基于集成学习技术对死亡时间的研究

Research on the post mortem interval based on ensemble learning technology

目的 利用集成多种机器学习算法对死亡时间进行推断。方法 收集近5年来中国西部某地546起死亡案例的尸体数据,利用KNN、隔离森林和SMOTE(Synthetic Minority Oversampling Technique)超采样算法对数据进行整理填充,开发出一种将决策树、随机森林和logistic回归集成在一起的机器学习算法,基于博弈论的sharpley指数对模型中重要影响因子进行比较,优化参数,进而推断死亡时间。结果 通过集成方法,在数据从不完整到完整的过程中,将决策树、随机森林和logistic回归三种准确度在70%~80%的弱分类器集成为准确度在99%以上的强分类器,通过17例案例的实际应用,误差在4 h以内的案例达41.18%。结论 利用集成算法可以有效的将多种机器学习算法集成为推断死亡时间准确度更高的算法,基于此建立的死亡时间推断模型可以应用于死亡时间推断。

Objective To estimate the post mortem interval(PMI) by integrating multiple machine learning algorithms. Methods Cadaver data from 546 death cases that happened in a western China locality over the past five years were collected, and the data were sorted and filled by over-sampling Algorithm, such as KNN, Isolated Forest and Synthetic Minority Oversampling Technique(SMOTE). A machine learning algorithm that integrates decision tree, random forest and logistic regression was developed to compare the important influencing factors in the model based on the Sharpley index,optimize the parameters, and thus infer the PMI. Results Through the integration method, three sets of weak classifiers with 70 % ~ 80 % accuracy, including decision trees, random forests and logistic regression, were combined together and turned into strong classifiers with over 99 % accuracy, when the incomplete data became complete. 41.18 % of the cases with an error of less than 4 hours were achieved through the application of 17 cases in practical fields. Conclusion The use of integration algorithm can effectively combine multiple machine learning algorithms into one with higher accuracy for PMI estimation, and the inference model based on this can be applied to PMI estimation in practice.