LASSO-Based Machine Learning Algorithm for Prediction of Lymph Node Metastasis in T1 Colorectal Cancer

Abstract

Purpose: The role of tumor-infiltrating lymphocytes (TILs) in predicting lymph node metastasis (LNM) in patients with T1 colorectal cancer (CRC) remains unclear. Furthermore, clinical utility of a machine learning-based approach has not been widely studied.

Materials and methods: Immunohistochemistry for TILs against CD3, CD8, and forkhead box P3 in both center and invasive margin of the tumor were performed using surgically resected T1 CRC slides. Three hundred and sixteen patients were enrolled and categorized into training (n=221) and validation (n=95) sets via random sampling. Using clinicopathologic variables including TILs, the least absolute shrinkage and selection operator (LASSO) regression model was applied for variable selection and predictive signature building in the training set. The predictive accuracy of our model and the Japanese criteria were compared using area under the receiver operating characteristic (AUROC), net reclassification improvement (NRI)/integrated discrimination improvement (IDI), and decision curve analysis (DCA) in the validation set.

Results: LNM was detected in 29 (13.1%) and 12 (12.6%) patients in training and validation sets, respectively. Nine variables were selected and used to generate the LASSO model. Its performance was similar in training and validation sets (AUROC, 0.795 vs. 0.765; p=0.747). In the validation set, the LASSO model showed better outcomes in predicting LNM than Japanese criteria, as measured by AUROC (0.765 vs. 0.518, p=0.003) and NRI (0.447, p=0.039)/IDI (0.121, p=0.034). DCA showed positive net benefits in using our model.

Conclusion: Our LASSO model incorporating histopathologic parameters and TILs showed superior performance compared to conventional Japanese criteria in predicting LNM in patients with T1 CRC.

Keywords: LASSO; Lymph node; Machine learning; T1 colorectal cancer; Tumor-infiltrating lymphocytes.

Fig. 1

Selection of significant parameters in clinicopathologic variables in the training set and definition of linear predictor. (A) Ten time cross-validation for tuning parameter selection in the LASSO model. (B) LASSO coefficient profiles. The LASSO was used for regression of high dimensional predictors. The method uses an L1 penalty to shrink some regression coefficients to exactly zero. The binomial deviance curve was plotted versus log (λ), where λ is the tuning parameter (A). LASSO coefficient profiles of clinicopathologic variables (B). LASSO, least absolute shrinkage and selection operator.

Fig. 2

Comparison of AUROC between LASSO model in the training and validation sets and Japanese criteria in the validation set. AUC, area under the curve; AUROC, area under the receiver operating characteristic; CI, confidence interval; LASSO, least absolute shrinkage and selection operator.

Fig. 3

Decision curve analysis of Japanese criteria and LASSO model in the training (A) and validation (B) set. The y-axis measures the net benefit. The green line represents the LASSO model. The red line represents the Japanese criteria. The gray line represents the assumption that all patients underwent surgeries. The black line represents the assumption that patients underwent no surgeries. The net benefit was calculated by subtracting the proportion of all patients who are false positive from the proportion who are true positive, weighting by the relative harm of forgoing treatment compared with the negative consequences of an unnecessary treatment. The decision curve showed that if the threshold probability of a patient or doctor is >10%, using the LASSO model in the current study to predict LNM adds more benefit than the treat-all-patients scheme or the treat-none scheme. For example, if the personal threshold probability of a patient is 20% (i.e., the patient would opt for surgery if his/her probability of LNM was > 20%), then the net benefit is 0.35 when using the LASSO model to make the decision of whether to undergo surgery, with added benefit than the treat-all scheme or the treat-none scheme. This decision curve analysis showed that the net benefit was comparable on the basis of the Japanese criteria and the treat-all or treat-none strategies. LASSO, least absolute shrinkage and selection operator; LNM, lymph node metastasis.