H S Tan1, N Liu2, R Sultana3, N-L R Han4, C W Tan1, J Zhang3, A T H Sia5, B L Sng6. 1. Department of Women's Anaesthesia, KK Women's and Children's Hospital, Singapore. 2. Duke-NUS Medical School, Singapore; Health Services Research Centre, Singapore Health Services, Singapore. 3. Duke-NUS Medical School, Singapore. 4. Division of Clinical Support Services, KK Women's and Children's Hospital, Singapore. 5. Department of Women's Anaesthesia, KK Women's and Children's Hospital, Singapore; Duke-NUS Medical School, Singapore. 6. Department of Women's Anaesthesia, KK Women's and Children's Hospital, Singapore; Duke-NUS Medical School, Singapore. Electronic address: sng.ban.leong@singhealth.com.sg.
Abstract
INTRODUCTION: Risk-prediction models for breakthrough pain facilitate interventions to forestall inadequate labour analgesia, but limited work has used machine learning to identify predictive factors. We compared the performance of machine learning and regression techniques in identifying parturients at increased risk of breakthrough pain during labour epidural analgesia. METHODS: A single-centre retrospective study involved parturients receiving patient-controlled epidural analgesia. The primary outcome was breakthrough pain. We randomly selected 80% of the cohort (training cohort) to develop three prediction models using random forest, XGBoost, and logistic regression, followed by validation against the remaining 20% of the cohort (validation cohort). Area-under-the-receiver operating characteristic curve (AUC), sensitivity, specificity, and positive and negative predictive values (PPV and NPV) were used to assess model performance. RESULTS: Data from 20 716 parturients were analysed. The incidence of breakthrough pain was 14.2%. Of 31 candidate variables, random forest, XGBoost and logistic regression models included 30, 23, and 15 variables, respectively. Unintended venous puncture, post-neuraxial analgesia highest pain score, number of dinoprostone suppositories, neuraxial technique, number of neuraxial attempts, depth to epidural space, body mass index, pre-neuraxial analgesia oxytocin infusion rate, maternal age, pre-neuraxial analgesia cervical dilation, anaesthesiologist rank, and multiparity, were identified in all three models. All three models performed similarly, with AUC 0.763-0.772, sensitivity 67.0-69.4%, specificity 70.9-76.2%, PPV 28.3-31.8%, and NPV 93.3-93.5%. CONCLUSIONS: Machine learning did not improve the prediction of breakthrough pain compared with multivariable regression. Larger population-wide studies are needed to improve predictive ability.
INTRODUCTION: Risk-prediction models for breakthrough pain facilitate interventions to forestall inadequate labour analgesia, but limited work has used machine learning to identify predictive factors. We compared the performance of machine learning and regression techniques in identifying parturients at increased risk of breakthrough pain during labour epidural analgesia. METHODS: A single-centre retrospective study involved parturients receiving patient-controlled epidural analgesia. The primary outcome was breakthrough pain. We randomly selected 80% of the cohort (training cohort) to develop three prediction models using random forest, XGBoost, and logistic regression, followed by validation against the remaining 20% of the cohort (validation cohort). Area-under-the-receiver operating characteristic curve (AUC), sensitivity, specificity, and positive and negative predictive values (PPV and NPV) were used to assess model performance. RESULTS: Data from 20 716 parturients were analysed. The incidence of breakthrough pain was 14.2%. Of 31 candidate variables, random forest, XGBoost and logistic regression models included 30, 23, and 15 variables, respectively. Unintended venous puncture, post-neuraxial analgesia highest pain score, number of dinoprostone suppositories, neuraxial technique, number of neuraxial attempts, depth to epidural space, body mass index, pre-neuraxial analgesia oxytocin infusion rate, maternal age, pre-neuraxial analgesia cervical dilation, anaesthesiologist rank, and multiparity, were identified in all three models. All three models performed similarly, with AUC 0.763-0.772, sensitivity 67.0-69.4%, specificity 70.9-76.2%, PPV 28.3-31.8%, and NPV 93.3-93.5%. CONCLUSIONS: Machine learning did not improve the prediction of breakthrough pain compared with multivariable regression. Larger population-wide studies are needed to improve predictive ability.
Authors: Nan Liu; Mingxuan Liu; Xinru Chen; Yilin Ning; Jin Wee Lee; Fahad Javaid Siddiqui; Seyed Ehsan Saffari; Andrew Fu Wah Ho; Sang Do Shin; Matthew Huei-Ming Ma; Hideharu Tanaka; Marcus Eng Hock Ong Journal: EClinicalMedicine Date: 2022-05-06