摘要
Acute Kidney Injury (AKI) is one of the most common acute and critical illnesses in general wards and intensive care units. Its high morbidity and high fatality rate have become a major global public health problem. There are often serious lags in clinical diagnosis of AKI. Early diagnosis and timely intervention and effective care become critical. The use of electronic medical record data to build an AKI risk prediction model has been proven to help prevent the occurrence of AKI. However, in actual clinical applications, the distribution of historical data and new data will continue to vary over time, resulting in a significant decrease in the performance of the model. How to solve the problem of model performance degradation over time will be a core challenge for the long-term use of predictive models in clinical applications. Aiming at the above problems, this paper studies the classic Transfer-Stacking model migration algorithm. Aiming at the lack of this algorithm, such as the loss of a large amount of feature information of the target domain and poor fit when integrating the model of the target domain, the Accumulate-Transfer-Stacking algorithm is proposed to improve it. Improvements include: 1) Optimize the input vector and model integration algorithm of Transfer-Stacking’s target domain model. 2) Optimize Transfer-Stacking from a single-source domain model to a multi-source domain model. The experimental results show that for the improved algorithm proposed in this paper when the data is sufficient and insufficient, the average AUC value of the model on the data of subsequent years is 0.89 and 0.87, and the average F1 Score value is 0.45 and 0.36. Moreover, this method is significantly better than the unimproved Transfer-Stacking algorithm and baseline method, and can effectively overcome the problem of data distribution heterogeneity caused by time factors.
Acute Kidney Injury (AKI) is one of the most common acute and critical illnesses in general wards and intensive care units. Its high morbidity and high fatality rate have become a major global public health problem. There are often serious lags in clinical diagnosis of AKI. Early diagnosis and timely intervention and effective care become critical. The use of electronic medical record data to build an AKI risk prediction model has been proven to help prevent the occurrence of AKI. However, in actual clinical applications, the distribution of historical data and new data will continue to vary over time, resulting in a significant decrease in the performance of the model. How to solve the problem of model performance degradation over time will be a core challenge for the long-term use of predictive models in clinical applications. Aiming at the above problems, this paper studies the classic Transfer-Stacking model migration algorithm. Aiming at the lack of this algorithm, such as the loss of a large amount of feature information of the target domain and poor fit when integrating the model of the target domain, the Accumulate-Transfer-Stacking algorithm is proposed to improve it. Improvements include: 1) Optimize the input vector and model integration algorithm of Transfer-Stacking’s target domain model. 2) Optimize Transfer-Stacking from a single-source domain model to a multi-source domain model. The experimental results show that for the improved algorithm proposed in this paper when the data is sufficient and insufficient, the average AUC value of the model on the data of subsequent years is 0.89 and 0.87, and the average F1 Score value is 0.45 and 0.36. Moreover, this method is significantly better than the unimproved Transfer-Stacking algorithm and baseline method, and can effectively overcome the problem of data distribution heterogeneity caused by time factors.
作者
Yunfei Xue
Yunfei Xue(Department of Computer Science, Jinan University, Guangzhou, China)
