A practical recurrence risk model based on Lasso-Cox regression for gastric cancer

Abstract

Introduction

Gastric cancer remains huge cancer threat worldwide. Detecting the recurrence of gastric cancer after treatment is especially important in improving the prognosis of patients. We aim to fit different risk models with different clinical variables for patients with gastric cancer, which further provides applicable guidance to clinical doctors for their patients.

Methods

We collected the primary data from the medical record system in Lanzhou University Second Hospital and further cleaned the primary data via assessing data integrity artificially; meanwhile, detailed conclusion criteria and exclusion criteria were made. We used R software (version 4.1.3) and SPSS 25.0 to analyze data and build models, in which SPSS was used to analyze the correlation and difference of different items in the training set and testing set, and different R packages were used to run LASSO regression, Cox regression and nomogram for variable selection, model construction and model validation.

Result

A total of 649 patients were included in our data analysis and model building. In LASSO regression selection, seven variables, pathological stage, tumor size, the number of total lymph nodes, the number of metastatic lymph nodes, intraoperative blood loss (IBL), the level of AFP and CA199, showed their correlation to the dependent variable. The multivariable Cox regression model fitted using these seven variables showed medium prediction ability, with an AUC of 0.840 in the training set and 0.756 in the testing set.

Conclusions

Pathological stage, tumor size, the number of total lymph nodes, the number of metastatic lymph nodes, IBL, the level of AFP and CA199 are significant in identifying recurrence risk for gastric cancer patients after radical gastrectomy.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00432-023-05346-1.

Introduction

According to global cancer statistics, there were approximately 19.3 million new cancer cases and 10.0 million cancer deaths in 2020. The global cancer burden continues to increase and is expected to be 28.4 million (a 47% rise from 2020) in 2040; in all of the cancer types, gastric cancer remains huge cancer threat worldwide; it ranks fifth for incidence and fourth for mortality globally (Sung et al. 2020). In addition to its high incidence, gastric cancer also shows a high recurrence rate, which is leading cause of gastric cancer death. Complete surgical resection is still the most important therapy for gastric cancer. Although surgical techniques have developed rapidly in recent years, long-term survival is still poor. According to a previous study, there are still about 80% of patients with gastric cancer die within a short period time from locoregional recurrence (87%) or distant metastasis (30%) (Kim et al. 2011). Thus, detecting the recurrence of gastric cancer after treatment is especially important in improving patients’ prognosis.

In recent years, people have realized that many important information exists in medical big data; meanwhile, machine learning shows its promising prospect in the data processing. Thus, previous researchers have designed many different clinical models based on machine learning and big medical data for clinical applications, of which, risk model has been found as a practical tool in clinical application, with which doctor can distinguish the different risk groups and give them different individual treatment plans. Various methods has been well used in cancer recurrence monitoring, including genomics, radiomics, clinical data and dietary habit. Jing Zhong et al. constructed a predictive model for patients with lung cancer with gene function clustering and machine learning. Ranjana Mitra et al. fitted a prognostic model with four significant risk genes derived from 40 differential genes, and the model exhibited excellent predictive capacity in clinical utility (Mitra et al. 2011). Ilda Patrícia Ribeiro et al. built a robust model with array comparative genomic hybridization data to predict HNSCC recurrence/metastasis, and this model showed excellent accuracy and feasibility (Ribeiro et al. 2017). Li Cheng et al. searched The Cancer Genome Atlas (TCGA) database and identified 44 upregulated and 117 downregulated genes in the recurrent samples, they constructed a random forest classifier with the subnetwork nodes as feature genes, which exhibited a 92% true positive rate when classifying recurrent and non-recurrent OC samples (Cheng et al. 2018). In some other studies, cohort study was conducted to identify the lifestyle risk factors related to cancer recurrence. According to a complete population-based cohort, Åsa Åkesson et al. found that age, FIGO stage III and adjuvant treatment were independent prognostic factors for recurrence of endometrioid endometrial cancer (Åkesson et al. 2023). In another multicenter cohort study, researcher fitted a prediction model for early systemic recurrence in breast cancer with some variables, which were significantly associated with distant recurrence, including sentinel lymph node tumor burden, age, pT classification, grade, progesterone receptor, adjuvant cytotoxic chemotherapy and adjuvant anti-human epidermal growth factor receptor 2 therapy, they indicated that the model could accurately predict early systemic recurrence and may facilitate decision-making related to treatment (Osako et al. 2022).

Although the predictive models based on genomics, radiomics and cohort study show excellent predictive capacity, a core problem is that the identification of significant variables needs to take a long time to complete or bring heavy economical burden to patients, thus more affordable and convenient models are urgently needed. In this study, we aim to fit different risk models with common clinical variables, which are all routinely tested in clinical business, and further provides convenient and applicable guidance to clinical doctors for their patients.

Data and method

Data preparation

We collected the primary data from the medical record system in Lanzhou University Second Hospital and further cleaned the primary data by assessing data integrity artificially. The inclusion criteria and exclusion criteria are as follows.

Inclusion criteria

1. Confirmed diagnosis of gastric cancer based on pathological results; the pathological result is limited to adenocarcinoma and signet-ring cell carcinoma.

2. No distal metastasis before operation.

3. Underwent radical gastrectomy operation and R0 resection.

4. Confirmed diagnosis of tumor recurrence after operation based on imaging results, including CT, ultrasound, MR or PET.

5. Complete clinical data, including gender, age, BMI, operation style, operation time, intraoperative blood loss (IBL), pathological stage, differentiation degree, tumor size, American Joint Committee of Cancer (AJCC) tumor-node-metastasis (TNM) classification (T stage, N stage), tumor location, total lymph nodes, metastatic lymph nodes, the level of AFP, CEA, CA125 and CA199 before operation, operation complication and chemotherapy status.

6. Complete survival data, including survival status (recurrence or not) and follow-up time.

Exclusion criteria

1. Patients diagnosed with a gastric benign lesion, sarcoma, neuroendocrine neoplasm or stromal tumor.

2. Underwent palliative surgery.

3. Incomplete clinical data or survival data.

Data analysis and model building

We used R software (version 4.1.3) and SPSS 25.0 in data analysis and model building. First, the caret package was used to randomly divide the cleaned data into the training set and testing set, conforming to the theoretical ratio of 7:3. Second, SPSS was used to analyze the correlation and difference of different items in the training set and testing set, in which continuous numerical variables were shown as a structure of ($\overline{x}\pm s$) and tested with t-test; the categorical variables were shown as n (%) and compared with Chi-square analysis. Third, the data frames were converted into different data matrixes, and the dummy variables were set in categorical variables, which were more than two categories. Subsequently, we used the Glmnet package to run the LASSO regression function, which was found to be excellent in variables selection. Fourth, the survival package was used to run multiple Cox regression and build a Cox model based on the variables selected in LASSO regression. Fifth, we built a nomogram model by using the nomogram package.

Model validation

Model validation was performed in the training set and testing set, respectively. First, we used the pROC package to fit different receiver characteristic curves (ROCs) in the training set and testing set, the area under the curve (AUC) was used to measure the predicting ability of the model built before. Second, the KM curve was used to test the cutoff value in the Cox regression model. Third, the lrm package was used to draw calibration curve and evaluate the calibration of the nomogram model. Fourth, we used rmda package to finish Decision Curve Analysis (DCA) to further evaluate the model in the training set and testing set.

Visualization

The print, plot and ggplot functions were used in model visualization.

Result

Characteristics of the study cohort

A total of 1851 patients with gastric cancer who received radical resection at Lanzhou University Second Hospital from January 1, 2014, to January 3, 2020, were enrolled in primary analysis, and the median follow-up was 709 days. After evaluating the data integrity artificially, 1202 patients were dropped out because of different degrees of data missing, and 649 patients were included in our data analysis and model building finally (Fig. 1), of whom 492 were men and 157 were women, 135 patients suffered from tumor recurrence and 514 patients did not, the tumor recurrence rate is 20.8% in the study cohort. These patients were randomly divided into the training set and testing set with a ratio of 7:3, such that 456 and 193 patients were included in the training set and testing set, respectively. We completed the related comparison and found no significant statistical difference for every included clinical item except for operation time. The baseline characteristics of all patients in two sets are shown in Table 1.

Fig. 1.

Flow diagram of study design

Table 1.

Characteristics of the 649 patients, who were diagnosed with gastric cancer and underwent radical gastrectomy, are divided into the training set and testing set randomly

Items ($\overline{\mathrm{x}}\pm \mathrm{s}$)/n (%)	Total patients cohort	Training set (n = 456)	Testing set (n = 193)	P value
Gender
Male	492 (75.8)	358 (78.5)	134 (69.4)	0.014
Female	157 (24.2)	98 (21.5)	59 (30.6)
Age	58.20 ± 9.83	58.28 ± 9.36	58.01 ± 10.82	0.747
BMI	22.18 ± 2.97	22.08 ± 3.41	22.43 ± 2.80	0.171
O. time	3.96 ± 1.14	3.89 ± 1.06	4.13 ± 1.30	0.025
Size	4.40 ± 2.43	4.46 ± 2.47	4.27 ± 2.32	0.375
TL	20.69 ± 9.57	20.71 ± 9.45	20.64 ± 9.87	0.934
ML	4.77 ± 6.45	4.35 ± 6.40	4.78 ± 6.58	0.442
Blood	137.81 ± 122.28	138.51 ± 119.93	136.17 ± 127.96	0.824
AFP	8.99 ± 75.21	11.29 ± 89.50	3.54 ± 6.16	0.067
CEA	9.72 ± 49.86	8.96 ± 47.42	11.52 ± 55.3	0.551
CA125	15.42 ± 41.39	15.91 ± 48.77	14.27 ± 11.96	0.644
CA199	37.37 ± 128.93	36.90 ± 126.82	38.47 ± 134.11	0.887
Time (days)	709.36 ± 489.48	715.96 ± 495.62	693.76 ± 475.54	0.598
O. style
Laparoscope	489 (75.3)	344 (75.4)	145 (75.1)	0.933
Open	160 (24.7)	112 (24.6)	48 (24.9)
P. style, n (%)
Adeno-	563 (86.7)	395 (86.6)	168 (87.0)	0.884
Mixed	86 (13.3)	61 (13.4)	25 (13.0)
Differentiation
Poorly	172 (26.5)	120 (26.3)	52 (26.9)	0.757
P-M	211 (32.5)	145 (31.8)	66 (34.2)
Moderately	201 (31.0)	148 (32.5)	53 (27.5)
H-M	38 (5.9)	25 (5.5)	13 (6.7)
Highly	27 (4.2)	18 (3.9)	9 (4.7)
P. stage
I	189 (29.1)	131 (28.7)	58 (30.1)	0.936
II	128 (19.7)	91 (20.0)	37 (19.2)
III	332 (51.2)	234 (51.3)	98 (50.8)
T. stage
1	121 (18.6)	81 (17.8)	40 (20.7)	0.755
2	99 (15.3)	69 (15.1)	30 (15.5)
3	29 (4.5)	22 (4.8)	7 (3.6)
4	400 (60.6)	284 (62.3)	116 (60.1)
N. stage
0	274 (42.2)	197 (43.2)	77 (39.9)	0.082
1	261 (40.2)	182 (39.9)	79 (40.9)
2	73 (11.2)	55 (12.1)	18 (9.3)
3	41 (6.3)	22 (4.8)	19 (9.8)
Location
Proximal	122 (18.8)	88 (19.3)	34 (17.6)	0.874
Gastric body	233 (35.9)	162 (35.5)	71 (36.8)
Distal	294 (45.3)	206 (45.2)	88 (45.6)
Complication
Yes	54 (8.3)	39 (8.6)	15 (7.8)	0.742
No	595 (91.7)	417 (91.4)	178 (92.2)
Chemotherapy
Yes	571 (88.0)	403 (88.4)	168 (87.0)	0.634
No	78 (12.0)	53 (11.6)	25 (13.0)
Recurrence
Yes	135 (20.8)	89 (19.5)	46 (23.8)	0.216
No	514 (79.2)	367 (80.5)	147 (76.2)

Complication serious diseases after operation, such as bleeding, respiratory infection, anastomotic fistula and gastroparesis

O. time operation time, O. style operation style, Time recurrence-free survival time, Blood intraoperative blood loss (IBL), P. stage pathological stage, P-M poorly-moderately differentiation, H-M highly-moderately differentiation. TL total lymph nodes, ML metastatic lymph nodes

Related risk variables selection in the training set

We set “family = binomial” in the Glmnet function because the dependent variable in our study is a dichotomous variable and then run Lasso regression; we chose the best choice of lambda (lambda.min = 0.021) and found the contribution coefficient of 16 variables were compressed to zero, and seven variables, pathological stage, tumor size, the number of total lymph nodes, the number of metastatic lymph nodes, intraoperative blood loss (IBL), the level of AFP and CA199, showed their correlation to the dependent variable. The variable selection is shown in Fig. 2.

Fig. 2.

variable selection by LASSO regression

Predictive model construction

Based on the seven variables selected in LASSO regression, we used multiple Cox regression to fit the survival model in the training set, and we got the coefficient for these seven variables (Table 2). Finally, we built the Cox model and got every patient’s risk index (RI); meanwhile, we fitted a nomogram to further present the recurrence risk for patients with gastric cancer (Fig. 3).

$$RI=0.864\ast \mathrm{P}.\mathrm{stage}+0.183\ast Size++0.088\ast ML+0.002\ast IBL+0.001\ast \mathrm{AFP}+0.001\ast CA199-0.032\ast TL$$

Table 2.

Multiple Cox regression analysis of the risk factors for the recurrence of gastric cancer in the training set

Variables	Coef	Exp (coef)	SE	Z value	P value	95% CI
P.stage	0.864	2.373	0.251	3.440	0.001	(1.4502, 3.8814)
Size	0.183	1.200	0.042	4.348	< 0.01	(1.1056, 1.3035)
TL	− 0.032	0.968	0.013	− 2.488	0.013	(0.9441, 0.9932)
ML	0.088	1.092	0.017	5.220	< 0.01	(1.0563, 1.1283)
AFP	0.001	1.001	0.001	1.413	0.158	(0.9997, 1.0020)
CA199	0.001	1.001	0.001	1.649	0.099	(0.9998, 1.0020)
IBL	0.002	1.002	0.001	1.997	0.046	(1.0000, 1.0033)

Fig. 3.

Nomogram model based on the variables selected by LASSO regression

Model validation

According to the RI, we got the cutoff value in the training group and testing group; they were 3.493 and 3.099, respectively. The patients in the training set were divided into a higher and a lower recurrence risk group with the cutoff value. We further drew the ROC curves to value the predictive model in the training set and testing set (Fig. 4). We found that the model showed excellent predictive capacity (AUC = 0.840 in the training group and AUC = 0.756 in the testing group). We fitted the KM curve in these two groups and found a significant statistical difference in recurrence-free time between these two groups (Fig. 5). We also drew DCA curve and calibration plot to evaluate the model. In the calibration plot (Fig. 6), the ideal diagonal 45-degree line represents perfect prediction. We observed a high consistency between the predicted results and the observed results from both the training set and testing set. DCA curve measures net benefit on the Y-axis by adding true positives and subtracting false positives, while threshold probability is represented on the X-axis. We found that all DCA curves of our constructed nomogram were consistently higher than two reference lines (Intervention for all, Intervention for none), this indicates that our built nomogram model performs well in clinical practice (Fig. 7). Therefore, we have concluded that our predictive model demonstrates a good fit in both the training group and testing group.

Fig. 4.

ROC curve in the training set and testing set (A, training set; B, testing set)

Fig. 5.

KM curve in the training set and testing set (A, training set; B, testing set)

Fig. 6.

Calibration curve in the training set and testing set (A, training set; B, testing set)

Fig. 7.

DCA curve in the training set and testing set (A, training set; B, testing set)

Discussion

According to the built predictive model, seven variables, pathological stage, tumor size, the number of total lymph nodes, the number of metastatic lymph nodes, IBL, the level of AFP and CA199, are closely related to the recurrence risk of the gastric cancer patients who underwent radical gastrectomy. In fact, researchers have indicated that these significant variables selected by Lasso regression are closely related to tumor recurrence, which further proves that methodology chosen in this study is feasible and effective in some aspects. In this study, we tried to reveal the correlation between these clinical items and the recurrence risk of gastric cancer and thus provide a convenient and visualized model for clinical doctors, with which they can provide personalized treatment plans for their patients.

Overfitting and multicollinearity are common phenomena in small samples or high dimensionality data studies, especially in linear model, which can affect the model’s prediction accuracy and application. Thus, different from the conventional single-factor analysis method, we used LASSO regression to select the related variables. The special character of LASSO regression is that LASSO regression can complete variable selection and regularization while fitting models, it can drop out some variables, which do not contribute much to dependent variable by compressing their impact coefficients to zero via giving an adjustable penalty coefficient lambda. Thus, LASSO regression shows apparent advantages in dropping out insignificant variables and avoiding overfitting compared with conventional analysis methods, especially in high dimensionality data. In this study, we included 23 conventional variables for fitting models in the first time. Although some of them seem to be significant in the prognosis of the patients with gastric cancer, including operation style, pathological style, T stage, N stage and tumor location, these variables were dropped out in LASSO regression finally. In multiple Cox regression, five variables were found as independent factors in gastric cancer recurrence, including pathological stage, tumor size, TL, ML and IBL, which is consistent with previous studies. Pathological stage has been well-known as prognostic factor of gastric cancer because higher pathological stage surely means the deeper depth of tumor invades and more metastatic lymphoid nodes. For IBL, people guessed that IBL could affect the prognosis by causing tumor cell spreading, systemic inflammation and impaired antitumor immunity, however, there is no clear mechanism to explain the correlation between IBL and tumor recurrence. What we can be sure of is that IBL can indeed increase tumor recurrence risk after resection. Arno Kornberg et al. indicated that IBL was associated with increased likelihood of tumor recurrence following liver transplant (LT) for hepatocellular carcinoma (HCC) (Kornberg et al. 2016). Bin Liu et al. found that excessive IBL was an independent predictor of HCC recurrence after LT, especially in patients with vascular invasion, and IBL > 4 L was independently associated with HCC recurrence for patients with vascular invasion (Liu et al. 2015); the similar results were also observed in colorectal cancer, pancreatic cancer and gastric cancer(Jiang et al. 2013; Kamei et al. 2009; Tamagawa et al. 2020). It is well known that lymphatic metastasis is an important way for tumor cell diffusion, malignant tumors can release some biological factors to induce lymphatic vessel expansion in primary tumors and in draining sentinel lymph nodes, thereby promoting lymphatic metastasis (Karaman and Detmar 2014). Hideya Kashihara et al. also proved this result in their study, they found that lymph node metastasis is an independent risk factor for gastric cancer recurrence after laparoscopic gastrectomy (Kashihara et al. 2017). In addition, serum marker is a convenient way for detecting the occurrence of cancer. Chikashi Shibata et al. demonstrated that CA199 was likely to be more useful for the detection of recurrence after curative gastrectomy for gastric cancer compared to CEA (Shibata et al. 2022). Ya-Kun Wang et al. observed that liver metastasis rate was significantly higher in the AFP ≥ 160 ng/mL gastric cancer group than in the AFP < 160 ng/mL (Wang et al. 2018).

In fact, some of the variables, which were dropped out in our studies could also impact the recurrence of gastric cancer in other researchers’ studies, including age, gender, weight, BMI, histological type, operation style, operation time, T stage, N stage and chemotherapy. Lelisho ME et al. explored the risk factors of gastric cancer by fitting an inverse Gaussian frailty model, and they observed that the gender, tumor size, treatment taken, vascular invasion, disease stage, helicobacter pylori infection and histological type were the determining prognostic factors (Lelisho et al. 2021). In another study, people indicated that extent of lymphadenectomy, depth of tumor invasion, lymph node metastasis and number of negative lymph nodes were closely related to gastric cancer recurrence; they further found that extent of lymphadenectomy and T4b status were independent predictors for locoregional recurrence, histological type and T4b status for peritoneal recurrence and N stages for distant metastasis (Jiao et al. 2020). R Buzzoni et al. concluded that T stage was closely related to the rate of locoregional recurrences, whereas N stage contributed to distant recurrence of gastric cancer (Buzzoni et al. 2006). Chengmao Zhou et al. fitted different models to predict the recurrence of gastric cancer with different machine learning algorithms, all of them showed excellent predicting capacity, and they found BMI, operation time, weight and age as the first four factors affecting postoperative recurrence of gastric cancer (Zhou et al. 2021). As we all know, patients with gastric cancer can benefit from adjuvant chemotherapy after radical resection, however, we did not observe the negative correlation between chemotherapy and tumor recurrence. The main reason for this phenomenon is that the patients who accept chemotherapy are always in an advanced tumor stage, and the patients without chemotherapy are in an early tumor stage, thus the recurrence rate in chemotherapeutic individuals is higher than that in non-chemotherapeutic individuals. In order to make our study more convinced, we further explored the difference between chemotherapeutic individuals and non-chemotherapeutic individuals. As shown in the attached Supplement Table 1, the recurrence rate in chemotherapeutic group is 20.8%, which is higher than that in non-chemotherapeutic group (14.1%), meanwhile, the chemotherapeutic individuals are characterized with higher pathological stage, T stage, N stage and poorer differentiation, which are found as risk factors in the prognosis of gastric cancer. In addition, nonstandard chemotherapeutic scheme designed by clinical doctors and irregularity of patients in accepting chemotherapy maybe another reason for it. Although these variables seems to be important in the prognosis of gastric cancer, they were dropped out by LASSO regression finally because of bare contribution to dependent variable, and thus simplifying the model to some extent.

We evaluated the predictive models via several curves. According to the cutoff value, we got a good score in AUC and calibration plot, and the recurrence-free time of the lower risk group is significantly longer than that in the higher risk group. Although all the results indicated that the model showed a good fit, obvious limitations still exist in our study. The main one was that this study was completed with single cohort data, which could lead to different biases in related variables selection and model construction. In addition, the lack of prospective validation is another problem. Thus, we will try to fit another model by including more samples from multiple study cohorts; meanwhile, we will build a prospective cohort for built model validation in the future.

Conclusions

The seven variables were selected by LASSO regression and verified by multivariable Cox regression and nomogram, including pathological stage, tumor size, the number of total lymph nodes, the number of metastatic lymph nodes, IBL, the level of AFP and CA199, are significant in identifying recurrence risk for gastric cancer patients after radical gastrectomy. The built multivariable Cox model and nomogram showed excellent predicting capacity and could help clinical doctors to develop personalized treatment plans for their patients.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

BH: contributed original draft preparation and data cleaning, FD did investigation and data cleaning. YL: performed methodology and supervision. All authors approved the submitted version.

Funding

This work was funded by Special Research Project of Lanzhou University Serving the Economic Social Development of Gansu Province (054000282), Major Science and Technology Special Project of Gansu Province (20ZD7FA003), Fundamental Research Funds for the Central Universities (lzujbky-2022-sp08); Medical Innovation and Development Project of Lanzhou University (lzuyxcx-2022-154, lzuyxcx-2022-141).

Availability of data and materials

Datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Conflict of interest

The authors declare no competing interest.

Ethics approval and consent to participate

The study protocol was approved by the Ethics Committee of the Second Hospital of Lanzhou University and conducted according to the principles of the Declaration of Helsinki. All methods were performed in accordance with the relevant guidelines and regulations. Written informed consent was obtained from all patients.

Consent for publication

Not applicable.