An increase in crop intensity could improve crop yield but may also lead to a series of environmental problems, such as depletion of ground water and increased soil salinity. The generation of high resolution(30 m) cr...An increase in crop intensity could improve crop yield but may also lead to a series of environmental problems, such as depletion of ground water and increased soil salinity. The generation of high resolution(30 m) crop intensity maps is an important method used to monitor these changes, but this is challenging because the temporal resolution of the 30-m image time series is low due to the long satellite revisit period and high cloud coverage. The recently launched Sentinel-2 satellite could provide optical images at 10–60 m resolution and thus improve the temporal resolution of the 30-m image time series. This study used harmonized Landsat Sentinel-2(HLS) data to identify crop intensity. The sixth polynomial function was used to fit the normalized difference vegetation index(NDVI) and enhanced vegetation index(EVI) curves. Then, 15-day NDVI and EVI time series were then generated from the fitted curves and used to generate the extent of croplands. Lastly, the first derivative of the fitted VI curves were used to calculate the VI peaks;spurious peaks were removed using artificially defined thresholds and crop intensity was generated by counting the number of remaining VI peaks. The proposed methods were tested in four study regions, with results showing that 15-day time series generated from the fitted curves could accurately identify cropland extent. Overall accuracy of cropland identification was higher than 95%. In addition, both the harmonized NDVI and EVI time series identified crop intensity accurately as the overall accuracies, producer’s accuracies and user’s accuracies of non-cropland, single crop cycle and double crop cycle were higher than 85%. NDVI outperformed EVI as identifying double crop cycle fields more accurately.展开更多
The purpose of this study was to compare the dose distribution of intensity-modulated ra- diotherapy （IMRT） in 7 and 5 fields as well as 3-D conformal radiotherapy （3D-CRT） plans for gastric cancer using dosimetri...The purpose of this study was to compare the dose distribution of intensity-modulated ra- diotherapy （IMRT） in 7 and 5 fields as well as 3-D conformal radiotherapy （3D-CRT） plans for gastric cancer using dosimetric analysis. In 15 patients with gastric cancer after D1 resection, dosimetric pa- rameters for IMRT （7 and 5 fields） and 3D-CRT were calculated with a total dose of 45 Gy （1.8 Gy/day） These parameters included the conformal index （CI）, homogeneity index （HI）, maximum dose spot for the planned target volume （PTV）, dose-volume histogram （DVH） and dose distribution in the organs at risk （OAR）, mean dose （Dmean）, maximal dose （Dmax） in the spinal cord, percentage of the normal liver volume receiving more than 30 Gy （V30） and percentage of the normal kidney volume receiving more than 20 Gy （V20）. IMRT （7 and 5 fields） and 3D-CRT achieved the PTV coverage. However, IMRT presented significantly higher CI and HI values and lower maximum dose spot distribution than 3D-CRT （P=0.001）. For dose distribution of OAR, IMRT had a significantly lower Dmean and Dmax in spinal cord than 3D-CRT （P=-0.009）. There was no obvious difference in V30 of liver and V20 of kidney between IMRT and 3D-CRT, but 5-field IMRT showed lower Dmean in the normal liver than other two plans （P=0.001）. IMRT revealed favorable tumor coverage as compared to 3D-CRT and IMRT plans. Specifically, 5-field IMRT plan was superior to 3D-CRT in protecting the spinal cord and liver, but this superiority was not observed in the kidney. Further studies are needed to compare differences among the three approaches.展开更多
目的构建可预测心脏骤停患者住院期间死亡风险的机器学习模型,并对其进行解释。方法提取美国重症监护医学信息数据库Ⅳ(Medical Information Mart for Intensive Care databaseⅣ,MIMIC-Ⅳ)2.0中心脏骤停患者转入ICU 24 h内首次临床资...目的构建可预测心脏骤停患者住院期间死亡风险的机器学习模型,并对其进行解释。方法提取美国重症监护医学信息数据库Ⅳ(Medical Information Mart for Intensive Care databaseⅣ,MIMIC-Ⅳ)2.0中心脏骤停患者转入ICU 24 h内首次临床资料及住院期间转归,基于机器学习算法构建6种可预测心脏骤停患者院内死亡风险的模型,包括XGBoost模型、轻量级梯度提升机(light gradient boosting machine,LGBM)模型、决策树(decision tree,DT)模型、K近邻(K-nearest neighbor,KNN)模型、Logistic回归模型、随机森林(random forest,RF)模型。采用受试者操作特征(receiver operator characteristic,ROC)曲线、临床决策曲线及校准曲线对模型进行评价,并采用Shapley加性解释(Shapley additive explanation,SHAP)算法评估不同临床特征对最优模型的影响,以增加模型的可解释性。结果共1465例符合纳入与排除标准的心脏骤停患者入选本研究。其中住院期间存活773例、死亡692例。经筛选,共纳入82个临床特征用于机器学习模型构建。模型评价结果显示,相较于其余5种模型,LGBM模型预测心脏骤停患者院内死亡的曲线下面积(area under the curve,AUC)更高[0.834(95%CI:0.688~0.894)],且相对于Logistic回归模型、XGBoost模型,其对死亡风险的预测准确性更高(校准度:0.166),临床决策性能更优,整体性能最佳。SHAP算法分析显示,对LGBM模型输出结果影响最大的3个临床特征分别为格拉斯哥睁眼反应评分、碳酸氢盐水平、白细胞计数。结论基于大型公共医疗卫生数据库建立的可预测心脏骤停患者住院期间死亡风险的机器学习模型中,LGBM模型性能最优,其可辅助临床进行更高效的疾病管理和更精准的医疗干预。展开更多
基金supported by the China Postdoctoral Science Foundation (2017M620075 and BX201700286)the National Natural Science Foundation of China (NSFC-61661136006)
文摘An increase in crop intensity could improve crop yield but may also lead to a series of environmental problems, such as depletion of ground water and increased soil salinity. The generation of high resolution(30 m) crop intensity maps is an important method used to monitor these changes, but this is challenging because the temporal resolution of the 30-m image time series is low due to the long satellite revisit period and high cloud coverage. The recently launched Sentinel-2 satellite could provide optical images at 10–60 m resolution and thus improve the temporal resolution of the 30-m image time series. This study used harmonized Landsat Sentinel-2(HLS) data to identify crop intensity. The sixth polynomial function was used to fit the normalized difference vegetation index(NDVI) and enhanced vegetation index(EVI) curves. Then, 15-day NDVI and EVI time series were then generated from the fitted curves and used to generate the extent of croplands. Lastly, the first derivative of the fitted VI curves were used to calculate the VI peaks;spurious peaks were removed using artificially defined thresholds and crop intensity was generated by counting the number of remaining VI peaks. The proposed methods were tested in four study regions, with results showing that 15-day time series generated from the fitted curves could accurately identify cropland extent. Overall accuracy of cropland identification was higher than 95%. In addition, both the harmonized NDVI and EVI time series identified crop intensity accurately as the overall accuracies, producer’s accuracies and user’s accuracies of non-cropland, single crop cycle and double crop cycle were higher than 85%. NDVI outperformed EVI as identifying double crop cycle fields more accurately.
文摘The purpose of this study was to compare the dose distribution of intensity-modulated ra- diotherapy （IMRT） in 7 and 5 fields as well as 3-D conformal radiotherapy （3D-CRT） plans for gastric cancer using dosimetric analysis. In 15 patients with gastric cancer after D1 resection, dosimetric pa- rameters for IMRT （7 and 5 fields） and 3D-CRT were calculated with a total dose of 45 Gy （1.8 Gy/day） These parameters included the conformal index （CI）, homogeneity index （HI）, maximum dose spot for the planned target volume （PTV）, dose-volume histogram （DVH） and dose distribution in the organs at risk （OAR）, mean dose （Dmean）, maximal dose （Dmax） in the spinal cord, percentage of the normal liver volume receiving more than 30 Gy （V30） and percentage of the normal kidney volume receiving more than 20 Gy （V20）. IMRT （7 and 5 fields） and 3D-CRT achieved the PTV coverage. However, IMRT presented significantly higher CI and HI values and lower maximum dose spot distribution than 3D-CRT （P=0.001）. For dose distribution of OAR, IMRT had a significantly lower Dmean and Dmax in spinal cord than 3D-CRT （P=-0.009）. There was no obvious difference in V30 of liver and V20 of kidney between IMRT and 3D-CRT, but 5-field IMRT showed lower Dmean in the normal liver than other two plans （P=0.001）. IMRT revealed favorable tumor coverage as compared to 3D-CRT and IMRT plans. Specifically, 5-field IMRT plan was superior to 3D-CRT in protecting the spinal cord and liver, but this superiority was not observed in the kidney. Further studies are needed to compare differences among the three approaches.
文摘目的构建可预测心脏骤停患者住院期间死亡风险的机器学习模型,并对其进行解释。方法提取美国重症监护医学信息数据库Ⅳ(Medical Information Mart for Intensive Care databaseⅣ,MIMIC-Ⅳ)2.0中心脏骤停患者转入ICU 24 h内首次临床资料及住院期间转归,基于机器学习算法构建6种可预测心脏骤停患者院内死亡风险的模型,包括XGBoost模型、轻量级梯度提升机(light gradient boosting machine,LGBM)模型、决策树(decision tree,DT)模型、K近邻(K-nearest neighbor,KNN)模型、Logistic回归模型、随机森林(random forest,RF)模型。采用受试者操作特征(receiver operator characteristic,ROC)曲线、临床决策曲线及校准曲线对模型进行评价,并采用Shapley加性解释(Shapley additive explanation,SHAP)算法评估不同临床特征对最优模型的影响,以增加模型的可解释性。结果共1465例符合纳入与排除标准的心脏骤停患者入选本研究。其中住院期间存活773例、死亡692例。经筛选,共纳入82个临床特征用于机器学习模型构建。模型评价结果显示,相较于其余5种模型,LGBM模型预测心脏骤停患者院内死亡的曲线下面积(area under the curve,AUC)更高[0.834(95%CI:0.688~0.894)],且相对于Logistic回归模型、XGBoost模型,其对死亡风险的预测准确性更高(校准度:0.166),临床决策性能更优,整体性能最佳。SHAP算法分析显示,对LGBM模型输出结果影响最大的3个临床特征分别为格拉斯哥睁眼反应评分、碳酸氢盐水平、白细胞计数。结论基于大型公共医疗卫生数据库建立的可预测心脏骤停患者住院期间死亡风险的机器学习模型中,LGBM模型性能最优,其可辅助临床进行更高效的疾病管理和更精准的医疗干预。
