The aim of this research is to develop a machine learning-based model to predict the risk of lower limb deep vein thrombosis in postoperative colorectal cancer patients. A retrospective study was conducted, including 429 colorectal cancer patients who underwent surgical treatment. Data were extracted from the hospital’s electronic medical record system, which included demographic details, medical history, treatment information, disease severity, blood test results, and postoperative complications. The SMOTE algorithm was employed to address the issue of class imbalance. LASSO regression, Xgboost, and random forest were applied for feature selection to identify the features most associated with the risk of lower limb deep vein thrombosis. Following this, a range of ML models, such as LR, RF, GB, MLP, XGB, and KNN, were developed and optimized using the 10-fold cross-validation approach. The performance of these models was assessed through a range of metrics, including accuracy, sensitivity, specificity, positive predictive value, negative predictive value, F1 score, Kappa score, AUC, calibration curve, clinical impact curve, and confusion matrix. To enhance the transparency and interpretability of the model, SHAP and LIME methods were used to explain the prediction results, clarifying the impact of each feature on the predictions and thereby offering useful references for clinicians. Figure 1 illustrates the overall workflow of the proposed system more clearly.

We retrospectively selected 429 colorectal cancer patients who visited the Department of Gastrointestinal Surgery at the First Affiliated Hospital of Southwest Medical University from January 2022 to January 2024. Exclusion criteria include: patients with a history of prolonged bed rest or restricted activity; patients with a history of venous thrombosis; patients with a history of coagulation disorders; patients using drugs affecting coagulation function; patients with malignancies outside the gastrointestinal tract; and patients preoperatively diagnosed with lower extremity deep vein thrombosis. (Exclusion criteria are shown in Figure 1 ). As this study is retrospective, patients are exempt from providing informed consent according to the ethics review board’s policy. The ethics committee has encrypted all personal information of patients involved in this study to prevent any leaks.

The study includes 44 variables related to demographic factors (gender, age), medical history (history of diabetes, hypertension, coronary artery disease, chronic obstructive pulmonary disease), physical characteristics (BMI), disease severity (clinical stage, histological grade, presence of cancer embolus, nerve invasion, vascular invasion), treatment information (surgical method, surgery duration, use of specific cancer treatments), laboratory values (white blood cell count, neutrophil count, lymphocyte count, monocyte count, NLR, hemoglobin, prealbumin, albumin, creatinine clearance, platelet count, prothrombin time PT, fibrinogen, thrombin time TT, D-dimer), and postoperative complications (postoperative high fever, anastomotic leak). Venous blood samples were collected within 24 hours of admission.

Patients were tested within 14 days postoperatively according to the diagnostic criteria for lower limb deep vein thrombosis. Specifically, color Doppler ultrasound showed an uneven echo solid mass in the lower limb, reduced or absent color blood flow and spectral signals, non-collapse of the venous lumen after compression, and venous incompressibility (9).

The structured database initially included 44 clinical variables. First, clinical variables with more than 30% missing data (n = 2) were excluded. The missing data were handled using 10-fold crossvalidation combined with the KNN imputation method. Subsequently, to prevent bias during later model training and improve interpretability, the Variance Inflation Factor (VIF) was employed to examine multicollinearity among the chosen features, ensuring all features’ VIF values were less than 10. Additionally, we also removed variables with nearly zero variance to simplify the model and enhance its robustness. In the end, 39 clinical features of patients were chosen to construct the predictive model. The SMOTE algorithm was used to address the class imbalance issue, balancing the dataset and avoiding bias. Subsequently, patient data were randomly divided into two datasets: (1) a training dataset (70%) for feature selection and model training, and (2) a testing dataset (30%) for model performance evaluation.

For predicting postoperative DVT occurrence in colorectal cancer patients, features were selected using training group samples through three machine learning models: LASSO regression, random forest, and XGboost. The results showed that 29, 15, and 15 feature vectors were selected in the three models, Ultimately, we selected 8 common feature variables from the three models: age, preoperative prealbumin, preoperative albumin, preoperative hemoglobin, CEA, PIKVA2, surgery time, and preoperative white blood cell count.

The machine learning task is to predict the probability distribution of patients developing lower extremity deep vein thrombosis based on these clinical variables. Model development involves experimenting with six machine learning algorithms: Logistic Regression (LR), Multilayer Perceptron (MLP), Extreme Gradient Boosting (XGBoost), Gaussian Process (GP), Random Forest (RF), and Naive Bayes (NB). During the training phase, we employed the 10-fold cross-validation method to train the models in order to achieve optimal predictive performance. To evaluate the predictive performance of each model, we primarily measured the Receiver Operating Characteristic (ROC) curve. In addition, we calculated sensitivity, specificity, accuracy, false positive (FP) rate, positive predictive value (PPV), negative predictive value (NPV), Brier score, F1 score, Decision Curve Analysis (DCA) curve, calibration curve, and Clinical Impact Curve (CIC) for a comprehensive assessment of the model’s performance.

All data analyses in this study were carried out using SPSS (27.0) and R language (version 4.3.3). Preliminary analysis of the dataset used descriptive statistics. Data points that followed a normal distribution were represented by mean ± standard deviation, whereas data points deviating from a normal distribution were shown as median (interquartile range). Subsequently, an independent samples t-test was employed to compare two groups of normally distributed data. In contrast, the Mann-Whitney U test was used for comparing two groups of non-normally distributed data. We resolved the sample imbalance problem by oversampling the minority classes using the SMOTE function from the DMwR2 package in R. To build the predictive model, the dataset was randomly split into a training subset comprising 70% of the total data and a testing subset making up 30% of the total data. Subsequently, various machine learning methods were executed using R, including logistic regression (glm package), Gaussian model (e1071 package), random forest (randomForest package), XGBoost (XGBoost package), feedforward neural network (nnet package), and naive Bayes model (e1071 package). Models were trained using the training subset data with these six ML algorithms. During the model training, a 10-fold cross-validation method was adopted to optimize the model parameters, aiming to prevent overfitting. Statistical significance was defined at the level of P<0.05.

We used the Shapley Additive Explanations (SHAP) algorithm and the Local Interpretable ModelAgnostic Explanations (LIME) algorithm to interpret the main feature contributions after machine learning model training. In particular, the SHAP algorithm assesses the average contribution of each feature value by computing its Shapley value within all possible combinations of features. By taking the weighted average of each feature value’s Shapley value, we can assess the impact of that feature on the overall prediction. Meanwhile, the LIME algorithm analyzes the model from a local perspective to explain the feature importance of specific predictions, providing an additional layer of interpretation and transparency. The combination of these two methods provides us with a multidimensional understanding of model interpretability.

This study encompassed 429 colorectal cancer patients who underwent surgical treatment. The median age of the patients was 67 years (range: 16-91), with 258 males (60.24%) and 171 females (39.76%). The original data from 429 cases includes 267 cases without lower extremity deep vein thrombosis (62.23%) and 162 cases with lower extremity deep vein thrombosis (37.77%). The baseline characteristics comparison of the two patient groups in the original data reveals that age, preoperative white blood cell count, preoperative lymphocyte count, preoperative hemoglobin, preoperative albumin, preoperative prealbumin count, preoperative glomerular filtration rate, gender, preoperative acute complete intestinal obstruction, and surgical method are all statistically significant (refer to Table 1 ).

A total of 1134 patients with colorectal cancer receiving surgical treatment were involved after data imbalance. Patients were split into a training group with 796 cases and a test group with 338 cases in a 7:3 ratio. LASSO regression, as a shrinkage estimation method, achieves variable selection and complexity adjustment by formulating an optimization objective function with a penalty term. This study utilized LASSO regression to identify features including age, surgical procedure, acute intestinal obstruction, nerve invasion, preoperative lymphocyte count, preoperative fibrinogen, preoperative prothrombin time, coronary artery disease, and diabetes ( Figure 2A ). Random forest builds multiple decision trees through the random selection of data subsets and features. Each feature’s importance score reflects its contribution to the model’s predictions, allowing the extraction of the most predictive features and the identification of characteristic factors. Features including age, preoperative prealbumin, preoperative albumin, preoperative hemoglobin, CA724, CEA, and CA242 were selected ( Figure 2B ). Xgboost improves prediction performance by constructing multiple weak learners and using an additive model approach. The importance of features is assessed by calculating gain, coverage, and frequency for each one, identifying factors like age, preoperative prealbumin, preoperative white blood cell count, preoperative hemoglobin, preoperative glomerular filtration rate, BMI, and preoperative prothrombin time ( Figure 2C ). By comparing the selection results of LASSO regression, Xgboost algorithm, and random forest algorithm, we identified the common subset of features selected by these three methods. These selected features were eventually used to construct the model, including age, preoperative prealbumin, preoperative albumin, preoperative hemoglobin, operation time, PIKVA2, CEA, and preoperative neutrophil count ( Figure 2D ).

In the training dataset, the RF model demonstrated excellent predictive performance with an AUC of 1.00, indicating very high prediction accuracy. In comparison, the AUC values for the remaining five models are as follows: XGB’s AUC is 0.996 (95% CI [0.994, 0.999]), GP’s AUC is 0.950 (95% CI [0.935, 0.966]), MLP’s AUC is 0.938 (95% CI [0.918, 0.958]), NB’s AUC is 0.882 (95% CI [0.859, 0.905]), and LR’s AUC is 0.814 (95% CI [0.785, 0.844]) ( Figure 3A ). The F1 scores of these models are as follows: RF 1.0, XGB 0.976, GP 0.878, MLP 0.889, NB 0.740, LR 0.720. In the testing dataset, the AUC values for XGB, GP, MLP, NB, LR, and RF are 0.936 (95% CI [0.907, 0.966]), 0.919 (95% CI [0.890, 0.949]), 0.884 (95% CI [0.843, 0.925]), 0.826 (95% CI [0.781, 0.871]), 0.806 (95% CI [0.760, 0.853]), and 0.973 (95% CI [0.959, 0.986]), respectively ( Figure 3B ). The F1 scores for XGB, GP, MLP, NB, LR, and RF are respectively 0.853, 0.816, 0.825, 0.693, 0.696, and 0.881. In this research, the accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and kappa value of each model were computed and compared ( Figures 3C, D ). The RF model performed excellently in the training dataset. Due to concerns about potential overfitting, the XGB model was ultimately selected as the optimal model.

In our study, we evaluated the predictive accuracy and calibration of the model by analyzing calibration curves for the training and test sets. The calibration curve results showed that the model in the training set had high predictive accuracy, with a Somers’ D coefficient of 0.992 and an area under the ROC curve of 0.996, indicating good discriminatory power ( Figure 4A ). Additionally, the regression calibration slope of the training set model is 0.9934, close to the ideal value of 1.000, and the intercept is -0.0175, demonstrating excellent calibration ability. The Brier score is 0.038, reflecting the high reliability of the model’s predictions. In contrast, the model’s discriminatory power in the test set decreased but still maintained a high level, with an area under the ROC curve of 0.936 and a Somers’ D coefficient of 0.873 ( Figure 4B ). Decision curves for the training set ( Figure 4C ) indicate that the model’s net benefit is significantly above the baseline strategy. On the test set ( Figure 4D ), the model likewise exhibits good net benefit, particularly in the threshold probability range of 0.1 to 0.95, where it maintains a high level of net benefit. The confusion matrix results show the performance differences of the model across different datasets. In the training set ( Figure 4E ), the model correctly identified 440 true negatives and 327 true positives, with 10 false positives and 19 false negatives, the true positive rate is 85.0%, and the true negative rate is 89.7%.In the test set ( Figure 4F ), the model correctly identified 119 true negatives and 178 true positives, misidentifying 20 false positives and 21 false negatives, with a true positive rate of 85.0% and a true negative rate of 89.7%. During the model development process, we considered applying a penalty to the confusion matrix to reduce Type II errors (false negatives). Specifically, we explored methods such as adjusting the classification threshold and using weighted loss functions to impose a higher penalty on false negatives during model training. However, after several experiments, we found that while these adjustments could reduce false negatives, they also led to an increase in false positives, which in turn affected the overall performance metrics of the model (such as AUC and accuracy). Therefore, we ultimately decided not to apply such penalties to maintain the overall balanced performance of the model. Finally, we plotted Clinical Impact Curves (CICs) to evaluate the net benefit of the model with the highest diagnostic value in terms of clinical utility and applicability. Clinical Impact Curves ( Figures 4G, H ) offer insights into the model’s capability to predict high-risk patients at various cost-benefit ratio thresholds. The test set’s curve indicates that when prediction score probabilities exceed 65%, the model’s predictions for postoperative colorectal cancer patients align closely with those who actually develop lower extremity deep vein thrombosis, confirming the model’s high clinical efficacy.

This study evaluated the relative importance of various factors affecting the susceptibility of colorectal cancer patients to developing lower extremity deep vein thrombosis post-surgery. Figure 5A visually represents this ranking, with each point indicating a sample and the color gradient from purple to yellow indicates the magnitude of sample feature values. The vertical axis shows the importance ranking of features alongside the correlation and distribution of feature values with SHAP values. Figure 5B illustrates the hierarchical significance of features in the XGB model. The vertical axis shows individual features ranked in descending order of importance, and the horizontal axis represents the average SHAP values. The analysis shows that age, preoperative albumin, preoperative white blood cell count, surgery duration, and preoperative hemoglobin are the top five ranked features in terms of importance, indicating their critical impact on the occurrence of DVT. To better understand the model’s decision-making process at the individual level, we performed detailed interpretability analyses using LIME on two representative samples(As illustrated in Figures 5C, D ). Through model visualization, we can discern the impact of each feature on the model predictions for these specific instances.