Patient Characteristics
A total of 17031 patients with NSCLC diagnosed from 2010 to 2015 were included in this study (Tables 1). Among these patients, 8661 (50.85%) were treated with radiotherapy while 8659 (50.84%) were not treated with radiotherapy. And 9369 (55.01%) were treated with chemotherapy while 7644 (44.88%) were not treated with chemotherapy. A median CSS of 14.0 months (range 1.0–119.0) was observed for the NSCLC patients.
Table 1. Descriptive analysis table of categorical variables
Variables
|
Categorical
|
Recode
|
Frequency
|
Percentage
|
Variables
|
Categorical
|
Recode
|
Frequency
|
Percentage
|
Marital Status
|
Unmarried
|
1
|
7751
|
45.51%
|
N stage
|
N0
|
1
|
6585
|
38.66%
|
|
Married
|
2
|
9262
|
54.38%
|
|
N1
|
2
|
1673
|
9.82%
|
Age
|
<60 years old
|
0
|
3686
|
21.64%
|
|
N2
|
3
|
6084
|
35.72%
|
|
≥60 years old
|
1
|
13327
|
78.25%
|
|
N3
|
4
|
2671
|
15.68%
|
Race
|
White
|
1
|
13160
|
77.27%
|
M stage
|
M0
|
1
|
9030
|
53.02%
|
|
Black
|
2
|
1512
|
8.88%
|
|
M1
|
2
|
7983
|
46.87%
|
|
Others
|
3
|
2341
|
13.75%
|
Surgery
|
Yes
|
1
|
4414
|
25.92%
|
Sex
|
Male
|
1
|
8768
|
51.48%
|
|
No
|
0
|
12599
|
73.98%
|
|
Female
|
2
|
8245
|
48.41%
|
lymph node examination
|
Yes
|
1
|
6135
|
36.02%
|
Primary site
|
Upper lobe
|
1
|
10205
|
59.92%
|
|
No
|
0
|
10878
|
63.87%
|
|
Middle lobe
|
2
|
776
|
4.56%
|
Radiotherapy
|
Yes
|
1
|
8661
|
50.85%
|
|
Lower lobe
|
3
|
5088
|
29.87%
|
|
No
|
0
|
8659
|
50.84%
|
|
Overlapping lesion
|
4
|
177
|
1.04%
|
Chemotherapy
|
Yes
|
1
|
9369
|
55.01%
|
|
Hard to distinguish
|
5
|
767
|
4.50%
|
|
No
|
0
|
7644
|
44.88%
|
Summary stage
|
localized
|
1
|
2973
|
17.46%
|
Tumor size
|
<70mm
|
1
|
14550
|
85.43%
|
|
Regional
|
2
|
4412
|
25.91%
|
|
≥70mm
|
2
|
2463
|
14.46%
|
|
Distant
|
3
|
9628
|
56.53%
|
bone metastasis
|
Yes
|
1
|
3540
|
20.79%
|
Histological type
|
Adenocarcinoma
|
1
|
11893
|
69.83%
|
|
No
|
0
|
13473
|
79.11%
|
|
Squamous lung cancer
|
2
|
4947
|
29.05%
|
brain metastasis
|
Yes
|
1
|
2588
|
15.20%
|
|
Large cell lung
|
3
|
173
|
1.02%
|
|
No
|
0
|
14425
|
84.70%
|
T stage
|
T1
|
1
|
3779
|
22.19%
|
liver metastasis
|
Yes
|
1
|
1360
|
7.99%
|
|
T2
|
2
|
6171
|
36.23%
|
|
No
|
0
|
15653
|
91.91%
|
|
T3
|
3
|
1047
|
6.15%
|
Survival status
|
Failure
|
1
|
13839
|
81.26%
|
|
T4
|
4
|
6016
|
35.32%
|
|
Censored
|
0
|
3174
|
18.64%
|
Survival curve crossing
Survival curves for all variables were plotted using the Kaplan-Meier method, and differences were found between groups by log-rank test (P<0.05) (Figure A.1). There was a SCC between the chemotherapy and non-chemotherapy patients (Figure 2a), and the survival time of the crossing point could be estimated to be 14 months. That is, the survival rate was higher in the chemotherapy-practicing patients before the crossing point, while the survival rate was higher in the non-chemotherapy patients after the crossing point. Although there was not a SCC of radiotherapy, the survival rate was lower with radiotherapy than without radiotherapy almost all the time (Figure 2b). Segmentation analysis based on the SCC of chemotherapy was performed in the following sections, the results after the crossing point were of more concern, because we wanted to figure out probable reasons for poor prognosis of chemotherapy and specific patients with survival time >14 months who would benefit from chemotherapy or radiotherapy. So only the results after the crossing point were shown in the following sections, the results before the crossing point were shown in the supplementary file (Appendix B).
Test the fitting requirement
The Cox model, the AFT model under LLD, and the AFT model under Weibull distribution were constructed both in the segmented and non-segmented analysis. Firstly, the time-dependent coefficient method was used to test the PH assumption for the Cox model. Therefore, three variables did not satisfy the PH assumption after the crossing point, including summary stage, T-stage, and surgery, but most of the variables did not satisfy the PH assumption in the non-segmented analysis (Table 2). Secondly, testing the fitting requirement for the AFT model under LLD by plotting curve, the segmented analysis and the non-segmented analysis both can be approximated as a straight line (Figure 3). Thirdly, testing the fitting requirement for the AFT model under LLD by plotting curve, the segmented analysis and the non-segmented analysis both can be approximated as a straight line (Figure 4). Therefore, the goodness-of-fit need to be compare between these survival model that both satisfied the fitting requirement.
Table 2. Results of the time-dependent coefficient test
With segmentation
|
Without segmentation
|
Time dependent variable
|
P
|
Time dependent variable
|
P
|
Primary site*T
|
0.001
|
Primary site*T
|
0.007
|
Summary stage* T
|
<0.001
|
Summary stage* T
|
<0.001
|
Histological type * T
|
0.182
|
Histological type * T
|
<0.001
|
T stage* T
|
<0.001
|
T stage* T
|
<0.001
|
N stage * T
|
0.016
|
N stage * T
|
<0.001
|
M stage * T
|
0.101
|
M stage * T
|
<0.001
|
Surgery * T
|
<0.001
|
Surgery * T
|
<0.001
|
lymph node examination * T
|
0.002
|
lymph node examination * T
|
0.189
|
Radiotherapy * T
|
0.363
|
Radiotherapy * T
|
0.157
|
Chemotherapy * T
|
0.075
|
Chemotherapy * T
|
<0.001
|
Tumor size* T
|
0.949
|
Tumor size* T
|
<0.001
|
Bone metastasis * T
|
0.123
|
Bone metastasis * T
|
0.001
|
Brain metastasis * T
|
0.006
|
Brain metastasis * T
|
0.002
|
Liver metastasis * T
|
0.017
|
Liver metastasis * T
|
0.011
|
Model comparison
It can be seen from the baseline table that the problem of imbalance between covariates was obvious. The figures of inter-group SMD before and after IPW were drawn to test the balance (Appendix C). In segmented analysis, there were still 8 covariates with inter-group SMD greater than 0.1 after weighting for chemotherapy, including race, primary site of tumor, summary stage, T stage, N stage, M stage, bone metastasis and tumor size. And there were still 7 covariates with inter-group SMD greater than 0.1 after weighting for radiotherapy, including primary site of tumor, summary stage, T stage, N stage, M stage, brain metastasis and brain metastasis. In the non-segmented analysis, the inter-group SMD of T stage remained greater than 0.1 after weighting for chemotherapy, and the inter-group SMD of summary stage, T stage and M stage remained greater than 0.1 after weighting for radiotherapy. Therefore, these covariates were included in the following prognostic analysis model.
AIC and Cox-Snell Residuals were applied to compare different survival models. While residual plots of the AFT model under LLD and the Cox model both showed a goodness-of-fit, the AFT model under LLD showed the smallest AIC values in both the chemotherapy-specific and radiotherapy-specific model. However, the AFT model under LLD was optimal for analyzing both chemotherapy and radiotherapy based on the combination of the two metrics (Table 3 and Figure 5).
Table 3. The AIC of prognostic analysis models
AIC
|
with segmentation
|
without segmentation
|
Cox model
|
AFT model under LLD
|
AFT model under Weibull distribution
|
Cox model
|
AFT model under LLD
|
AFT model under Weibull distribution
|
model for chemotherapy
|
213266.26
|
108539.95
|
109576.87
|
552809.31
|
231422.00
|
237030.08
|
model for radiotherapy
|
218491.58
|
111511.49
|
112616.18
|
567125.33
|
237586.66
|
241660.56
|
Case study
Since the results of goodness-of-fit showed the best performance of the AFT model under LLD, it was used in the segmented analysis. And the Cox model was constructed in the non-segmented analysis for comparison.
In the segmented analysis, we found out that the survival time for radiotherapy was shorter than non-radiotherapy (=0.930, 95%CI: 0.881-0.980), and the survival time for chemotherapy was shorter than non-chemotherapy (=0.925, 95%CI: 0.877-0.973). In the non-segmented analysis, we found out the for radiotherapy was not statistically significant (Table 4). Even so, the of radiotherapy (=0.922) and chemotherapy (=0.628) both showed them to be a protective factor for prognosis. The comparison revealed that the non-segmented Cox model was unable to find out that radiotherapy and chemotherapy had the effect of shortening survival time compared with non-radiotherapy and non-chemotherapy in NSCLC patients with survival time >14 months.
Table 4. Comparing the results of the segmented analysis and the non-segmented analysis
|
segmented analysis (AFT model under LLD)
|
non-segmented analysis (Cox model)
|
Variable
|
|
95%CI lower bound
|
95%CI upper bound
|
P
|
|
95%CI lower bound
|
95%CI upper bound
|
P
|
Radiotherapy (No for reference)
|
|
|
|
|
|
|
|
|
Yes
|
0.930
|
0.881
|
0.980
|
<0.05
|
0.974
|
0.922
|
1.030
|
0.354
|
Chemotherapy (No for reference)
|
|
|
|
|
|
|
|
|
Yes
|
0.925
|
0.877
|
0.973
|
<0.05
|
0.657
|
0.628
|
0.688
|
<0.05
|
Subgroup analyses
In the KM curves, we found a lower survival rate for chemotherapy after the crossing point, and a lower survival rate for radiotherapy all the time. In the segmented analysis, we also found out that radiotherapy and chemotherapy shortened the survival time compared to non-radiotherapy and non-chemotherapy. Therefore, we performed subgroup analyses to find specific patients with survival time >14 months who would benefit from radiotherapy or chemotherapy. Since the PH assumption was not satisfied by multiple variables in the subgroup analyses, the PO assumption showed a linear relationship (Figure 6), all indicating a good fit, then the AFT model under LLD was constructed.
The study found longer survival time with chemotherapy compared to no chemotherapy in patients with regional stage (=1.305), with distant stage (=1.149), with squamous lung cancer (=1.246), with large cell lung cancer (=7.275), with T3 stage (=1.721), with N1 stage (=1.296), with N2 stage (=1.096), with N3 stage (=1.880), with M1 stage (=1.090), with non-surgery (=1.116), with non-lymph node examination (=1.090), with tumor size ≥ 70mm (=1.463), with liver metastasis (=1.476) (Table 5). Among them, chemotherapy had the best effect on prolonging the survival time of NSCLC patients with large cell lung cancer.
The study also found longer survival time with radiotherapy compared to no radiotherapy in patients with primary site hard to distinguish (=1.534), with N2 stage (=1.113), with N3 stage (=1.212), with non-surgery (=1.195) (Table 5). Among them, radiotherapy had the best effect on prolonging the survival time of NSCLC patients with large cell lung cancer.
Table 5. Results of subgroup analyses
|
Specific patient
|
|
95%CI lower bound
|
95%CI upper bound
|
P
|
Benefiting from chemotherapy (non-chemotherapy for reference)
|
Regional stage
|
1.305
|
1.206
|
1.403
|
<0.05
|
Distant stage
|
1.149
|
1.079
|
1.220
|
<0.05
|
Squamous lung cancer
|
1.246
|
1.151
|
1.342
|
<0.05
|
Large cell lung
|
7.275
|
5.723
|
8.827
|
<0.05
|
T3
|
1.721
|
1.542
|
1.899
|
<0.05
|
N1
|
1.296
|
1.170
|
1.424
|
<0.05
|
N2
|
1.096
|
1.013
|
1.179
|
<0.05
|
N3
|
1.880
|
1.756
|
2.003
|
<0.05
|
M1
|
1.090
|
1.013
|
1.167
|
<0.05
|
Non-surgery
|
1.116
|
1.061
|
1.172
|
<0.05
|
Non-lymph node examination
|
1.090
|
1.027
|
1.152
|
<0.05
|
Tumor size ≥70mm
|
1.463
|
1.293
|
1.633
|
<0.05
|
Liver metastasis
|
1.476
|
1.289
|
1.662
|
<0.05
|
Benefiting from radiotherapy (non-radiotherapy for reference)
|
primary site hard to distinguish
|
1.534
|
1.294
|
1.775
|
<0.05
|
N2
|
1.113
|
1.026
|
1.199
|
<0.05
|
N3
|
1.212
|
1.082
|
1.341
|
<0.05
|
Non-surgery
|
1.195
|
1.138
|
1.252
|
<0.05
|
Note: did not receive radiotherapy or chemotherapy was considered as a reference