Frontiers in medicine · 2026
Preoperative risk stratification for postoperative inflammatory-nutritional deterioration after colorectal cancer surgery: an interpretable machine learning study.
СтатьяНазвание и аннотация: Перевод пока не подготовлен
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Zhang Yinchao, Hu Yingkun, Cao Xinke, Zan Xinquan, Tian Zhiqiang, Shen Wei
Аннотация
BACKGROUND: Postoperative inflammatory-nutritional deterioration may adversely affect recovery after colorectal cancer surgery. This study aimed to develop and internally validate an interpretable machine learning model for preoperative prediction of postoperative modified Glasgow Prognostic Score (mGPS) = 2.
METHODS: This single-centre retrospective study included consecutive patients who underwent elective radical resection for colorectal cancer between 2020 and 2025. The cohort was divided into training and internal validation sets using stratified sampling. Recursive feature elimination identified eight predictors: age, prealbumin, albumin, cholinesterase, lymphocyte percentage, platelet count, fibrinogen and serum calcium. Eight conventional machine learning algorithms, TabICLv2 and SuperLearner ensembles were evaluated using discrimination, calibration and decision curve analyses. SHapley Additive exPlanations were used for model interpretation.
RESULTS: Among 1196 patients, 507 (42.4%) developed postoperative mGPS=2. In the internal validation cohort, TabICLv2 achieved an area under the receiver operating characteristic curve of 0.753 (95% confidence interval, 0.701-0.804) and a Brier score of 0.197. SuperLearner-W and SuperLearner-SR showed comparable discrimination, with areas under the curve of 0.752 and 0.751, respectively. Lymphocyte percentage, platelet count and fibrinogen were the leading contributors to model predictions. Learning-curve analysis revealed suboptimal convergence behaviour, with relatively flat cross-validation trajectories and persistent separation from training performance across increasing training fractions. A web-based research prototype was developed to provide individualised risk estimates, patient-level feature attribution and an applicability-domain assessment indicating the level of support from the model-development data.
CONCLUSION: An interpretable model based on routinely available preoperative variables provided moderate discrimination for postoperative inflammatory-nutritional deterioration after colorectal cancer surgery. The suboptimal learning-curve convergence indicates that the model is not suitable for clinical deployment without further development and independent external validation.
METHODS: This single-centre retrospective study included consecutive patients who underwent elective radical resection for colorectal cancer between 2020 and 2025. The cohort was divided into training and internal validation sets using stratified sampling. Recursive feature elimination identified eight predictors: age, prealbumin, albumin, cholinesterase, lymphocyte percentage, platelet count, fibrinogen and serum calcium. Eight conventional machine learning algorithms, TabICLv2 and SuperLearner ensembles were evaluated using discrimination, calibration and decision curve analyses. SHapley Additive exPlanations were used for model interpretation.
RESULTS: Among 1196 patients, 507 (42.4%) developed postoperative mGPS=2. In the internal validation cohort, TabICLv2 achieved an area under the receiver operating characteristic curve of 0.753 (95% confidence interval, 0.701-0.804) and a Brier score of 0.197. SuperLearner-W and SuperLearner-SR showed comparable discrimination, with areas under the curve of 0.752 and 0.751, respectively. Lymphocyte percentage, platelet count and fibrinogen were the leading contributors to model predictions. Learning-curve analysis revealed suboptimal convergence behaviour, with relatively flat cross-validation trajectories and persistent separation from training performance across increasing training fractions. A web-based research prototype was developed to provide individualised risk estimates, patient-level feature attribution and an applicability-domain assessment indicating the level of support from the model-development data.
CONCLUSION: An interpretable model based on routinely available preoperative variables provided moderate discrimination for postoperative inflammatory-nutritional deterioration after colorectal cancer surgery. The suboptimal learning-curve convergence indicates that the model is not suitable for clinical deployment without further development and independent external validation.