Risk prediction models based on machine learning for emergence delirium in elderly patients undergoing spine surgery: development and validation study
Authors:
Open Access
- 26 June 2026
- journal article
- Published by Springer Nature in BMC Anesthesiology
- Vol. 26 (1)
Abstract
Emergence delirium (ED) is a common complication in elderly patients undergoing surgery for degenerative spinal disease (DSD) and is associated with adverse postoperative outcomes. Risk prediction may support perioperative risk stratification and individualized monitoring. This study aimed to develop a machine learning based model to predict ED risk in this population. This single-center prospective observational cohort study enrolled 430 elderly patients with DSD undergoing spinal surgery at Shanghai Changzheng Hospital between January and October 2024 using convenience sampling. Eligible patients were screened and consented before surgery, and predictors were collected prospectively across five perioperative domains: patient characteristics, preoperative, intraoperative, anesthesia-related, and early postoperative factors. Emergence delirium was assessed as the outcome in the PACU. Feature selection was performed using least absolute shrinkage and selection operator (LASSO) regression. Six machine learning (ML) models were developed and internally validated, and model interpretability was assessed using SHapley Additive exPlanations (SHAP). Among 38 features examined, LASSO regression identified nine key predictors, including age, education level, postoperative pain, and midazolam exposure, among others. Among six ML models evaluated in the testing set, logistic regression achieved the best overall performance, with an area under the curve of 0.832, AUPRC of 0.693, accuracy of 0.769, specificity of 0.858, sensitivity of 0.610, F1-score of 0.650, and Brier score of 0.163. Global feature importance ranking was determined using the SHAP summary plot, showing that the most influential features were education, age, and Midazolam. Individual feature importance, as depicted by the SHAP beeswarm plot, indicated that intraoperative bleeding, operation duration, and age had the most significant impact on prediction. In addition, a potential synergistic pattern between operation duration and intraoperative blood loss indicated a synergistic increase in emergence delirium risk when both were elevated. An interpretable logistic regression model showed acceptable overall performance for predicting emergence delirium risk in elderly patients undergoing spinal surgery. The model may support anesthesia-informed perioperative risk stratification and individualized monitoring; however, given its moderate sensitivity, it should be used as an adjunct to clinical judgment rather than as a stand-alone screening tool.Keywords
Funding Information
- the National Social Science Foundation of China (24CRK009)
- Social Science Cultivation Project of Naval Medical University (2023SK026)
This publication has 29 references indexed in Scilit:
- Artificial Intelligence and Machine Learning in Clinical MedicineNew England Journal of Medicine, 2023
- Machine Learning: Algorithms, Real-World Applications and Research DirectionsSN Computer Science, 2021
- Prevalence of spine degeneration diagnosis by type, age, gender, and obesity using Medicare dataScientific Reports, 2021
- Subjective and Objective Measures of Symptoms, Function, and Outcome in Patients With Degenerative Spine DiseaseArthritis Care & Research, 2020
- Risk factors for postoperative delirium after spinal surgery: a systematic review and meta-analysisAging Clinical and Experimental Research, 2019
- Impact of timing and duration of postoperative delirium: a retrospective observational studySurgery, 2018
- A simple frailty questionnaire (FRAIL) predicts outcomes in middle aged African AmericansThe Journal of nutrition, health and aging, 2012
- Comparison of Data Structures for Computing Formal ConceptsLecture Notes in Computer Science, 2009
- Geriatric Nutritional Risk Index: a new index for evaluating at-risk elderly medical patientsThe American Journal of Clinical Nutrition, 2005
- Clinical importance of changes in chronic pain intensity measured on an 11-point numerical pain rating scalePain, 2001