Machine learning models can predict cancer-associated disseminated intravascular coagulation in critically ill colorectal cancer patients

Abstract

Background: Due to its complex pathogenesis, the assessment of cancer-associated disseminated intravascular coagulation (DIC) is challenging. We aimed to develop a machine learning (ML) model to predict overt DIC in critically ill colorectal cancer (CRC) patients using clinical features and laboratory indicators.

Methods: This retrospective study enrolled consecutive CRC patients admitted to the intensive care unit from January 2018 to December 2023. Four ML algorithms were used to construct predictive models using 5-fold cross-validation. The models' performance in predicting overt DIC and 30-day mortality was evaluated using the area under the receiver operating characteristic curve (ROC-AUC) and Cox regression analysis. The performance of three established scoring systems, ISTH DIC-2001, ISTH DIC-2018, and JAAM DIC, was also assessed for survival prediction and served as benchmarks for model comparison.

Results: A total of 2,766 patients were enrolled, with 699 (25.3%) diagnosed with overt DIC according to ISTH DIC-2001, 1,023 (36.9%) according to ISTH DIC-2018, and 662 (23.9%) according to JAAM DIC. The extreme gradient boosting (XGB) model outperformed others in DIC prediction (ROC-AUC: 0.848; 95% CI: 0.818-0.878; p < 0.01) and mortality prediction (ROC-AUC: 0.708; 95% CI: 0.646-0.768; p < 0.01). The three DIC scores predicted 30-day mortality with ROC-AUCs of 0.658 for ISTH DIC-2001, 0.692 for ISTH DIC-2018, and 0.673 for JAAM DIC.

Conclusion: The results indicate that ML models, particularly the XGB model, can serve as effective tools for predicting overt DIC in critically ill CRC patients. This offers a promising approach to improving clinical decision-making in this high-risk group.

Keywords: anticoagulation; colorectal cancer; disseminated intravascular coagulation; intensive care unit; machine learning model.

FIGURE 1

Schematic of the study design. CRC, colorectal cancer; DIC, disseminated intravascular coagulation; ICU, intensive care unit; ISTH, International Society on Thrombosis and Haemostasis; JAAM, Japanese Association for Acute Medicine; ML, machine learning.

FIGURE 2

Cox regression for the prediction of 30-day mortality in critically ill CRC patients with different DIC underlying mechanisms. ISTH DIC-2018, DIC score using cut-off scores published in 2018. Daily repeated scoring was performed during ICU stays. A patient was annotated as overt DIC if they had a positive DIC score that day. All HRs were adjusted for age and gender. CI, confidence intervals; CRC, colorectal cancer; DIC, disseminated intravascular coagulation; HRs, hazard ratios; ISTH, International Society on Thrombosis and Haemostasis. *P-value of less than 0.05.

FIGURE 3

The performance of the RF Plot (A), SVM Plot (B), XGB Plot (C), and LR Plot (D) models in the training cohort, with DIC diagnosed by ISTH DIC-2018 as the outcome. To evaluate the performance of the four machine learning models, we plotted the classification based on the optimal threshold and ROC and PR curves. AUCs were also calculated with 95% CIs. The optimal threshold points of the PR curves were plotted, along with their respective sensitivities and positive predictive values. CI, confidence interval; DIC, disseminated intravascular coagulation; ISTH, International Society on Thrombosis and Haemostasis; LR, logistic regression; PPV, positive predictive value; PR, precision-recall curve; RF, random forest; ROC-AUC, the area under the receiver operating characteristic curve; SEN, sensitivity; SVM, supporting vector machine; XGB, XGBoost; Youden index: = sensitivity + specificity - 1.

FIGURE 4

The performance of the RF Plot (A), SVM Plot (B), XGB Plot (C), and LR Plot (D) models in the validation cohort, with DIC diagnosed by ISTH DIC-2018 as the outcome. To evaluate the performance of the four machine learning models, we plotted the classification based on the optimal threshold, and the ROC and PR curves. AUCs were also calculated with 95% CIs. The optimal threshold points of the PR curves were plotted, along with their respective sensitivities and positive predictive values. CI, confidence interval; DIC, disseminated intravascular coagulation; ISTH, International Society on Thrombosis and Haemostasis; LR, logistic regression; PPV, positive predictive value; PR, precision-recall curve; RF, random forest; ROC-AUC, the area under the receiver operating characteristic curve; SEN, sensitivity; SVM, supporting vector machine; XGB, XGBoost; Youden index: = sensitivity + specificity - 1.

FIGURE 5

Kaplan-Meier plot of estimated 30-day mortality according to different ML models in training (A, C, E, G) and validation (B, D, F, H) cohort. Any differences in the incidence were evaluated with a log-rank test. Plot A-H was grouped by predictions of RF, SVM, XGB, and LR models on the first day of ICU stay, respectively. CI, confidence interval; DIC, disseminated intravascular coagulation; HR, hazard ratio; ICU, intensive care unit; LR, logistic regression; ML, machine learning; RF, random forest; SVM, supporting vector machine; XGB, XGBoost.

FIGURE 6

Analysis and interpretation of the XGB Model. Plot (A) displays the SHAP analysis results for the model training set. The variables are characterized by their mean absolute SHAP values. Distributions of the top ten ranked variables are displayed across individual patients. Each point in the figure denotes the SHAP value for a particular patient. The y-axis shows the ranking of each variable’s impact on the model prediction. The x-axis displays the SHAP value. Blue indicates lower variable values, while red indicates higher values. Plot (B) illustrates the average contribution of each feature to the model output as determined by the SHAP analysis. INR, international normalized ratio; PLT, platelet counts; PT, prothrombin time; PTT, activated partial thromboplastin time; SHAP, Shapley Additive Explanations; XGB, XGBoost.