ICU Pain Risk Estimation from Bedside Signals

We are researching a model that estimates whether a nurse-assessed pain score (CPOT) is elevated, using only the biosignals that ICU bedside monitors record automatically. The work is joint research with Tohoku University, conducted under ethics approval from Tohoku University Hospital (approval number 2025-1-395).

Background

Many patients in intensive care units experience severe pain. Behavioral instruments such as the Critical-Care Pain Observation Tool (CPOT) are widely used, but they require nurses and physicians to repeatedly observe and score patients at the bedside — a labor-intensive process in which delayed assessment can mean delayed treatment. If signs of elevated pain scores could be estimated from automatically recorded signals such as heart rate and blood pressure, the system could help prioritize reassessment and reduce the burden on clinical staff.

What we did

The research has two pillars.

Curating an ML-ready clinical dataset. From records collected in the Tohoku University Hospital ICU, we built a dataset that precisely aligns 23,092 CPOT assessments from 891 patients with the preceding 180 minutes of bedside signals — 16 signal types including heart rate, blood pressure and respiratory rate, sampled at 1-minute intervals — totaling more than 60,000 hours of recordings. Public ICU databases rarely offer this level of temporal alignment between pain scores and signals, making a minute-synchronized dataset of this scale a valuable foundation for time-series learning.

A model built on physiological grouping. The model divides the 16 signals into three physiological groups — circulatory, respiratory and body temperature — extracts shared features per group, and integrates them with a Transformer. Because it uses no manually assessed inputs such as age, sex or sedation scores (RASS), inference is fully automated. In evaluation it achieved the highest AUROC (0.7124) among the compared models. Removing the grouping lowered accuracy, and removing the circulatory group had the largest impact — consistent with the clinical observation that pain is accompanied by autonomic and hemodynamic changes.

Positioning and next steps

This is a retrospective study, and the model output should be treated as a score rather than a calibrated probability. The model is not a diagnostic tool for pain; it aims to help identify patients who may warrant bedside reassessment. External validation, calibration and prospective workflow evaluation are required before clinical use. It is one of our ongoing efforts in the medical and healthcare domain, building AI research on biosignals together with medical institutions.

Links

• Study information: Tohoku University Hospital research disclosure (approval 2025-1-395)