International journal of surgery (London, England) · 2025
Data imputation and domain-adaptive prediction of 1-year postoperative mortality in geriatric hip fracture patients following arthroplasty from multi-center study.
СтатьяМногоцентровое исследованиеНазвание и аннотация: Перевод в очереди
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.
Lu Xingchen, Zeng Cong, Gong Zhiwei, Jin Runze, Zhang Zhaoyuan, Liu Kesi, Hu Ning, Xiao Jingjing, Xiong Yan
Аннотация
BACKGROUND: Accurate postoperative mortality prediction of geriatric hip fracture patients is an important but challenging task. However, due to the influence factors, such as differences in medical care and patients' contraindications, the clinical data from different centers is often incomplete and has very poor consistency. Therefore, this paper aims at building a robust multi-center postoperative mortality prediction model prior to surgery even with aforementioned "dirty data" problems.
METHOD: A total of 744 geriatric hip fracture patients from three different medical centers were included in this study. A total of 22 features were collected which contained demographic characteristics, clinical investigations, and 1-year death events. Four statistical methods and four machine learning methods were utilized to fill in the missing laboratory indexes. Four non-adaptive methods and two domain-adaptive machine learning methods were selected to build the mortality prediction model. A 5-fold stratified cross-validation strategy was employed to ensure identical class distributions across splits, and all predictions were evaluated using bootstrap resampling with 1000 iterations.
RESULTS: Random Forest Imputer was the best model for data imputation, as the root mean square error and mean absolute deviation achieved (0.143, 0.098) for Chongqing Daping Hospital and Xinqiao Hospital (DXH), and (0.137, 0.097) for the First Affiliated Hospital (FAH) of Chongqing Medical University. For mortality prediction trained with each database and validated internally, the best possible model was XGB where the AUC could achieve 0.80 (95% CI: [0.69, 0.91]) for DXH and 0.83 (95% CI: [0.67, 0.94]) for FAH respectively. When external test was performed in another center, the domain-adaptive Nearest Neighbor-based Importance Weighting model achieved the best performance with AUC 0.69 (95% CI: [0.54, 0.84]) for DXH and 0.83 (95% CI: [0.70, 0.94]) for FAH.
CONCLUSION: The methods based on machine learning techniques were the most suited for the imputation of missing values. Using the domain adaptation methods, the built 1-year postoperative mortality prediction models for geriatric hip fracture patients could be applied in different medical centers with a desirable performance increase.
METHOD: A total of 744 geriatric hip fracture patients from three different medical centers were included in this study. A total of 22 features were collected which contained demographic characteristics, clinical investigations, and 1-year death events. Four statistical methods and four machine learning methods were utilized to fill in the missing laboratory indexes. Four non-adaptive methods and two domain-adaptive machine learning methods were selected to build the mortality prediction model. A 5-fold stratified cross-validation strategy was employed to ensure identical class distributions across splits, and all predictions were evaluated using bootstrap resampling with 1000 iterations.
RESULTS: Random Forest Imputer was the best model for data imputation, as the root mean square error and mean absolute deviation achieved (0.143, 0.098) for Chongqing Daping Hospital and Xinqiao Hospital (DXH), and (0.137, 0.097) for the First Affiliated Hospital (FAH) of Chongqing Medical University. For mortality prediction trained with each database and validated internally, the best possible model was XGB where the AUC could achieve 0.80 (95% CI: [0.69, 0.91]) for DXH and 0.83 (95% CI: [0.67, 0.94]) for FAH respectively. When external test was performed in another center, the domain-adaptive Nearest Neighbor-based Importance Weighting model achieved the best performance with AUC 0.69 (95% CI: [0.54, 0.84]) for DXH and 0.83 (95% CI: [0.70, 0.94]) for FAH.
CONCLUSION: The methods based on machine learning techniques were the most suited for the imputation of missing values. Using the domain adaptation methods, the built 1-year postoperative mortality prediction models for geriatric hip fracture patients could be applied in different medical centers with a desirable performance increase.