Extracting social determinants of health from dental records: feasibility study

Understanding social determinants of health (SDoH) in dental patients is important for supporting population health and clinical decision making but it remains unclear how well these are recorded in electronic dental records (EDRs) and if SDoH can be extracted from clinical notes. We retrospectively examined 100 clinical notes from 112,881 records of patient encounters that included an examination item in the Nepean Local Health District (LHD), with ethics approval from the LHD. To create gold standard annotation for a rule-based natural language processing (NLP) system, we applied expert domain knowledge, WordNet and word2vec embeddings, and the Unified Modelling Language System to identify 20,168 records that were more likely to include SDoH. Of these, we randomly sampled 100 records and 5 dental students annotated them using a schema adapted from the World Health Organization categorisation of SDoH. The 100 records were from 100 unique patients, 44% of these clinical encounters included a comprehensive oral examination. The most common SDoH was available for sugar (67%), tobacco (24.5%), alcohol (4.3%), housing (3.2%), and employment (1%), and 37% of clinical notes had no SDoH information available even though they included concepts that suggested SDoH. Information was complete for 9.5% of records describing sugar consumption, and 7.9% for tobacco, but complete details were rarely described for alcohol and housing. Clinical notes often include SDoH but it is important to understand the limits of what can be extracted to inform the feasibility of population health studies, and the design of technologies that might support improved consistency and completeness of SDoH in EDRs.