论文部分内容阅读
目的:对比线性回归模型与四种机器学习算法对临床医学研究生学习行为的预测效能,探讨不同预测模型的优缺点和适用性。方法:以全国6 922名临床医学研究生的调查数据为例,通过自评学习行为量表获得综合得分;在训练集中,分别利用Lasso线性回归和人工神经网络、决策树、Bootstrap随机森林、提升树四种监督式机器学习算法建立预测模型;对验证集数据进行预测并比较不同模型的预测效能。结果:临床医学研究生学习行为综合得分为(3.31±0.54),总体达标率为74.02%。在线性回归模型中,年龄、学校级别、学位类型、学习兴趣、压力和满意度对学习行为的影响差异有统计学意义;在对验证集的预测中,线性回归模型的敏感度为0.484,特异度为0.914,准确率为0.801。四种机器学习算法的各项指标均高于线性回归模型,其中随机森林的提升度最高。结论:线性回归模型对研究生学习行为的预测效果良好,机器学习在预测准确性上优于线性回归模型,但传统线性回归模型在计算效率和可解读性上具有一定优势。“,”Objective:To compare the prediction efficiency of traditional linear regression model and four machine learning models on the learning behavior of clinical medical postgraduates, and to explore the pros and cons and applicability of different prediction models.Methods:A total of 6,922 clinical medical postgraduates were surveyed, their comprehensive learning behavior scores were obtained through the learning behavior scale. In the training set, Lasso linear regression and artificial neural network, decision tree, Bootstrap random forest, and lifting tree were used to build prediction models respectively. The above models were used to predict the validation set data and compare the prediction efficiency.Results:The comprehensive learning behavior score of clinical medical postgraduates was (3.31±0.54) points, and the overall compliance rate was 74.02%. In the linear regression model, the influence of age, school level, degree type, learning interest, pressure and satisfaction on learning behavior were statistically significant. In the prediction of validation set, the sensitivity, specificity, and accuracy of the linear regression model were 0.484, 0.914, and 0.801, respectively. The indexes of the four machine learning models were higher than those of the traditional linear regression model, and the Bootstrap random forest had the highest elevation.Conclusion:The linear regression model has a good prediction effect on learning behavior, and machine learning is superior to linear regression model in terms of accuracy of prediction. However, traditional linear regression models are superior to machine learning models in computational efficiency and interpretability.