Computational identification of salient cancer care topics and themes in oncology social work notes

Social work notes provide a multifaceted narrative of a patient's cancer journey. Topic modeling is a computational technique that can be used to organize notes into thematic clusters and study topics in oncological social work at scale. In this study, we used the BERTopic framework to perform topic modeling of cancer social work notes and identify roles and themes frequently documented in oncological social work.

We collected 106 676 social work notes from 25 341 patients diagnosed with cancer at the University of California San Francisco. Notes were split at the sentence level, and BERTopic model parameters were optimized experimentally. For each BERTopic cluster, 2 annotators independently assigned labels to 250 randomly selected sentences and keyword sets, with disagreements adjudicated by a third reviewer. Given the same sentences and keywords, 3 large language models independently labelled each cluster. Finally, human-generated and large language model-generated labels were harmonized.

Topic modeling of social work notes revealed 15 topics, classified into 3 distinct subthemes: (1) social worker assessments of patient psychosocial needs and service provision; (2) patient-centric factors, experiences, and the impact of cancer diagnosis on lifestyle; and (3) family and caregiver experiences when caring for someone with cancer.

Topic modeling supplemented by large language model interpretation reveals complex psychosocial themes in patients receiving social work services during cancer care experience. We envision these methods supporting research that clusters oncological social work notes into thematic cohorts for qualitative studies on how social work care differs across patient strata and how topics are differentially distributed across cancer subtypes.