Machine learning based quantitative pain assessment for the perioperative period.

Summary

Using photoplethysmography from 242 patients, an XGBoost-based model achieved AUROC 0.819 intraoperatively and 0.927 postoperatively for pain assessment, outperforming a commercial surgical pain index postoperatively. Interpretable features such as waveform skewness and diastolic phase rate (intraop) and systolic area/baseline fluctuation (postop) drove performance.

Key Findings

• XGBoost-based models achieved AUROC 0.819 (intraoperative) and 0.927 (postoperative) for pain detection.
• Outperformed a commercial surgical pain index postoperatively (0.927 vs 0.577 AUROC).
• Feature importance indicated waveform skewness and diastolic phase rate decrease (intraop) and systolic phase area/baseline fluctuation (postop) as key predictors.

Clinical Implications

PPG-based ML models could augment or replace proprietary nociception indices, enabling broader, cost-effective pain monitoring intraoperatively and in PACU. Integration into monitors may improve analgesic titration and reduce under/over-treatment.

Limitations

• Single-center dataset with no external validation; generalizability uncertain.
• Pain labels combined NRS and clinical criteria; potential labeling noise.

Future Directions

External, multicenter validation; integration into anesthesia workstations; prospective trials testing analgesic titration guided by the model vs standard care.