Page 21 - 《南京医科大学学报》自然科学版2026年第2期

第46卷第2期 杨 玥，葛 愿，李明辉，等. 基于机器学习的心脏术后衰弱预测模型的构建与验证［J］.
2026年2月 南京医科大学学报（自然科学版），2026，46（2）：173-180，187 ·177 ·

A Important variables in LASSO E Important variables in XGBoost H Variable importance plot of RF
Valves Age Age
treatment
Albumin LVDd
CABG
LVDd LVEF
Age LVEF Hypertension
Hypertension Albumin
LVDd
Surgical duration Sex
LVEF Sex CABG
BMI Smoking
Albumin
Smoking
Surgical duration
BMI
Drinking Valves
treatment
0 0.25 0.50 0.75 0 0.1 0.2 0.3 0.4 0 5 10 15 20
Coefficient Importance score Importance score
B ROC curve of LASSO F ROC curve of XGBoost I ROC curve of RF
1.00 1.00 Cut⁃off=0.309
Cut⁃off=0.243 1.00 Cut⁃off=0.238
0.75 0.75 0.75
Sensitivity 0.50 Sensitivity 0.50 Sensitivity 0.50

0.25 0.25 0.25
AUC=0.795 AUC=0.839 AUC=0.838
95%CI：0.698-0.892 95%CI：0.757-0.921 95%CI：0.756-0.920
0 0 0
0 0.25 0.50 0.75 1.00 0 0.25 0.50 0.75 1.00 0 0.25 0.50 0.75 1.00
1-Specificity 1-Specificity 1-Specificity
C G J The relationship between
LASSO cross⁃validation results Beeswarm plot
tree number and OOB
Age 0.407
0.28 Albumin 0.225 0.6
Classification error 0.24 Surgical duration 0.083 Feature value Error rate 0.4 Error rate
LVDd 0.191
0.26
LVEF 0.175
High
Hypertension 0.141
Alive
Low
Sex 0.076
Death
0.22
BMI 0.044
0.20 Smoking 0.034 0.2 OOB
Drinking 0.033
-6 -4 -2 -1.0 -0.5 0 0.5 1.0
Min.：0.049 6｜1se.：0.059 8｜ SHAP value
Variables at min.：7 （Impact on model output） 0 100 200 300 400 500
Number of trees
D LASSO regression path Variable K LASSO XGBoost
Age Hypertension
2
Albumin LAD
Sex LVDd
1 CABG LVPW 0 1 1
Coefficient 0 Cerebral infarction PSQI 2 4 4 Common variables：
COPD
BMI
-1
Valves treatment
Surgical duration
Drinking Smoking 0 LVEF VIF：1.221
Age VIF：1.024
-2 LVEF Diabetes
1×10 -3 1×10 -2 1×10 -1 Albumin VIF：1.010
RF LVDd VIF：1.233
logλ
A-D：The optimal parameter λ selection in the LASSO model employed 10⁃fold cross⁃validation using the one⁃standard⁃error rule（λ.1se=0.059 8），
which resulted in the selection of seven variables. E-G：The XGBoost model selected the top 10 most important variables. Panel G presented the SHAP
beeswarm plot，demonstrating feature importance and the directional impact of features on the model’s predictions. H-J：The RF model selected the top
10 most important variables，trained with 500 decision trees to ensure stability，and 10⁃fold cross⁃validation was applied to optimize hyperparameters
and prevent overfitting. Panel J displayed the out⁃of⁃bag（OOB）error curve of the RF model，illustrating the relationship between the model’s error rate
and the number of decision trees. K：A Venn diagram compared the overlap of variables selected by the three different methods.
图2 机器学习分析结果
Figure 2 Machine learning analysis results