Machine learning assisted radiomics in predicting postoperative occurrence of deep venous thrombosis in patients with gastric cancer

Abstract

Background

Gastric cancer patients are prone to lower extremity deep vein thrombosis (DVT) after surgery, which is an important cause of death in postoperative patients. Therefore, it is particularly important to find a suitable way to predict the risk of postoperative occurrence of DVT in GC patients. This study aims to explore the effectiveness of using machine learning (ML) assisted radiomics to build imaging models for prediction of lower extremity DVT occurrence in GC patients after surgery.

Methods

Included in this retrospective study were eligible patients who underwent surgery for GC. CT imaging data from these patients were collected and divided into a training set and a validation set. The least absolute shrinkage and selection operator (LASSO) algorithm was applied to reduce the dimensionality of variables in the training set. Four machine learning algorithms, known as random forest (RF), extreme gradient boosting (XGBoost), support vector machine (SVM) and naive Bayes (NB), were used to develop models for predicting the risk of lower extremity DVT occurrence in GC patients. These models were subsequently validated using the internal validation set and an external validation cohort.

Results

LASSO analysis identified 10 variables, based on which four ML models were established, which were then incorporated with the clinical characteristics to predict lower extremity DVT occurrence in the training set. Among these models, RF and NB demonstrated the highest predictive performance, achieving an AUC of 0.928, while SVM and XGBoost achieved a slightly lower AUC of 0.915 and 0.869, respectively.

Conclusion

ML algorithms based on imaging information may prove to be novel non-invasive models for predicting postoperative occurrence of DVT in GC patients.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12885-025-13630-1.

Introduction

Cancer patients generally present hypercoagulable state, making venous thromboembolism (VTE) a significant cause of death [1]. Deep vein thrombosis (DVT) of the lower extremities is a common and serious condition among cancer patients, though it is often preventable [2]. Approximately 20% VTE events are related to malignant tumors, with numerous factors affecting the risk of cancer-related thrombosis [3, 4]. Despite the widespread use of VTE scores and prophylactic anticoagulants for high-risk patients, many cancer patients still develop lower extremity DVT in clinical practice [5, 6]. Therefore, it is crucial to develop more accurate methods for predicting the occurrence of lower extremity DVT in cancer patients.

Gastric cancer (GC) remains a leading cause of cancer-related mortality globally, with surgical resection as a cornerstone of curative intent [7]. However, postoperative complications such as lower extremity DVT significantly impact patient survival and recovery, necessitating improved risk stratification strategies [8, 9]. Recent advances in systemic therapies—including immunotherapy, targeted agents, and combination regimens—have reshaped the perioperative landscape for GC [10, 11]. Similarly, biomarker-driven therapies targeting HER2, Claudin 18.2, and FGFR2b are expanding treatment options, yet their interplay with postoperative complications, such as DVT, remains underexplored [12, 13]. Furthermore, the emergence of liquid biopsy-based surveillance [14] and personalized adjuvant approaches [15] underscores the need for holistic models that integrate oncological treatment responses with surgical risk prediction. Based on the stratification of different treatment options, various predictive models have been constructed for predicting the prognosis of GC patients with satisfactory predictive performance [16, 17]. However, few studies have reported prognostic models to predict DVT complications after surgery in GC patients. Machine learning (ML) is considered a potential tool for this purpose [18].

Cachexia is a syndrome characterized by weight and muscle loss (with or without loss of adipose tissue) which cannot be completely reversed by conventional nutritional support [19]. It has been established that sarcopenia is associated with tumor prognosis and treatment response, but few studies have explored whether sarcopenia is linked to the occurrence of lower extremity DVT in patients with malignant tumors [20, 21]. Most retrospective studies currently use muscle information from CT scans of the third lumbar vertebra to define the occurrence of sarcopenia [22]. Previous studies have developed models for disease prediction using imaging omics in various cancer types, including breast cancer [23, 24], lung cancer [25, 26], and liver cancer [27]. These models can be constructed independently or integrated with other omics data to enhance their predictive capabilities [28]. However, limited research has focused on radiomics-based models for predicting DVT. To address this gap, we developed radiomics ML models by utilizing muscle information from CT scans at the third lumbar vertebra to predict the risk of postoperative occurrence of DVT in GC patients, with the goal to improve preventive strategies against DVT.

This study aims to extract muscle information from the third lumbar vertebra CT scans and use ML algorithms to construct four models for predicting lower extremity DVT occurrence in GC patients, hoping that they could provide new non-invasive methods for guiding the prevention of lower extremity DVT in GC patients.

Materials and methods

Study population

In this study, we retrospectively analyzed data from 1201 GC patients who underwent surgical treatment at the First Affiliated Hospital of Wenzhou Medical University and the First People’s Hospital of Wenzhou (Wenzhou, China) between January 2019 and January 2022. The diagnosis of GC was based on clinical symptoms, laboratory indicators, radiological findings and histopathological results. Inclusion criteria were patients (1) aged ≥ 18 years; (2) who were diagnosed with GC by gastroscopy and histological examination before surgery and received surgical treatment; (3) whose preoperative imaging examination showed no distant metastasis in the liver and abdominal cavity; (4) with no history of DVT; (5) with no serious disease that may affect coagulation function and without using drugs that may affect coagulation function; and (6) with complete abdominal CT imaging examination and complete lower extremity vascular ultrasound data. Exclusion criteria were patients with (1) a history of venous thrombosis; (2) a history of using medications that may affect coagulation function, such as aspirin, warfarin, or low molecular weight heparin; and (3) underlying diseases such as acute or chronic renal failure, heart disease, cirrhosis, acute illness, other autoimmune diseases, or chronic infection, and who missed relevant data at admission. Based on these inclusion and exclusion criteria, 106 patients from the First Affiliated Hospital of Wenzhou Medical University were included, of whom 56 developed postoperative thrombosis and 50 did not. Additionally, 60 patients were admitted to Wenzhou First People’s Hospital for external validation, of whom 30 developed thrombosis and the remaining 30 were thrombosis free. This study was approved by the ethics committees of the said hospitals. Due to its retrospective nature, written informed consent from patients was not required. The flowchart of the study is shown in Fig. 1.

Fig. 1.

Flow diagram outlining patient enrollment, classification, and the study design

Diagnostic criteria for lower extremity DVT

The study conformed to the diagnostic guidelines for lower extremity DVT as outlined in the ESMO Clinical Practice Guideline publication, ‘Venous Thromboembolism in Patients with Cancer.‘ [29] Following these recommendations, we utilized Doppler ultrasound of the lower extremity veins, and our findings demonstrated that this technique yielded no false-positive results. Patients were classified into two groups: those who developed lower extremity DVT (Thrombus group) and those who did not (No thrombus group).

Region of interest (ROI) segmentation

Radiomics analysis was conducted in accordance with standardized protocols established in previous studies. CT images of the third lumbar vertebra were independently reviewed by two experienced radiologists, who were blinded to the clinical data. ROI was manually delineated using 3D Slicer software. Radiomic features were then comprehensively extracted from the ROI using Python (version 3.7.0) and the PyRadiomics library (version 3.1.0).

Statistical analysis

This study included 15 basic clinical information variables and 107 imaging features. Continuous data are expressed as the mean or median and interquartile range, depending on their distribution. Data were compared using independent t-tests or Wilcoxon rank-sum tests based on the normality of the variables. Categorical data are expressed as frequency (percentage) and compared using chi-square or Fisher’s exact tests.

The least absolute shrinkage sum selection operator (LASSO) variable selection algorithm was then employed in the training set to identify the most relevant imaging variables and calculate individual coefficients. The LASSO algorithm was used to select the most significant imaging variable effects and reduce the dimensionality of the variables.

Subsequently, the 10 significant variables identified by the LASSO algorithm were included in the ML model for prediction. The training set was used to develop models for each ML algorithm, while the test set was used to evaluate the accuracy of each algorithm. The ML models were built using a system framework that incorporated four different ML algorithms: random forest (RF), extreme gradient boosting (XGBoost), support vector machine (SVM), and Naive Bayes (NB). The receiver operating characteristic (ROC) curve was used to evaluate the discriminative ability of all four ML models.

Visualization of the ML models was presented using SHapley Additive exPlanation (SHAP) analysis, which explains which variables are most important for model predictions and their contributions to overall model performance for a particular prediction. A two-sided p-value < 0.05 was considered statistically significant. All statistical analyses were performed using R software (version 4.0.2), including the tidyverse package (version 1.3.1) and the fastshap package (version 0.0.7).

Results

Information features for ML datasets

A total of 106 patients were included in this study, of whom 56 patients developed venous thrombosis and the remaining 50 patients had no venous thrombosis. The basic clinical characteristics of these patients are shown in Table 1. There were no significant differences in age, gender, body mass index (BMI), tumor size, and tumor TNM grade between the two groups.

Table 1.

The clinical characteristic of patients

Clinical characteristics	Cases	Patients
	(N = 106)	Thrombus (N = 56)	No Thrombus (N = 50)			p-value
Total	106	56	50
Gender				0.1560
Male	71	18	33
Female	35	12	23
Age (years)					0.0937
< 60	24	8	16
≥ 60	82	48	34
Smoking					0.6257
Yes	49	23	26
No	57	33	24
Alcohol
Yes	15	7	8		0.0668
No	91	49	42
BMI
< 24	35	18	17		0.0662
≥ 24	75	38	33
T stage					0.2048
T1 + T2	55	26	29
T3 + T4	51	30	21
N stage					0.0903
Yes	60	31	29
No	46	25	21
M stage					0.1087
Yes	82	45	37
No	24	11	13
Tumor size (cm)					0.1678
< 5	84	39	45
≥ 5	22	16	6
CA199(U/ml)					0.3743
< 37	83	42	41
≥ 37	23	14	9
CA125 (Uml)					0.0844
< 35	83	45	38
≥ 35	23	11	12
CEA (ng/ml)					0.1855
< 5	73	41	32
≥ 5	33	15	18
Neutrophil absolute value (mean,×10^9/L)	4.300	4.400	4.240		0.6706
Lymphocyte absolute count (mean,×10^9/L)	1.65	1.563	1.768		0.1605
Plasma albumin (g/L)					0.1560
< 35	37	18	19
≥ 35	69	38	31

The patients were divided into a training set (n = 75) and a validation set (n = 31) at a 3:1 ratio. The characteristics of the patients in both sets are shown in Table 2. There were no significant differences in age, gender, BMI, tumor size, and tumor TNM grade between the two sets.

Table 2.

The clinical characteristic of patients in the training and validation sets

Clinical characteristics	Cases
	(N = 106)	Training set (N = 75)	Validation set (N = 31)			p-value
Total	106	75	31
Gender				0.2472
Male	71	53	18
Female	35	22	13
Age (years)					0.4132
< 60	24	12	12
≥ 60	82	63	19
Smoking					0.8341
Yes	49	28	21
No	57	47	10
Alcohol
Yes	15	9	6		0.0668
No	91	66	25
BMI
< 24	35	20	15		0.4818
≥ 24	75	55	16
T stage					0.8608
T1 + T2	55	34	21
T3 + T4	51	41	10
N stage					0.3305
Yes	60	39	21
No	46	36	10
M stage					0.0873
Yes	82	54	28
No	24	21	3
Tumor size (cm)					0.2911
< 5	84	61	24
≥ 5	22	14	8
CA199(U/ml)					0.2952
< 37	83	67	24
≥ 37	23	16	7
CA125 (Uml)					0.0844
< 35	83	68	15
≥ 35	23	7	16
CEA (ng/ml)					0.5104
< 5	73	56	17
≥ 5	33	19	14
Neutrophil absolute value (mean,×10^9/L)	4.300	4.1700	4.240		0.2712
Lymphocyte absolute count (mean,×10^9/L)	1.65	1.687	1.573		0.1248
Plasma albumin (g/L)					0.4795
< 35	37	22	15
≥ 35	69	53	16

A total of 60 patients from Wenzhou People’s Hospital were included in this study for external validation, including 30 patients with venous thrombosis and 30 without. The basic clinical characteristics of these patients are summarized in Supplementary Table 1. No significant differences were observed between the two groups regarding age, gender, BMI, tumor size, or TNM grade.

Variable selection

A total of 107 imaging features were extracted from the CT images of the third lumbar vertebra muscle (Supplementary Table 2). The LASSO algorithm was used to identify the 10 most relevant imaging features in the training set, which are: lmc1, lmc2, Correlation, LargeAreaLowGrayLevelEmphasis, LongRunHighGrayLevelEmphasis, Kurtosis, ShortRunLowLevelEmphasis, Elongation, InterquartileRange and MCC. (Supplementary Table 3).

Additionally, we incorporated 10 clinical features to develop clinically relevant ML models. These features included platelet count, CEA, age, neutrophil count, albumin, lymphocyte count, height, weight, BMI, CA125 (Cancer Antigen 125), sex, and CA199 (Carbohydrate Antigen 19 − 9).

Diagnostic performance of ML models

The 10 selected imaging variables were incorporated into four ML models to predict the occurrence of lower extremity DVT (Fig. 2). Among these models, the NB model demonstrated the best prediction performance, with an area under the curve (AUC) of 0.882. The SVM and XGBoost models showed comparable prediction performance, with AUCs of 0.784 and 0.797, respectively. The RF model had a slightly lower AUC of 0.850.

Fig. 2.

ROC curves of four machine learning models constructed using imaging features in the training set. (A-B) ROC curves of NB and SVM using 10 imaging features to predict lower limb DVT. (C-D) ROC curves of XGBoost and RF using 10 imaging features to predict lower limb DVT

Subsequently, we utilized clinical variables to develop four ML models for predicting the occurrence of DVT. Among these, the NB model demonstrated the best predictive performance, achieving an AUC value of 0.843. The SVM and RF models showed comparable performance, with an AUC value of 0.784 and 0.797, respectively. The XGBoost model exhibited a slightly lower AUC value of 0.608 (Figure S1).

Finally, we combined the imaging features with the clinical features and used the above four ML methods to build models and achieved satisfactory prediction results. Among them, RF and NB performed the best, offering an AUC value of 0.928, followed by SVM (AUC value = 0.915) and XGBoost (AUC value = 0.869) (Fig. 3). Additionally, we validated the ML model that combined the clinical features with the imaging features in an external validation cohort. The result showed that the SVM model performed the best, offering an AUC value of 0.830, followed by the RF (0.826), XGBoost (0.821) and NB models (0.705) (Figure S2). In general, combination of the imaging features with the clinical features achieved the best prediction effect.

Fig. 3.

ROC curves of four machine learning models constructed using imaging features combined with clinical features in the training set. (A-B) ROC curves of NB and SVM using 10 imaging features combined with clinical features to predict lower limb DVT. (C-D) ROC curves of XGBoost and RF using 10 imaging features combined with clinical features to predict lower limb DVT

Feature importance

To further investigate the significance of each variable in the combined model, we employed IncNodePurity and %IncMSE to assess the importance of variables in the RF model. As illustrated in Fig. 4, despite discrepancies in the feature importance rankings between the two metrics, lmc2 emerged as the most critical imaging feature, and platelet count as the most important clinical feature.

Fig. 4.

Variable importance of ML models. (A) Using lncnodepurity to define variable importance in the RF model. (B) Using %lncMSE to define variable importance in the RF model. Abbreviations: BMI, body mass index; CA125, Cancer Antigen 125; CEA, Carcinoembryonic Antigen; CA199, Carbohydrate Antigen 19 − 9

Visualization of radiomics-clinical models

SHAP provides a quantitative approach to interpreting the NB and RF method. The SHAP summary plot offers a visually concise representation by illustrating the range and distribution of feature importance relative to the output of the model, linking the value of each feature to its corresponding impact. Features are ranked in descending order of global importance. Each point on the plot represents a SHAP value for a specific feature in an individual patient, which is displayed horizontally and stacked vertically to indicate the density of identical SHAP values. The points are color-coded based on the feature’s value, transitioning from low (blue) to high (red). Similarly, in the RF and NB models, the imaging features lmc2 and lmc1 emerged as the key determinants for distinguishing the occurrence of DVT. Likewise, among the clinical features, platelet count was a crucial factor for identifying DVT in both NB and RF models (Fig. 5).

Fig. 5.

SHAP analysis in NB and RF models. (A) SHAP of the RF model in the training set. (B) SHAP of the NB model in the training set. Abbreviations: BMI, body mass index; CA125, Cancer Antigen 125; CEA, Carcinoembryonic Antigen; CA199, Carbohydrate Antigen 19 − 9

Subgroup analysis

Subsequently, we performed a subgroup analysis. After carefully evaluating both the training and validation sets, we selected the RF to construct a subgroup analysis model. Knowing that age over 70 years is a significant risk factor for DVT, patients were divided into two groups: those aged ≧ 70 years and those < 70 years. In patients aged ≧ 70 years, the RF model achieved an AUC value of 0.975 versus 0.917 in patients aged < 70 years. When patients were stratified by gender, the RF model yielded an AUC value of 0.899 in males and 0.952 in females (Fig. 6).

Fig. 6.

Subgroup analysis of the RF model (A-B) The RF models were constructed in patients with age < 70 (A) and age ≥ 70 (B). (C-D) The RF models were constructed in patients in male (C) and female (D)

Discussion

As an emerging tool, ML has gained widespread application in predicting disease prognosis and evaluating treatment efficacy [30, 31]. In our study, we developed four novel ML models by utilizing the muscle data from third lumbar CT scans to predict the risk of lower extremity venous thrombosis in GC patients. All these models offer improved accuracy in predicting lower extremity venous thrombosis and may contribute to more comprehensive prevention strategies for GC patients in clinical settings [32].

From 107 key features identified by Lasso regression, we selected 10 most relevant ones to construct four ML models by using both imaging data and clinical features. The result showed that all the four models exhibited strong predictive performance for predicting lower extremity DVT, of which the NB and RF models performed the best, offering an AUC value of 0.928. Although many previous studies have primarily focused on the overall prognosis or post-treatment outcomes in GC patients, few studies have addressed the prediction of lower extremity DVT occurrence in GC patients after surgery. Although VTE-related scores have been commonly used in clinical practice, they are limited in their ability to accurately distinguish the risk of DVT in similar postoperative patients [33]. In this study, we developed ML-based prediction models by utilizing CT imaging features and clinical characteristics of the third lumbar vertebra. The result showed that they could predict the onset of postoperative DVT effectively, offering a more precise prediction independent of traditional VTE scores, thus providing a new valuable tool for clinical use. However, current gaps include unclear biological links between radiomic features (e.g., lmc2) and thrombosis, limited external validation, and the exclusion of dynamic postoperative factors. Addressing these issues will require multicenter studies, mechanistic biomarker research, and the integration of real-time monitoring. In our future work, we plan to develop an even more accurate model for predicting the risk of postoperative DVT across different surgical types, thus facilitating more refined and stratified approaches to DVT prevention in clinical settings. In addition, we envision a future where preoperative CT scans serve dual purposes: guiding surgical planning and predicting complications like DVT. This dual use could maximize resource efficiency and improve patient outcomes, embodying the promise of precision medicine.

Tumors are closely associated with DVT, and GC is recognized as a high-risk tumor for DVT [34]. Surgery further amplifies this risk, making GC patients particularly susceptible to DVT postoperatively [35]. Although the current guidelines emphasize that pharmacological intervention can help prevent DVT, the associated bleeding risk must also be considered [36, 37]. Therefore, more precise stratification of DVT risk and targeted prophylaxis for patients at the highest risk are essential to maximizing clinical benefits [37]. ML has been increasingly utilized in the development of various predictive models and demonstrated strong efficacy in this domain [38, 39]. For instance, Shuai Jin et al. used clinically available variables to build ML models that could effectively predict DVT in cancer patients [39]. Similarly, Bing Xue et al. constructed models using clinical data to identify postoperative DVT risk [40], and Biao Chen et al. applied ML to assess perioperative thrombosis risk in abdominal emergency general surgery [41]. However, most existing models rely heavily on clinical indicators and often overlook the overall postoperative nutritional status of cancer patients. To tackle this issue, we have developed four ML models by utilizing relevant imaging data extracted from CT scans of the third lumbar vertebrae muscles to predict the occurrence of postoperative DVT in GC patients. This approach not only expands the scope of ML application but also holds significant clinical and practical value.

Few studies have investigated the correlation between muscle mass and the incidence of lower extremity DVT following surgery. Takashi Hirase et al. reported that sarcopenia, measured by the psoas muscle index (PMI) could predict the occurrence of DVT after complex thoracolumbar spine repair surgery [42]. Yuta Torii et al. showed that sarcopenia or reduced muscle mass was associated with the resolution of short-term DVT in patients treated with direct oral anticoagulants [43]. Schmeusser et al. found that low skeletal muscle mass was a risk factor for poor survival in patients with non-metastatic renal cell carcinoma with DVT [43]. Although all these studies suggested a correlation between sarcopenia and thrombosis, no study has addressed the relationship between muscle mass and the risk of DVT occurrence in GC patients after surgery. Using a sarcopenia method, we constructed four predictive models by extracting imaging data from the third lumbar vertebra muscles of 106 GC patients who underwent surgery, and categorized them into a DVT group and a non-DVT group. Muscle features were extracted from CT images and integrated with the clinical characteristics of the patients. The medical images were transformed into mineable data, thus providing more detailed insights than traditional clinical indicators alone [44]. Using ML, we developed four novel prediction models, all of which showed potential as effective tools for preventing DVT.

This study has several limitations. First, the sample size is relatively small, and our findings have not been validated with a larger, centrally collected dataset or patient data from multiple centers, which may limit their generalizability. Secondly, a significant number of patients were excluded due to the absence of imaging data, which may potentially produce selection bias. Thirdly, as it is a retrospective study, we were unable to establish a definitive causal relationship between DVT and sarcopenia. Prospective studies with larger cohorts are required to validate this relationship. Finally, our study did not account for other variables related to gastric cancer surgery, such as surgical details, the duration of the procedure, key intraoperative factors (e.g., transfusion data and urine output), or commonly used medications like vasopressors, inotropic drugs, and those prescribed by consultants, knowing that these factors may influence the performance of the models.

Conclusion

In this study, we developed four models by integrating muscle features extracted from CT scans with the clinical features to predict the postoperative incidence of lower extremity DVT in GC patients, achieving an AUC value of 0.928, 0.915, 0.869 and 0.928 respectively. They may prove to be novel non-invasive models for predicting postoperative occurrence of DVT in GC patients, thus providing evidence-based cluse for clinicians to take appropriate measures to prevent the occurrence DVT in such patients.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Author contributions

Yuan Zeng and Yuhao Chen extracted, collected and analyzed data. Fujun Yu designed and supervised the experiment. Dandan Zhu and Jun Xu prepared table. Xiangting Zhang, Huiya Ying, Xian Song, Yixiao Wang and Ruoru Zhou reviewed the results, interpreted data, and wrote the manuscript. All authors have made an intellectual contribution to the manuscript and approved the submission.

Funding

The project was supported by the Zhejiang Provincial Natural Science Foundation of China (No. LY23H030003), the Zhejiang Provincial Medical and Health Science and Technology Plan Project (No. 2024KY1264) and the Wenzhou Municipal Science and Technology Bureau (No. Y20220024).

Data availability

The data that support the findings of this study are available from the corresponding author Fujun Yu, upon reasonable request.

Declarations

Ethics approval and consent to participate

This study was approved by the Ethics Committee of the First Affiliated Hospital of Wenzhou Medical University and conducted in accordance with the ethical standards stipulated in the Declaration of Helsinki. The patient’s written informed consent has been obtained, and their anonymous information will be published in this article.

Competing interests

The authors declare no competing interests.