Leveraging Artificial Intelligence to Uncover Symptom Burden in Palliative Care: Analysis of Nonscheduled Visits Using a Phi-3 Small Language Model.

Purpose: This study aimed to differentiate nonscheduled visits (NSVs) in an outpatient palliative care setting that are driven by or accompanied by uncontrolled symptoms from those that are administrative or routine, such as prescription refills and examination readings. A small language model (SLM) was used to enhance the detection and management of symptoms, thus improving health care resource allocation.

Methods: A retrospective analysis was performed on 25,867 patient visits to an outpatient palliative care unit, including 7,036 NSVs. A stratified random sample of 384 NSVs was reviewed to determine the presence of symptoms, using physician audits as the gold standard. A Phi-3-based SLM was validated against these audits to assess its accuracy in detecting the symptoms. The validated SLM was then applied to the entire NSV data set to identify symptom patterns. Multivariate linear regression was used to analyze the association of age, cancer type, and insurance category with the presence of symptoms.

Results: SLM demonstrated high sensitivity (99.4%) and accuracy (95.3%) in identifying symptom-driven NSVs. The analysis revealed that 85.7% of the NSVs were driven by symptoms, indicating a significant hidden burden of unmanaged symptoms. The study found that certain demographic and clinical factors, including younger age groups and specific cancer types, were significantly associated with an increased symptom burden.

Conclusion: This study highlights the substantial burden of symptom-driven NSVs in palliative care and demonstrates the effectiveness of using a SLM to identify and manage symptoms. Implementing such models in clinical practice can improve patient care by optimizing the allocation of health care resources and tailoring interventions to the needs of patients with advanced illnesses.