Automatic Extraction of Social Determinants of Health from Medical Notes of Chronic Lower Back Pain Patients

Background. Adverse social determinants of health (SDoH), or social risk factors, such as food insecurity and housing instability, are known to contribute to poor health outcomes and inequities. Our ability to study these linkages is limited because SDoH information is more frequently documented in free-text clinical notes than structured data fields. To overcome this challenge, there is a growing push to develop techniques for automated extraction of SDoH. In this study, we explored natural language processing (NLP) and inference (NLI) methods to extract SDoH information from clinical notes of patients with chronic low back pain (cLBP), to enhance future analyses of the associations between SDoH and low back pain outcomes and disparities. Methods. Clinical notes (n=1,576) for patients with cLBP (n=386) were annotated for seven SDoH domains: housing, food, transportation, finances, insurance coverage, marital and partnership status, and other social support, resulting in 626 notes with at least one annotated entity for 364 patients. We additionally labelled pain scores, depression, and anxiety. We used a two-tier taxonomy with these 10 first-level ontological classes and 68 second-level ontological classes. We developed and validated extraction systems based on both rule-based and machine learning approaches. As a rule-based approach, we iteratively configured a clinical Text Analysis and Knowledge Extraction System (cTAKES) system. We trained two machine learning models (based on convolutional neural network (CNN) and RoBERTa transformer), and a hybrid system combining pattern matching and bag-of-words models. Additionally, we evaluated a RoBERTa based entailment model as an alternative technique of SDoH detection in clinical texts. We used a model previously trained on general domain data without additional training on our dataset. Results. Four annotators achieved high agreement (average kappa=95%, F1=91.20%). Annotation frequency varied significantly dependent on note type. By tuning cTAKES, we achieved a performance of F1=47.11% for first-level classes. For most classes, the machine learning RoBERTa-based NER model performed better (first-level F1=84.35%) than other models within the internal test dataset. The hybrid system on average performed slightly worse than the RoBERTa NER model (first-level F1=80.27%), matching or outperforming the former in terms of recall. Using an out-of-the-box entailment model, we detected many but not all challenging wordings missed by other models, reaching an average F1 of 76.04%, while matching and outperforming the tested NER models in several classes. Still, the entailment model may be sensitive to hypothesis wording and may require further fine tuning. Conclusion. This study developed a corpus of annotated clinical notes covering a broad spectrum of SDoH classes. This corpus provides a basis for training machine learning models and serves as a benchmark for predictive models for named entity recognition for SDoH and knowledge extraction from clinical texts.