Rhee Wounsuk, Chang Sam Yeol, Chang Bong-Soon, Kim Hyoungmin
Ministry of Health and Welfare, Government of the Republic of Korea, 13, Doum 4-ro, Sejong, 30113, Republic of Korea.
Siebel School of Computing and Data Science, University of Illinois Urbana-Champaign, 201 N. Goodwin Avenue, Urbana, IL, 61801, USA.
BMC Med Inform Decis Mak. 2025 Jul 28;25(1):279. doi: 10.1186/s12911-025-03125-1.
Lumbar spinal stenosis is one of the most common surgery-requiring conditions of the spine in the aged population. As elderly patients often present with multiple comorbidities and limited physiological reserve, individualized risk assessment using comprehensive geriatric assessment is crucial for optimizing surgical outcomes.
Patients 65 years or older who underwent elective surgery for lumbar spinal stenosis between June 2015 and December 2018 were prospectively enrolled, resulting in 261 eligible patients of age 72.3 ± 4.8 years (male 108, female 153). Twenty-seven experienced complications of Clavien-Dindo grade 2 or higher within 30 days, and 79 received transfusion during hospital stay. The cohort was split into train-validation (n = 208) and test (n = 53) sets. A total of 48 features, including demographics, comorbidity, nutrition, and perioperative status, were collected. Logistic regression, support vector machine (SVM), random forest, XGBoost, and LightGBM were trained using five-fold cross-validation. AUROC and AUPRC were considered the primary performance metrics, and the results were compared with those estimated with ACS-NSQIP scoring system. A set of Compact models incorporating a smaller number of features was also trained, and SHAP analysis was conducted to evaluate the models' interpretability.
The reduced number of features did not result in the drop of AUROC and AUPRC for all machine learning algorithms (P > 0.05). when compared to the ACS-NSQIP scoring system, which achieved a test AUROC of 0.38 (95% confidence interval [CI], 0.13-0.73) and 0.22 (95% CI, 0.10-0.36) on the first two tasks, the Compact model showed significantly greater AUROC values nearing or surpassing 0.90. Decision tree-based algorithms demonstrated larger test AUROC than logistic regression and generally agreed on the most influential features for each task.
Advanced machine learning models have consistently shown greater performance and interpretability over conventional methodologies, implying their potential for a more individualized risk assessment of the aged population.
Not applicable as this research is not a clinical trial.
腰椎管狭窄症是老年人群中最常见的需要手术治疗的脊柱疾病之一。由于老年患者常伴有多种合并症且生理储备有限，使用综合老年评估进行个体化风险评估对于优化手术效果至关重要。
前瞻性纳入2015年6月至2018年12月期间接受择期腰椎管狭窄症手术的65岁及以上患者，共261例符合条件的患者，年龄为72.3±4.8岁（男性108例，女性153例）。27例在30天内发生Clavien-Dindo 2级或更高等级的并发症，79例在住院期间接受了输血。将该队列分为训练验证组（n = 208）和测试组（n = 53）。共收集了48项特征，包括人口统计学、合并症、营养状况和围手术期状态。使用五折交叉验证对逻辑回归、支持向量机（SVM）、随机森林、XGBoost和LightGBM进行训练。将曲线下面积（AUROC）和精确召回率曲线下面积（AUPRC）作为主要性能指标，并将结果与使用美国外科医师学会国家外科质量改进计划（ACS-NSQIP）评分系统估计的结果进行比较。还训练了一组包含较少特征的精简模型，并进行SHAP分析以评估模型的可解释性。
对于所有机器学习算法，特征数量的减少并未导致AUROC和AUPRC下降（P>0.05）。与ACS-NSQIP评分系统相比，该系统在前两项任务中的测试AUROC分别为0.38（95%置信区间[CI]，0.13 - 0.73）和0.22（95%CI，0.10 - 0.36），精简模型的AUROC值显著更高，接近或超过0.90。基于决策树的算法显示出比逻辑回归更大的测试AUROC，并且在每个任务的最具影响力特征上总体一致。
先进的机器学习模型始终表现出比传统方法更好的性能和可解释性，这意味着它们在对老年人群进行更个体化风险评估方面具有潜力。
本研究并非临床试验，因此不适用。
