目的基于脓毒症患者院内死亡预测中因机器学习样本类别不平衡导致敏感度过低的问题,构建一种新的基于平衡随机森林(Balanced Random Forest,BRF)算法的预测模型。方法从MIMIC-Ⅲ公开数据库中获取符合脓毒症(Sepsis-3.0)标准患者的17个...目的基于脓毒症患者院内死亡预测中因机器学习样本类别不平衡导致敏感度过低的问题,构建一种新的基于平衡随机森林(Balanced Random Forest,BRF)算法的预测模型。方法从MIMIC-Ⅲ公开数据库中获取符合脓毒症(Sepsis-3.0)标准患者的17个时间序列变量数据,截取入住ICU后最初48 h的数据,计算出17个变量的714个统计特征,将其用于模型构建和性能评估。利用曲线下面积(Area Under Curve,AUC)和修正的几何平均值(Adjusted Geometric-Mean,AGM)进行超参调优。除BRF算法模型外,还与传统的逻辑回归和随机森林(Random Forest,RF)算法模型性能进行比较。结果最终筛选出10270例有ICU住院经历的脓毒症患者,院内总体死亡率为18.04%。各种模型性能测试结果表明,基于样本类别不平衡的机器学习模型的预测敏感度显著提高,其中RF-AGM模型最低,为0.1826(95%CI:0.1351~0.2322),BRF-AGM模型提高到0.7110(95%CI:0.6537~0.7677),利用新的预测模型,将会发现更多面临死亡的患者并及时给予救治。BRF-AGM模型的AUC、AGM和特异性分别达到了0.7994(95%CI:0.7696~0.8288)、0.7282(95%CI:0.7046~0.7519)和0.7349(95%CI:0.7101~0.7590)。结论BRF-AGM模型在ICU脓毒症患者死亡预测方面具有巨大应用潜力,可以避免临床医生延误治疗患者,这对改善患者预后具有重要意义,但BRF-AGM模型的临床效用还需要前瞻性多中心研究来进一步评估。展开更多
Due to the NP-hardness of the two-sided assembly line balancing(TALB) problem, multiple constraints existing in real applications are less studied, especially when one task is involved with several constraints. In thi...Due to the NP-hardness of the two-sided assembly line balancing(TALB) problem, multiple constraints existing in real applications are less studied, especially when one task is involved with several constraints. In this paper, an effective hybrid algorithm is proposed to address the TALB problem with multiple constraints(TALB-MC). Considering the discrete attribute of TALB-MC and the continuous attribute of the standard teaching-learning-based optimization(TLBO) algorithm, the random-keys method is hired in task permutation representation, for the purpose of bridging the gap between them. Subsequently, a special mechanism for handling multiple constraints is developed. In the mechanism, the directions constraint of each task is ensured by the direction check and adjustment. The zoning constraints and the synchronism constraints are satisfied by teasing out the hidden correlations among constraints. The positional constraint is allowed to be violated to some extent in decoding and punished in cost function. Finally, with the TLBO seeking for the global optimum, the variable neighborhood search(VNS) is further hybridized to extend the local search space. The experimental results show that the proposed hybrid algorithm outperforms the late acceptance hill-climbing algorithm(LAHC) for TALB-MC in most cases, especially for large-size problems with multiple constraints, and demonstrates well balance between the exploration and the exploitation. This research proposes an effective and efficient algorithm for solving TALB-MC problem by hybridizing the TLBO and VNS.展开更多
文摘目的基于脓毒症患者院内死亡预测中因机器学习样本类别不平衡导致敏感度过低的问题,构建一种新的基于平衡随机森林(Balanced Random Forest,BRF)算法的预测模型。方法从MIMIC-Ⅲ公开数据库中获取符合脓毒症(Sepsis-3.0)标准患者的17个时间序列变量数据,截取入住ICU后最初48 h的数据,计算出17个变量的714个统计特征,将其用于模型构建和性能评估。利用曲线下面积(Area Under Curve,AUC)和修正的几何平均值(Adjusted Geometric-Mean,AGM)进行超参调优。除BRF算法模型外,还与传统的逻辑回归和随机森林(Random Forest,RF)算法模型性能进行比较。结果最终筛选出10270例有ICU住院经历的脓毒症患者,院内总体死亡率为18.04%。各种模型性能测试结果表明,基于样本类别不平衡的机器学习模型的预测敏感度显著提高,其中RF-AGM模型最低,为0.1826(95%CI:0.1351~0.2322),BRF-AGM模型提高到0.7110(95%CI:0.6537~0.7677),利用新的预测模型,将会发现更多面临死亡的患者并及时给予救治。BRF-AGM模型的AUC、AGM和特异性分别达到了0.7994(95%CI:0.7696~0.8288)、0.7282(95%CI:0.7046~0.7519)和0.7349(95%CI:0.7101~0.7590)。结论BRF-AGM模型在ICU脓毒症患者死亡预测方面具有巨大应用潜力,可以避免临床医生延误治疗患者,这对改善患者预后具有重要意义,但BRF-AGM模型的临床效用还需要前瞻性多中心研究来进一步评估。
基金Supported by National Natural Science Foundation of China(Grant Nos.51275366,50875190,51305311)Specialized Research Fund for the Doctoral Program of Higher Education of China(Grant No.20134219110002)
文摘Due to the NP-hardness of the two-sided assembly line balancing(TALB) problem, multiple constraints existing in real applications are less studied, especially when one task is involved with several constraints. In this paper, an effective hybrid algorithm is proposed to address the TALB problem with multiple constraints(TALB-MC). Considering the discrete attribute of TALB-MC and the continuous attribute of the standard teaching-learning-based optimization(TLBO) algorithm, the random-keys method is hired in task permutation representation, for the purpose of bridging the gap between them. Subsequently, a special mechanism for handling multiple constraints is developed. In the mechanism, the directions constraint of each task is ensured by the direction check and adjustment. The zoning constraints and the synchronism constraints are satisfied by teasing out the hidden correlations among constraints. The positional constraint is allowed to be violated to some extent in decoding and punished in cost function. Finally, with the TLBO seeking for the global optimum, the variable neighborhood search(VNS) is further hybridized to extend the local search space. The experimental results show that the proposed hybrid algorithm outperforms the late acceptance hill-climbing algorithm(LAHC) for TALB-MC in most cases, especially for large-size problems with multiple constraints, and demonstrates well balance between the exploration and the exploitation. This research proposes an effective and efficient algorithm for solving TALB-MC problem by hybridizing the TLBO and VNS.