Leveraging natural language processing to identify social determinants of health for patients with Alzheimer’s disease and related dementias from electronic medical records

Abstract

BackgroundSocial determinants of health (SDoH) have a significant impact on the well-being of individuals with Alzheimer’s disease and related dementias (ADRD). ADRD patients from socioeconomically disadvantaged groups are generally subject to higher risks of adverse health events than those from advantaged groups. However, information related to SDoH is not collected systematically in most medical institutions. Unstructured electronic medical records (EMRs), including free-text clinical narratives and nurse and social worker notes, may include salient information that can be used to identify individuals who may need assistance. In this study, we sought to develop a rule-based, natural language processing (NLP) algorithm that identifies from patients’ unstructured EMRs their housing instability, transportation problems, food and medication insecurity, solitariness, and financial exploitation.MethodWe randomly selected 1,000 medical notes from Michigan Medicine between 2016 and 2019 and partitioned them into 750 notes for training and 250 notes for testing. Training notes were preprocessed by a prototype algorithm and annotated by registered nurses. We then trained the algorithm by resolving all human and machine disagreements. The post-training NLP algorithm was evaluated on the testing notes. The most promising medical notes that found information about SDoH were social worker notes.ResultThe NLP algorithm achieved 85.3% of note-wise accuracy. Determinant-wise accuracy scores were 88.6% (housing), 91.1% (transportation), 86.7% (food), 92.0% (medication), 87.4% (solitariness) and 95.3% (financial exploitation). Determinant-wise F1-scores were 90.6% (housing), 86.9% (transportation), 91.3% (food), 95.7% (medication), 82.3% (solitariness) and 93.8% (financial exploitation).ConclusionOur NLP algorithm is a stand-alone module that can run on real-time EMRs and extract information about SDoH with a high level of accuracy. Findings can be used by clinicians, care managers, and social workers to pragmatically and efficiently reach out to ADRD patients who may need assistance with food, transportation, or housing. In doing this, adverse health events may be reduced among this vulnerable patient population. Future efforts will focus on making this algorithm generalizable for use at other institutions.