AI caught every spiritual pain note in 500 palliative nursing records but flagged many false ones

The cost was a high rate of false alarms. Specificity was 0.75 for Gemini and 0.61 for Copilot, and positive predictive value was 0.34 and 0.25, so most flagged records did not actually show spiritual pain. The authors say this could cause alert fatigue for clinicians and that the tool cannot work as a standalone diagnostic system.

Spiritual care is often under-provided, and the authors note that more non-specialists will need to address it as palliative care demand grows. They suggest a two-stage process, with AI screening followed by clinician verification, but say that needs testing. All results come from one hospital with pre-filtered records.

The team prompted both models in Japanese with a definition of spiritual pain from the SpiPas nursing assessment sheet plus example sentences. No model training was needed. Gemini and Copilot were used because they were accessible inside the Tohoku University network, which met personal information protection requirements.

For subdomains of spiritual pain, sensitivity ranged from 0.74 to 0.91 and positive predictive value from 0.11 to 0.57. In the error analysis of Gemini's 111 false positives, other causes included overinterpreting patients' words (15), interpreting physical symptoms as spiritual pain (15), and mistaking perception of prognosis (5) or psychological symptoms (2).

The authors say deteriorating function and delirium can trigger spiritual pain, so even human reviewers may struggle to separate them. They propose clearer prompts, stricter definitions and step-by-step prompting to cut errors, and suggest two-stage screening with clinician verification.

The authors list several limits: a single Japanese hospital, only two models, a small sample with multiple records per patient, and pre-filtering by search terms. They say these factors may have led to overestimated performance.