Comparative study of XGBoost and logistic regression for predicting sarcopenia in postsurgical gastric cancer patients

Abstract

The use of machine learning (ML) techniques, particularly XGBoost and logistic regression, to predict sarcopenia among postsurgical gastric cancer patients has gained significant attention in recent research. Sarcopenia, characterized by the progressive loss of skeletal muscle mass and strength, is a serious concern in these patients due to its association with poor postoperative outcomes, including increased morbidity and mortality. In this study, machine learning was used to establish a risk prediction model for sarcopenia in patients with gastric cancer undergoing gastrectomy to facilitate early intervention and reduce the incidence of postoperative complications. Gastric cancer patients who underwent surgery at a tertiary comprehensive hospital in Nanjing (China) from January 2022 to December 2023 were retrospectively included in this study, and their clinical and follow-up data were collected. The XGBoost model and multivariate logistic regression analysis model were used to screen the factors related to postoperative outcomes, and the results of the two models were compared. The area under the receiver operating characteristic (ROC) curve (AUC), sensitivity and specificity were calculated to evaluate the predictive value of the XGBoost model. The SHAP (SHapley Additive exPlanations) method was used to explain the XGBoost model and determine the impact of features on the prediction model. A total of 231 postoperative gastric cancer patients were included in this study, of whom 128 (55.4%) developed sarcopenia. The results of the univariate analysis and LASSO (Least Absolute Shrinkage and Selection Operator) regression were cross-validated, and 5 key study variables were ultimately determined: serum albumin, comorbid diabetes, operation style, nutritional score, and ECOG (Eastern Cooperative Oncology Group) performance status score. The XGBoost model has slightly better AUC (0.987, 95% CI: 0.976-0.998) than the logistic regression model (0.918, 95% CI: 0.873-0.963) in the training set. The SHAP analysis showed that in the XGBoost model, diabetes, nutritional score, and serum albumin have a greater impact on the sarcopenia risk prediction after gastric cancer surgery, especially the impact of diabetes and nutritional score is the most significant, followed by the ECOG performance status score, and operation style has the least impact. In summary, the machine learning-based sarcopenia prediction model constructed in this study provides a valuable decision support tool for clinical screening and intervention of sarcopenia.

Introduction

Gastric cancer (GC) remains one of the most prevalent malignant neoplasms affecting the digestive tract, ranking fifth globally in incidence and fourth in mortality^1,2. Recent advances in prognostic modeling have demonstrated the clinical utility of nomograms in predicting recurrence patterns, with studies by Liu et al. revealing distinct predictors for early and late recurrence, including nutritional markers and inflammatory indices^3,4. While these models have enhanced risk stratification for tumor recurrence, postoperative complications such as sarcopenia have emerged as critical determinants of long-term survival and quality of life⁵. Sarcopenia, characterized by progressive skeletal muscle loss, affects 20–40% of GC patients postoperatively and is strongly associated with prolonged hospital stays, reduced chemotherapy tolerance, and increased mortality^6–8. Current evidence predominantly focuses on preoperative predictors including age, tumor stage, and baseline body composition^9–11, yet neglects dynamic postoperative variables highlighted in recurrence models^3,4, such as nutritional deterioration (e.g., prealbumin < 70.1 mg/L), physical inactivity, and systemic inflammation. This oversight is particularly significant given that 53% of patients exhibit > 10% body weight loss within six months post-gastrectomy, exacerbating sarcopenia progression^3,12. To address this gap, we developed a dual-phase predictive framework integrating postoperative nutritional parameters, activity metrics, and inflammatory profiles with machine learning algorithms. Utilizing logistic regression for interpretability and XGBoost for nonlinear pattern recognition, this study aims to establish a clinically actionable model for sarcopenia risk stratification, thereby enabling targeted nutritional and rehabilitative interventions to mitigate functional decline. Clinical trial number: not applicable.

Methods

Study population

This study selected research variables through literature review and retrospective analysis of medical records. A total of 231 patients who underwent gastric cancer surgery and were admitted to the Department of Gastrointestinal Surgery at a tertiary comprehensive hospital in Nanjing (China) from January 2022 to December 2023 were selected for this study. Data information was collected through medical records review, questionnaire survey, and on-site evaluation. Inclusion criteria were as follows: (1) patients aged 18–80 years old¹³; (2) all patients underwent endoscopic examination, imaging evaluation, and pathological confirmation before surgery to confirm the diagnosis; the seventh edition tumor staging system released by the American Joint Committee on Cancer (AJCC) in 2010 was used to classify the tumor as stage I, II, or III; (3) patients between 1 and 12 months after gastric cancer surgery¹⁴; (4) patients received a combination chemotherapy regimen of oxaliplatin and tigio. Exclusion criteria were as follows: (1) patients who already had sarcopenia before surgery; (2) patients with severe comorbidities, such as heart disease, liver disease, or other important organ failures; (3) patients had received treatment for pancreatic cancer or colorectal cancer and other digestive system tumors in the past; (4) patients with incomplete medical records; (5) serious complications occurred after the second surgery.

Treatment methods

All patients underwent surgical resection, which was planned and completed by the same surgical team. A total of 200 patients (87%) underwent laparoscopic gastrectomy, 31 patients (13%) underwent open gastrectomy, of whom 64 patients (28%) underwent proximal resection, 93 patients (40%) underwent distal resection, and 74 patients (32%) underwent total gastrectomy.

Diagnostic criteria for sarcopenia

Based on the recommendations of the 2019 Asian Working Group for Sarcopenia (AWGS 2019), muscle mass should be assessed using dual-energy X-ray absorptiometry (DXA) or multifrequency bioelectrical impedance analysis (BIA), with height-based correction¹⁵. Given the available medical facilities at our institution, we utilized the InBody 270 device to analyze body composition and calculate the skeletal muscle mass of the limbs. According to BIA criteria, a skeletal muscle mass of less than 7.0 kg/m² in men and less than 5.7 kg/m² in women is considered indicative of sarcopenia. Additionally, muscle strength was measured using a Hand Grip Dynamometer (Xiangshan electronic grip strength meter), with values below 28 kg for men and 18 kg for women classified as weak. Alternatively, physical function was assessed using the 6-meter walking speed test, where a walking speed below 1.0 m/s was used as a criterion for sarcopenia.

Observation indicators

The baseline data of each patient before the first surgery was organized. Demographic data includes: gender, age, body mass index (BMI), surgical approach, tumor stage, number of chronic diseases, serum albumin, and postoperative period. Disease-related factors include: combination of diabetes, number of other chronic diseases, tumor stage, scope of surgical resection. Physiological and biochemical indicators include: serum albumin concentration, white blood cell count, platelet count, lymphocyte count. Factors related to treatment include: surgical approach, postoperative period. Other indices include: nutritional score, ECOG (Eastern Cooperative Oncology Group Functional Status Score) performance status score.

Statistical analysis

Continuous variables with normal distribution are presented as the mean ± standard deviation, while non-normal distribution is presented as the median and interquartile range (IQR). Categorical variables are expressed as frequency and percentage. Continuous variables were tested using a t-test or Mann Whitney U test. Categorical variables were tested using the chi square test or Fisher’s exact probability method. A test level of α = 0.05 and P < 0.05 were used for statistical difference, bilateral test. In the modeling process, single factor analysis and LASSO (Least Absolute Shrinkage and Selection Operator) regression analysis were used to introduce significant variables related to Sarcopenia in postoperative gastric cancer patients into multivariate logistic regression analysis. Variables that have a significant impact on the predictive model were selected using the forward stepwise method. Based on the partial regression coefficients of these variables, an equation for evaluating the risk of Sarcopenia in postoperative gastric cancer patients was constructed. In addition, using the rms R software package (version 4.3.1) to generate nomograms, and the risk assessment of the patient is presented graphically. Simultaneously utilizing XGBoost machine learning to construct a predictive model and conducting explanatory analysis on the model using the SHAP (SHapley Additive exPlanations) method. Model validation and evaluation: the bootstrap method (repeated sampling 100 times) was used to perform internal validation of the model. The discriminative ability of the model was evaluated using the area under the receiver operating characteristic (ROC) curve (AUC) of the subjects, while the calibration of the model was measured by the Hosmer-Lemeshow test. In general, the closer the AUC value is to 1, the stronger the discriminative ability of the model. If the AUC is greater than 0.75, it indicates that the model has good discriminative ability. In the Hosmer Lemeshow test, a P-value greater than 0.05 usually indicates a good fit of the model. In addition, the predicted curve being close to the reference curve also indicates that the model has good calibration.

Results

General Information Analysis of the Research Object (Table 1).

Table 1.

General information analysis of the research object.

A total of 231 patients were included in this study, including 177 males (77%), 54 females (23%), with an average age of 66 (60, 70) years, 87 diabetes patients (38%), 103 non-sarcopenia patients (45%), 128 sarcopenia patients (55%), 154 patients (67%) within 6 months after surgery, 77 patients (33%) within 6–12 months after surgery. Among sarcopenia patients, 81 patients (63%) had sarcopenia within 6 months, and 37 patients (37%) had sarcopenia within 6–12 months. There was no statistically significant difference between the two groups in terms of postoperative period and platelet count (P > 0.05). There were significant differences in gender, age, BMI, surgical resection range, operation style, tumor stage, diabetes, white blood cell count, absolute lymphocyte count (ALC), serum albumin, nutritional score, ECOG performance status score, et al. (P < 0.05).

Univariate analysis of postoperative Sarcopenia in gastric cancer patients (Table 2).

Table 2.

Univariate analysis of postoperative sarcopenia in gastric cancer patients.

Variable	β	P	OR	95%CI
				Lower	Upper
Age	0.047	0.004	1.049	1.016	1.085
Gender	0.775	0.014	2.172	1.174	4.080
BMI	3.205	< 0.001	24.663	7.301	154.035
Diabetes	3.554	< 0.001	34.940	14.505	104.475
Chronic disease number	0.092	0.326	1.096	0.914	1.320
Postoperative time	0.345	0.224	1.412	0.812	2.479
Operation style	−1.106	0.007	0.331	0.143	0.723
Stages
I			1.000
II	1.674	0.007	5.333	1.664	19.038
III	0.397	0.365	1.487	0.638	3.613
Surgical resection range
Proximal resection			1.000
Distal resection	1.610	< 0.001	5.005	2.550	10.124
Total gastrectomy	0.934	0.008	2.544	1.281	5.159
WBC	−0.020	0.539	0.980	0.917	1.045
LY	−0.003	0.099	0.997	0.993	1.001
ALC	−0.343	0.090	0.710	0.462	1.013
Albumin	−0.140	< 0.001	0.870	0.812	0.927
ECOG	1.674	< 0.001	5.334	3.284	9.375
Nutrition	1.027	< 0.001	2.792	2.122	3.772

Using 231 training datasets, the presence or absence of sarcopenia was used as the dependent variable, and 15 selected influencing factors were used as independent variables for univariate analysis. The relevant values are shown in Table 2. The results showed that there were significant differences in gender, tumor stage II, tumor stage III, distal resection, total gastrectomy, operation style, age, serum albumin, BMI, ECOG performance status score, nutritional score, and diabetes (P < 0.05), which could be regarded as an independent influencing factors for postoperative sarcopenia in gastric cancer patients.

LASSO regression analysis results (Fig. 1).

Fig. 1.

 LOSSA Regression Variable Selection.

The LASSO regression model was utilized to select and obtain the eight simplest model coefficients, namely, combined diabetes, white blood cell count, absolute lymphocyte count, blood platelet count, operation style, serum albumin, nutrition score, and ECOG performance status score.

Screening of influencing factors (Fig. 2).

Fig. 2.

Venn diagram.

To obtain a more accurate and simple prediction model, 12 variables were screened out by single factor analysis, and 8 variables were screened out by LASSO regression analysis. Using a Venn diagram, the intersection of the two was taken to obtain 5 variables, including serum albumin, combined diabetes, operation style, nutrition score and ECOG performance status score.

Multivariate logistic regression analysis (Table 3).

Table 3.

Multivariate logistic regression analysis.

Variable	Coef	S.E.	Wald	P	OR	95%CI
						Lower	Upper
Intercept	0.189	2.310	0.080	0.935	1.208	0.120	12.174
Albumin	−0.098	0.055	−1.78	0.075	0.907	0.858	0.958
Diabetes	2.730	0.601	4.540	< 0.001	15.336	8.407	27.975
Operation style	−1.624	0.710	−2.290	0.022	0.197	0.097	0.401
Nutrition	0.943	0.283	3.340	0.001	2.568	1.936	3.407
ECOG	0.283	0.437	0.650	0.517	1.327	0.857	2.055

Using 231 training datasets, the presence or absence of sarcopenia was used as the dependent variable, and the five variables selected above were used as independent variables. Multifactor logistic regression analysis was performed. The results showed that serum albumin, combined diabetes, operation style, nutrition score, and ECOG performance status score were closely related to the occurrence of postoperative sarcopenia in patients with gastric cancer.

Constructing a logistic regression model.

The final model obtained is as follows:

Logit (P) = 0.189 − 0.098×serum albumin + 2.730×combined diabetes-1.624×operation style + 0.943×nutrition score + 0.283×ECOG performance status score.

According to this model, a nomogram of sarcopenia in patients with gastric cancer after surgery was drawn (Fig. 3). By analyzing the categorical variables in the figure, the scores of each index can be obtained separately, and by calculating the sum of these scores, the possibility of sarcopenia in patients after gastric cancer surgery can be obtained.

Fig. 3.

Nomogram for Predicting the Risk of Sarcopenia in Postoperative Gastric Cancer Patients.

The receiver operating characteristic (ROC) curve (Fig. 4).

Fig. 4.

ROC curve of Logistic model.

The results of this model show that the AUC of the training set ROC curve is 0.918 (Fig. 6), the 95% confidence interval is 0.873–0.963, the sensitivity is 0.895, and the specificity is 0.849. The AUC of the test set ROC curve is 0.900, the 95% confidence interval is 0.831–0.969, the sensitivity is 0.818, and the specificity is 0.838. These results indicate that the model has a good ability to distinguish whether Sarcopenia occurs in postoperative gastric cancer patients.

Fig. 6.

Calibration curve of Logistic model.

Decision curve (Fig. 5).

Fig. 5 .

Decision Curve of Logistic Model.

The decision curve analysis (DCA) is performed to assess the clinical utility of the prediction model using both training and testing datasets, as shown in Fig. 5. The horizontal axis represents the threshold probability, and the vertical axis represents the net benefit of zero for all study subjects who are not sarcopenia patients. The gray curve represents the net benefit assuming that all included study subjects are sarcopenia patients, and the red and blue curves represent the net benefit predicted using this column chart model for patients with sarcopenia. When the threshold probability is between 10 and 98%, the net benefit of using this prediction model to evaluate the probability of Sarcopenia in gastric cancer postoperative patients is greater than that of patients who do not use it. This indicates that the prediction model has good clinical utility.

Calibration curve (Fig. 6).

The calibration curves are drawn using clinical data from the training and testing datasets separately, as shown in Fig. 6. The horizontal axis represents the predicted probability of patients developing Sarcopenia, while the vertical axis represents the actual proportion of patients developing Sarcopenia. The curve shows that the predicted value of the model for predicting Sarcopenia in postoperative gastric cancer patients is basically consistent with the actual value. The calibration ability of the model was evaluated by the Hosmer-Lemeshow goodness-of-fit test, and the test value of the model was=15.116, P = 0.057, revealing that there was no statistically significant difference between the predicted value and the actual value of the prediction model (P > 0.05), indicating that the model has a good fitting degree.

XGBoost machine learning model.

Performance Evaluation of XGBoost Machine Learning Model.

The performance evaluation related variables of the model are shown in Table 3. The AUC of the XGBoost model in predicting the risk of Sarcopenia in postoperative gastric cancer patients (Fig. 7) is higher (training set: 0.987, 95% CI: 0.976–0.998; test set: 0.909, 95% CI: 0.840–0.977). The XGBoost model performs well in predicting the risk of Sarcopenia in postoperative gastric cancer patients, with high sensitivity, specificity, and prediction accuracy. The accuracy of the XGBoost model in predicting the risk of Sarcopenia in postoperative gastric cancer patients was evaluated by analyzing the calibration curves (Fig. 8) and clinical decision curves (Fig. 9) of the training and testing datasets.

Fig. 7.

ROC curve of XGBoost mode.

Fig. 8.

Calibration curve of XGBoost model.

Fig. 9.

Decision Curve of XGBoost Model.

Explanation of the XGBoost Machine Learning Model (Fig. 10).

Fig. 10.

Importance chart of feature variables.

The SHAP summary diagram of the XGBoost model shows the impact of feature variables on the prediction model. The included feature variables are sorted in descending order of the SHAP mean absolute value. In descending order, they are “diabetes”, “nutrition score”, “serum albumin”, “ECOG performance status score”, and “operation style”. According to the prediction model, the higher the SHAP value of the feature variable, the greater the likelihood of developing sarcopenia. Based on the importance diagram of the characteristic variables in the following figure, it can be concluded that:

According to the SHAP interpretative analysis, diabetes, nutritional score and serum albumin are the main influencing factors for sarcopenia in patients with gastric cancer after surgery. The combination of diabetes and high nutrition score will increase the risk of sarcopenia, while high serum albumin will reduce the risk of sarcopenia. Second, the ECOG performance status score has a certain impact on the risk of developing sarcopenia, that is, the worse the activity of gastric cancer patients, the higher the risk of postoperative sarcopenia. The impact of operation style is relatively small.

Explanation of machine learning models at the patient level (Fig. 11).

Fig. 11.

Predicted SHAP values for patients.

To demonstrate how the XGBoost model evaluates the contribution of individual patient features, the SHAP method is used to try to explain the individual predictions of two patients. The color represents the contribution of each feature, thus blue indicates that the feature has a negative impact on prediction (arrow to the left, SHAP value decreases), and red indicates that the feature has a positive impact on prediction (arrow to the right, SHAP value increases). The length of the color bar represents the strength of the contribution, and E [f (x)] represents the SHAP reference value, which means the average predicted value of the model is 0.465. For a “true positive” group of patients, the XGBoost model predicts a SHAP value of 3.100 for sarcopenia, exceeding the reference value, thus indicating that the model predicts the presence of sarcopenia in the patient. The predicted SHAP value of the XGBoost model for the “true negative” group is −1.960, which is lower than the reference value, indicating that the model predicts that the patient does not have Sarcopenia. The predicted results are consistent with the actual situation.

Discussion

This study included 231 patients and analyzed various factors associated with sarcopenia to evaluate their impact on postoperative recovery. The study found a high proportion of male patients (77%), with most patients concentrated around the age of 66, which is consistent with the high incidence of sarcopenia in the elderly population¹⁶.

The study also revealed that diabetes occurred in 38% of patients and was closely related to the development of sarcopenia. After gastric cancer surgery, patients are prone to unstable blood sugar levels in a short period of time, often resulting in a state of hyperglycemia. The metabolic disorders and inflammatory reactions caused by hyperglycemia are underlying mechanisms that may lead to muscle loss. From the perspective of metabolic disorder, diabetes patients often have metabolic abnormalities, such as high blood sugar and blood lipid. High blood sugar levels may affect the uptake and utilization of glucose by muscle cells, resulting in insufficient energy supply to muscles, which in turn affects normal metabolism and growth of muscles, ultimately leading to a decrease in muscle mass¹⁷. Meanwhile, abnormal lipid metabolism may also disrupt the balance of lipid metabolism within muscles, resulting in adverse effects on muscle cells¹⁸. Regarding the inflammatory response, a chronic inflammatory status exists in diabetes patients for a long time. The continuous release of inflammatory cytokines will damage muscle cells, inhibit the synthesis of muscle proteins, and promote the decomposition of muscle proteins, ultimately leading to the occurrence and development of sarcopenia. In future research, we will further study the relationship between patients’ blood glucose values and myopenia, not just limited to whether the patients have diabetes.

Sarcopenia, characterized by the loss of skeletal muscle mass and strength, is a major issue for patients undergoing gastric cancer surgery. Various studies have identified multiple risk factors associated with the development of sarcopenia in these patients, particularly focusing on preoperative conditions, surgical procedures, and postoperative outcomes. One of the main findings is that there is a significant difference in the incidence of sarcopenia among cancer patients. A study showed that 52.5% of cancer patients suffer from sarcopenia, emphasizing the importance of using standardized diagnostic criteria to ensure accurate incidence and risk assessment¹⁹. The association between sarcopenia and postoperative adverse outcomes has been fully demonstrated, with sarcopenia patients having a higher incidence of complications, longer hospital stays, and increased medical costs^20,21. This study collected relevant data from cancer postoperative patients and excluded patients who had previously suffered from sarcopenia. The study found that postoperative nutritional status, functional status, and treatment-related factors have a significant impact on the occurrence and development of sarcopenia. These findings fill the gap in previous research and help medical professionals take effective measures in a timely manner, develop effective management strategies, take early prevention measures, make early diagnosis, and perform early intervention.

Our present study had some limitations. This is a single-center, retrospective study, where all patients underwent surgery performed by the same surgical team and received the same chemotherapy regimen. This design minimizes the potential influence of geographical variations, surgical skills, and chemotherapy regimens on postoperative muscle loss in patients. In the future, we will continue to collect data from multiple centers to further refine the model.

The association between sarcopenia and postoperative adverse outcomes has been well-documented across various malignancies. A recent meta-analysis by Liu et al. (2024) highlighted that preoperative sarcopenia significantly correlates with worse overall survival (HR = 1.53, P < 0.00001) and progression-free survival (HR = 1.55, P < 0.00001) in pancreatic cancer patients undergoing curative-intent surgery, underscoring its prognostic value beyond gastrointestinal malignancies like gastric cancer²². While their findings did not establish a direct link between sarcopenia and major complications in pancreatic cancer, our study identified diabetes, nutritional status, and functional decline as critical modifiable risk factors for sarcopenia in gastric cancer patients, suggesting cancer-specific nuances in sarcopenia’s clinical impact. These differences may stem from variations in tumor biology, surgical stress, and metabolic demands across cancer types. For instance, pancreatic cancer’s aggressive nature and high catabolic burden may amplify muscle loss, whereas gastric cancer’s association with malnutrition and postoperative dietary restrictions might exacerbate sarcopenia through distinct pathways^6,22. Nevertheless, the consistent evidence across studies reinforces the need for standardized sarcopenia screening and intervention protocols in oncology.

Conclusions

In this study, we found that serum albumin, diabetes mellitus, operation style, nutritional score, and ECOG performance status score, were the influencing factors for the risk of developing sarcopenia after gastric cancer. Nursing interventions focus on nutrition and exercise, both of which are increasingly being studied. Such assessing the nutritional status of the patient, combining the patient’s serum albumin and other nutritional indicators, developing an individualized diet plan for the patient, increasing the intake of protein, amino acids, and trace elements, and, if necessary, oral nutritional supplements (ONS) for patients who can eat ad libitum²³. Patients with diabetes mellitus should be given diabetic dietary guidance and urged to perform accurate blood glucose monitoring. At the same time, exercise guidance can be used to incorporate resistance exercise combined with aerobic exercise into the daily rehabilitation training of patients²⁴, so as to continuously improve the quality of nursing services and the connotation of specialties.

Author contributions

All authors contributed to the study’s conception and design. Preparation of materials, data collection, and analysis were performed by S.S., Y.L.L., X.P.W.,J.J.M.,X.N. and A.L. The initial draft of the manuscript was written by Y.J.G. and X.Z. All authors provided valuable input and feedback on previous versions of the manuscript. All authors read and approved the final manuscript.

Funding

This study was funded by Nanjing Health Science and Technology Development Special Fund Project [YKK21230] and Project of Chinese Hospital Reform and Development Institute, Nanjing University [NDYG2023057].

Data availability

The datasets used in the manuscript are available from the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Ethics approval and consent to participate

The study was approved by the Ethics Committee of Nanjing Jiangning Hospital (Approval No. 2023-03-018-K01). All procedures performed in this study were in accordance with the ethical standards of the institutional and/or national research committee. This study is a retrospective study and has an approval of an informed consent waiver.