subsequent RNA-seq analysis provided further evidence for these genes, highlighting specific cell-type expression patterns and potential roles in AD pathology, such as the downregulation of SST.
Limitations and Future Work
- Correlational Nature of Findings: The study acknowledges that the gene-SDoH connections inferred are correlational and do not establish causality.
- Confounding Factors: Biological and environmental variables could confound the observed relationships.
- Need for In Vitro and Clinical Validation: The authors suggest that in vitro functional tests (gene knockdown/overexpression) and clinical validation are necessary to confirm causality and real-world evidence.
- Generalizability: Further validation on independent datasets is needed to confirm the generalizability of the findings.
- Demographic Data Limitations: Incomplete or imbalanced demographic data in the datasets could skew analyses. Future work should explore imputation techniques and validation on diverse datasets.
- LLM Backbone: While GPT-40 proved effective, exploring more recent and powerful domain-specific LLMs could further enhance the framework.
- Integration Challenges: Harmonizing data from heterogeneous sources and ensuring data quality remains a challenge. Future work could explore improved harmonization methods.
- Incorporating Other AD KGs: The paper notes that other large AD KGs, like AlzKB, have yet to incorporate SDoH information, presenting an opportunity for future integration.
Summary updated