3.1 Patient Characteristics
A total of 106 lung adenocarcinoma patients were included in this study. The TNM stage was IIIb or IV, and EGFR gene mutation was positive by gene test. There were 43 males and 63 females. There was no significant difference between sex and age groups, but there was significant difference between the length before treatment and the length after 2 months of treatment. Table 1 shows the clinical features of the patients.
Table1. Summary of clinical characteristics
characteristics
|
response group
(n=53)
|
nonresponse
group(n=53)
|
Statistic
|
P value
|
gender
|
|
|
0.978
|
0.323
|
male
|
29(54.7%)
|
34(64.2%)
|
-1.968
|
0.049
|
female
long diameter before treatment
|
24(45.3%)
36(29,45)
|
19(35.8%)
32(24,40)
|
long diameter after 2 months
|
20(15,27)
|
28(19,35)
|
2.989
|
0.003
|
age
|
65.6±8.07
|
69.2±10.01
|
1.820
|
0.072
|
3.2 Analysis of radiomics features
396 initial radiomics features were extracted from CT plain scan images, and the delta radiomics were calculated. Subsequently, LASSO regression and 10 cross-validation dimensionality reduction were performed on the training set data, and three radiomics features with non-zero coefficients were extracted as sumEntropy
LongRunLowGreyLevelEmphasis_angle0_offset1, LongRunLowGreyLevelEmphasis_angle0_offset4. The screening process and coefficients are shown in Figure1 and Table 2. The comparison of delta radiomics between the response group and nonresponse group is shown in Figure 2 and Table 3.
Table 2 Description of Radiomics Feature Coefficients
Delta radiomic features
|
coefficient
|
constant
|
-0.1134597
|
sumEntropy
|
0.1871893
|
LongRunLowGreyLevelEmphasis_angle0_offset1
|
0.1314608
|
LongRunLowGreyLevelEmphasis_angle0_offset4
|
0.1393257
|
Table3.The comparison of delta radiomics between the response group and nonresponse group
delta radiomics
|
response group
|
nonresponse group
|
statistic
|
p
|
sumEntropy
|
0.760±0.895
|
0.155±0.729
|
-3.751
|
<0.001
|
LongRunLowGreyLevelEmphasis_angle0_offset1
|
0.535[0.208,1.019]
|
0.121[0.020,0.359]
|
-3.949
|
<0.001
|
LongRunLowGreyLevelEmphasis_angle0_offset4
|
0.588[0.128,1.229]
|
0.302[0.068,0.724]
|
-3.753
|
<0.001
|
Index1: sumEntropy
Index2: LongRunLowGreyLevelEmphasis_angle0_offset1
Index3: LongRunLowGreyLevelEmphasis_angle0_offset4
3.3 Analysis of diagnostic efficacy of radiomics models
3.3.1 Analysis of Diagnostic Efficacy of Logistic Regression Model
The diagnostic performance of the logistic regression model in evaluating the efficacy of EGFR-TKI targeted drugs in the treatment of lung adenocarcinoma is shown in the table below. As shown below. The AUC of the training group and the testing group were 0.778 and 0.773, respectively, and the 95% CIs were 0.670-0.886, 0.593-0.952, respectively. The sensitivity was 0.83, 0.67, the specificity was 0.86, 0.80, the positive predictive value was 0.85, 0.77, the negative predictive value was 0.83, 0,71, the accuracy was 0.84, 0.73, and the F1 was 0.84, 0.71, respectively, as shown in Figure 3 and Table 4.
Table 4 Diagnostic performance of logistic regression models
|
|
|
training
|
testing
|
AUC
|
0.778
|
0.773
|
95 CI%
|
0.670-0.886
|
0.593-0.952
|
sensitivity
|
0.667
|
0.765
|
specificity
|
0.743
|
0.6
|
PPV
|
0.743
|
0.684
|
NPV
|
0.667
|
0.692
|
accuracy
|
0.703
|
0.688
|
F1
|
0.703
|
0.722
|
3.3.2 Analysis of Diagnostic Efficiency of Decision Tree Model
The diagnostic performance of the decision tree model in evaluating the efficacy of EGFR-TKI targeted drugs in the treatment of lung adenocarcinoma is shown in the table below. As shown below. The AUC of the training group and the testing group were 0.834 and 0.653, respectively, and the 95% CIs were 0.751-0.917, 0.465-0.841, respectively. The sensitivity was 0.743, 0.706, the specificity was 0.857, 0.400, the positive predictive value was 0.852, 0.571, the negative predictive value was 0.750, 0.545, the accuracy was 0.797, 0.562, and the F1 was 0.84, 0.71, respectively, as shown in Figure 4 and shown in Table 5.
Table 5 Diagnostic performance of decision tree models
|
training
|
testing
|
AUC
|
0.834
|
0.653
|
95 CI%
|
0.751-0.917
|
0.465-0.841
|
sensitivity
|
0.743
|
0.706
|
specificity
|
0.857
|
0.400
|
PPV
|
0.852
|
0.571
|
NPV
|
0.750
|
0.545
|
accuracy
|
0.797
|
0.562
|
F1
|
0.794
|
0.632
|
3.3.3 Analysis of Diagnostic Efficiency of Support Vector Machine Model
The diagnostic performance of the support vector machine model in evaluating the efficacy of EGFR-TKI targeted drugs in the treatment of lung adenocarcinoma is shown in the table below. The AUC of the training group and the validation group were 0.817 and 0.745, respectively, and the 95% CIs were 0.719-0.915 and 0.565-0.926, respectively. As shown below. The sensitivity was 0.564, 0.588, the specificity was 0.942, 0.800, the positive predictive value was 0.917, 0.769, the negative predictive value was 0.660, 0.631, the accuracy was 0.743, 0.687, and the F1 was 0.698, 0.667, respectively, as shown in Figure 5 and shown in Table 6.
Table 6 Diagnostic performance of SVM models
|
training
|
testing
|
AUC
|
0.817
|
0.745
|
95 CI%
|
0.719-0.915
|
0.565-0.926
|
sensitivity
|
0.564
|
0.588
|
specificity
|
0.942
|
0.800
|
PPV
|
0.917
|
0.769
|
NPV
|
0.660
|
0.631
|
accuracy
|
0.743
|
0.687
|
F1
|
0.698
|
0.667
|
3.3.4 Analysis of Diagnostic Efficiency of Adaptive Enhancement Model
The diagnostic performance of the adaptive enhancement model in evaluating the efficacy of EGFR-TKI targeted drugs in the treatment of lung adenocarcinoma is shown in the table below. The AUC of the training group and the testing group were 0.730 and 0.414, respectively, and the 95% CIs were 0.628-0.833, 0.242-0.586, respectively. The sensitivity was 0.718, 0.706, the specificity was 0.743, 0.467, the positive predictive value was 0.757, 0.600, the negative predictive value was 0.703, 0.583, the accuracy was 0.730, 0.594, and the F1 was 0.737, 0.649, respectively, as shown in Figure 6 and shown in Table 7.
Table.7 Diagnostic performance of adaptive boosting models
|
training
|
testing
|
AUC
|
0.730
|
0.414
|
95 CI%
|
0.628-0.833
|
0.242-0.586
|
sensitivity
|
0.718
|
0.706
|
specificity
|
0.743
|
0.467
|
PPV
|
0.757
|
0.600
|
NPV
|
0.703
|
0.583
|
accuracy
|
0.730
|
0.594
|
F1
|
0.737
|
0.649
|
3.4 Nomogram analysis
A logistic regression-based learning algorithm builds a Nomogram model (as shown in Figure 7). The three radiomics parameters index1, index2, and index3 represent sumEntropy respectively
LongRunLowGreyLevelEmphasis_angle0_offset1
The points corresponding to LongRunLowGreyLevelEmphasis_angle0_offset4) are added together to obtain the total score. The higher the total score, the higher the probability that the corresponding targeted therapy is effective. After the Nomogram prediction model was verified by Boostrap method for 1000 times of internal sampling, it was concluded that the predicted value of the model had a good coincidence with the actual clinical observation value, and was close to a 45° oblique line (as shown in Figure 8), indicating that the The calibration of the model is good; and the decision curve shows that the nomogram model has a greater clinical benefit in evaluating the efficacy of EGFR-TKI targeted drugs in the treatment of lung adenocarcinoma, as shown in Figure 9.