Prediction of acute postoperative pain based on intraoperative nociception level (NOL) index values: the impact of machine learning-based analysis

Abstract

The relationship between intraoperative nociception and acute postoperative pain is still not well established. The nociception level (NOL) Index (Medasense, Ramat Gan, Israel) uses a multiparametric approach to provide a 0-100 nociception score. The objective of the ancillary analysis of the NOLGYN study was to evaluate the ability of a machine-learning aglorithm to predict moderate to severe acute postoperative pain based on intraoperative NOL values. Our study uses the data from the NOLGYN study, a randomized controlled trial that evaluated the impact of NOL-guided intraoperative administration of fentanyl on overall fentanyl consumption compared to standard of care. Seventy patients (ASA class I-III, aged 18-75 years) scheduled for gynecological laparoscopic surgery were enrolled. Variables included baseline demographics, NOL reaction to incision or intubation, median NOL during surgery, NOL time-weighted average (TWA) above or under manufacturers' recommended thresholds (10-25), and percentage of surgical time spent with NOL > 25 or < 10. We evaluated different machine learning algorithms to predict postoperative pain. Performance was assessed using cross-validated area under the ROC curve (CV-AUC). Of the 66 patients analyzed, 42 (63.6%) experienced moderate to severe pain. NOL post-intubation (42.8 (31.8-50.6) vs. 34.8 (25.6-41.3), p = 0.05), median NOL during surgery (13 (11-15) vs. 11 (8-13), p = 0.027), percentage of surgical time spent with NOL > 25 (23% (18-18) vs. 20% (15-24), p = 0.036), NOL TWA < 10 (2.54 (2.1-3.0) vs. 2.86 (2.48-3.62), p = 0.044) and percentage of surgical time spent with NOL < 10 (41% (36-47) vs. 47% (40-55), p = 0.022) were associated with moderate to severe PACU pain. Corresponding ROC AUC for the prediction of moderate to severe PACU pain were 0.65 [0.51-0.79], 0.66 [0.52-0.81], 0.66 [0.52-0.79], 0.65 [0.51-0.79] and 0.67 [0.53-0.81]. Penalized logistic regression achieved the best performance with a 0.753 (0.718-0.788) CV-AUC. Our results, even if limited by the small number of patients, suggest that acute postoperative pain is better predicted by a multivariate machine-learning algorithm rather than individual intraoperative nociception variables. Further larger multicentric trials are highly recommended to better understand the relationship between intraoperative nociception and acute postoperative pain.Trial registration Registered on ClinicalTrials.gov in October 2018 (NCT03776838).

Keywords: Acute postoperative pain; Machine learning; Nociception level index; Nociception monitoring; Prediction.