Detecting spiritual pain in electronic medical records using large language models

Spiritual pain is one dimension of total pain in terminally ill patients with cancer. Reducing the burden of assessing spiritual pain in electronic medical records is essential for improving the quality of spiritual care. Large language models may offer a solution. This study explored whether large language models can be used to develop an automatic system for accurately identifying spiritual pain in patients with terminal cancer.

We used unstructured nursing records from patients with cancer who died in an inpatient palliative care unit at a university hospital in Japan between July and December 2024. A total of 500 records generated within 3 weeks before patients' death were randomly selected for analysis. Records were labeled for the presence of spiritual pain. Gemini-2.5-flash and Microsoft Copilot, accessible within the Tohoku University network environment, were utilized. Prompts included definitions of spiritual pain and example sentences.

The analyzed records were derived from 109 patients with cancer (57 (52.3%) male; mean age, 70.9 years). Manual review identified spiritual pain in 56 records (11.2%). Sensitivity for Gemini and Copilot was 1.00; specificity was 0.75 and 0.61; positive predictive value was 0.34 and 0.25; and negative predictive value was 1.00 and 1.00, respectively. Error analysis of false positives indicated that the models frequently misclassified decreased physical functions and delirium as spiritual pain.

This study demonstrates that large language models can identify spiritual pain with performance comparable to, or exceeding, that of previous studies utilizing classical natural language processing methods.