基于不完整数据的IHB-LightGBM心脏病预测模型

提出基于不完整数据的IHB-LightGBM(Improved Hyperband-Light Gradient Boosting Machine)心脏病预测模型。首先,在Hyperband算法超参数采样的基础上引入了权重值,并通过蓄水池法按特征权重对其进行排序,从而筛选出最优参数以提高算法的参数寻优能力;其次,针对心脏病数据样本小且属性缺失的问题,使用K近邻算法对不完整数据进行缺失值插补,再将处理得到的完整数据进行归一化,使数据映射至0～1范围内;最后,对LightGBM采用改进后的IHB优化算法进行全局参数寻优,建立IHB-LightGBM心脏病预测模型。使用UCI心脏病数据集进行实验,结果表明IHB算法的参数寻优效果优于贝叶斯、随机搜索等优化算法,IHB-LightGBM模型在各项评价指标也上明显高于随机森林、极端随机树等算法,可以获得更快的预测速度和更高的预测精度。     

机器学习算法在肝细胞癌微血管侵犯术前预测中的应用价值

目的探讨机器学习算法在肝细胞癌微血管侵犯(MVI)术前预测中的应用价值。方法采用回顾性描述性研究方法。收集2015年5月至2018年12月福建医科大学孟超肝胆医院收治的277例肝细胞癌患者的临床病理资料;男235例,女42例;年龄为(56±10)岁,年龄范围为33~80岁。患者术前均行磁共振成像检查。227例肝细胞癌患者通过计算机产生随机数方法以7∶3比例分为训练集193例和验证集84例。应用逻辑回归列线图,支持向量机(SVM)、随机森林(RF)、人工神经网络(ANN)和轻量级梯度提升机(LightGBM)机器学习算法构建MVI术前预测模型。观察指标:(1)训练集及验证集患者临床病理资料分析。(2)影响训练集患者肿瘤MVI危险因素分析。(3)机器学习算法预测模型构建及其术前预测肿瘤MVI准确性比较。正态分布的计量资料以±s表示,组间比较采用配对t检验。计数资料以绝对数表示,组间比较采用χ2检验。单因素和多因素分析采用Logistic回归模型。结果(1)训练集及验证集患者临床病理资料分析:训练集和验证集患者性别(男,女)分别为157、36例和78、6例,两组比较,差异有统计学意义(χ2=6.028,P<0.05)。(2)影响训练集患者肿瘤MVI危险因素分析:训练集193例患者中,MVI阳性108例,MVI阴性85例。单因素分析结果显示:年龄、肿瘤数目、肿瘤直径、卫星病灶、肿瘤边界、甲胎蛋白(AFP)、碱性磷酸酶(ALP)和纤维蛋白原水平是影响肿瘤MVI的相关因素(比值比=0.971,2.449,1.368,4.050,2.956,4.083,2.532,1.996,95%可信区间为0.943~1.000,1.169~5.130,1.180~1.585,1.316~12.465,1.310~6.670,2.214~7.532,1.016~6.311,1.323~3.012,P<0.05)。多因素分析结果显示:AFP>20μg/L、肿瘤多发、肿瘤直径越大、肿瘤边界不光滑是影响肿瘤MVI的独立危险因素(比值比=3.680,3.100,1.438,3.628,95%可信区间为1.842~7.351,1.334~7.203,1.201~1.721,1.438~9.150,P<0.05),而年龄越大,MVI发生风险越低(比值比=0.958,95%可信区间为0.923~0.994,P<0.05)。(3)机器学习算法预测模型构建及其术前预测肿瘤MVI准确性比较:①应用多因素分析结果筛选指标,包括年龄、AFP、肿瘤数目、肿瘤直径、肿瘤边界,构建逻辑回归列线图,SVM、RF、ANN及LightGBM机器学习算法预测模型,一致性分析结果显示逻辑回归列线图预测模型稳定性较好。逻辑回归列线图、SVM、RF、ANN、LightGBM机器学习算法预测模型训练集和验证集曲线下面积(AUC)分别为0.812、0.794、0.807、0.814、0.810和0.784、0.793、0.783、0.803、0.815,SVM、RF、ANN、LightGBM机器学习算法AUC分别与逻辑回归列线图AUC比较,差异均无统计学意义[(95%可信区间为0.731~0.849,0.744~0.860,0.752~0.867,0.747~0.862,Z=0.995,0.245,0.130,0.102,P>0.05)和(95%可信区间为0.690~0.873,0.679~0.865,0.702~0.882,0.715~0.891,Z=0.325,0.026,0.744,0.803,P>0.05)]。②应用RF、LightGBM机器学习算法自行筛选临床病理因素指标构建预测模型。根据指标对预测模型重要度排序,选择重要度>0.01的指标,包括年龄、肿瘤直径、AFP、白细胞(WBC)、血小板、总胆红素、天冬氨酸氨基转移酶、γ-谷氨酰转移酶、ALP和纤维蛋白原,构建RF机器学习算法预测模型;挑选重要度>5.0的指标,包括年龄、肿瘤直径、AFP、WBC、ALP和纤维蛋白原,构建LightGBM机器学习算法预测模型;由于ANN及SVM机器学习算法不具备筛选指标能力,应用单因素分析结果筛选指标,包括年龄、肿瘤数目、肿瘤直径、卫星病灶、肿瘤边界、AFP、ALP和纤维蛋白原水平,构建SVM、ANN机器学习算法预测模型。SVM、RF、ANN、LightGBM机器学习算法预测模型训练集和验证集AUC分别为0.803、0.838、0.793、0.847和0.810、0.802、0.802、0.836,分别与逻辑回归列线图AUC比较,差异均无统计学意义[(95%可信区间为0.740~0.857,0.779~0.887,0.729~0.848,0.789~0.895,Z=0.421,0.119,0.689,1.517,P>0.05)和(95%可信区间为0.710~0.888,0.700~0.881,0.701~0.881,0.740~0.908,Z=0.856,0.458,0.532,1.306,P>0.05)]。结论机器学习算法可用于术前预测肝细胞癌MVI,但其应用价值尚需多中心大样本数据进一步验证。     

基于机器学习方法的非编码RNA-蛋白质相互作用的预测

目的非编码RNA-蛋白质的相互作用(noncoding RNA-protein interactions,ncRPI)具有重要的生物学意义,目前预测其相互作用已成为当下研究非编码RNA (noncoding RNA,ncRNA)和蛋白质功能的重要途径之一。方法本研究基于ncRNA和蛋白质的序列信息提取特征,运用卷积自编码器预处理原始数据,训练三个机器学习模型:LightGBM(LBM)、随机森林(random forest,RF)和极端梯度增强算法(extreme gradient boosting,XGB),预测ncRNA与蛋白质的相互作用。结果在RPI369和RPI488两个数据集做5倍交叉验证,LBM、RF与XGB三个模型在两个数据集均达到较高的预测准确率,在RPI369数据集三个模型的预测准确率分别为0.757(LBM)、0.791(RF)、0.791(XGB),在RPI488数据集三个模型的预测准确率分别为0.918(LBM)、0.908(RF)、0.918(XGB);三个模型在RPI1807、RPI2241、RPI13254大数据集也取得较高的AUC(area under curve)值,在RPI1807三个模型的AUC值均为0.99,在RPI2241三个模型最低AUC值为0.87,在RPI13254三个模型最低AUC值为0.81,都表现出较好的预测准确性。结论机器学习方法能够预测ncRNA与蛋白质是否存在相互作用。     

Hardness Prediction of Laser Powder Bed Fusion Product Based on Melt Pool Radiation Intensity

The quality stability and batch consistency of laser powder bed fusion products are key issues that must be solved in additive manufacturing. The melt pool radiation intensity data of laser powder bed fusion contain a significant amount of forming process information, and studies have shown that the analysis of melt pool radiation intensity using data-driven methods can achieve online quality judgment; however, there are still speed and accuracy problems. In this study, we propose a data-driven model for hardness predictions of laser powder bed fusion products based on process parameters fused with power spectrum features of melt pool intensity data, which quickly and accurately predicts the microhardness of laser powder bed fusion specimens and can make constructive guidance for closed-loop feedback quality regulation in practical production. The effects of three integrated learning models, Random Forest, XGBoost and LightGBM, are also compared. The results indicate that random forest has the highest prediction accuracy in this dataset; however, it has the limitation of slow training and prediction speeds. The LightGBM algorithm has the fastest training and prediction speeds, about 1.4% and 4.4% of the random forest, respectively; however, the prediction accuracy is lower than that of random forest and XGBoost. XGBoost has the best overall comparative performance with adequate training and prediction speeds, about 23.7% and 37.9% of the random forest, respectively, while ensuring a specified prediction accuracy, which is suitable for application in engineering practices.     

Prediction of lipid nanoparticles for mRNA vaccines by the machine learning algorithm

Lipid nanoparticle (LNP) is commonly used to deliver mRNA vaccines. Currently, LNP optimization primarily relies on screening ionizable lipids by traditional experiments which consumes intensive cost and time. Current study attempts to apply computational methods to accelerate the LNP development for mRNA vaccines. Firstly, 325 data samples of mRNA vaccine LNP formulations with IgG titer were collected. The machine learning algorithm, lightGBM, was used to build a prediction model with good performance (R² > 0.87). More importantly, the critical substructures of ionizable lipids in LNPs were identified by the algorithm, which well agreed with published results. The animal experimental results showed that LNP using DLin-MC3-DMA (MC3) as ionizable lipid with an N/P ratio at 6:1 induced higher efficiency in mice than LNP with SM-102, which was consistent with the model prediction. Molecular dynamic modeling further investigated the molecular mechanism of LNPs used in the experiment. The result showed that the lipid molecules aggregated to form LNPs, and mRNA molecules twined around the LNPs. In summary, the machine learning predictive model for LNP-based mRNA vaccines was first developed, validated by experiments, and further integrated with molecular modeling. The prediction model can be used for virtual screening of LNP formulations in the future.     

Prediction of Response of Hepatocellular Carcinoma to Radioembolization: Machine Learning Using Preprocedural Clinical Factors and MR Imaging Radiomics

PurposeTo create and evaluate the ability of machine learning–based models with clinicoradiomic features to predict radiologic response after transarterial radioembolization (TARE).Materials and Methods82 treatment-naïve patients (65 responders and 17 nonresponders; median age: 65 years; interquartile range: 11) who underwent selective TARE were included. Treatment responses were evaluated using the European Association for the Study of the Liver criteria at 3-month follow-up. Laboratory, clinical, and procedural information were collected. Radiomic features were extracted from pretreatment contrast-enhanced T1-weighted magnetic resonance images obtained within 3 months before TARE. Feature selection consisted of intraclass correlation, followed by Pearson correlation analysis and finally, sequential feature selection algorithm. Support vector machine, logistic regression, random forest, and LightGBM models were created with both clinicoradiomic features and clinical features alone. Performance metrics were calculated with a nested 5-fold cross-validation technique. The performances of the models were compared by Wilcoxon signed-rank and Friedman tests.ResultsIn total, 1,128 features were extracted. The feature selection process resulted in 12 features (8 radiomic and 4 clinical features) being included in the final analysis. The area under the receiver operating characteristic curve values from the support vector machine, logistic regression, random forest, and LightGBM models were 0.94, 0.94, 0.88, and 0.92 with clinicoradiomic features and 0.82, 0.83, 0.82, and 0.83 with clinical features alone, respectively. All models exhibited significantly higher performances when radiomic features were included (P = .028, .028, .043, and .028, respectively).ConclusionsBased on clinical and imaging-based information before treatment, machine learning–based clinicoradiomic models demonstrated potential to predict response to TARE.