AUTHOR=Piao Zheng Hua , Ge Rong , Lu Lu TITLE=An artificial intelligence prediction model outperforms conventional guidelines in predicting lymph node metastasis of T1 colorectal cancer JOURNAL=Frontiers in Oncology VOLUME=13 YEAR=2023 URL=https://www.frontiersin.org/journals/oncology/articles/10.3389/fonc.2023.1229998 DOI=10.3389/fonc.2023.1229998 ISSN=2234-943X ABSTRACT=Background

According to guidelines, a lot of patients with T1 colorectal cancers (CRCs) undergo additional surgery with lymph node dissection after being treated by endoscopic resection (ER) despite the low incidence of lymph node metastasis (LNM).

Aim

The aim of this study was to develop an artificial intelligence (AI) model to more effectively identify T1 CRCs at risk for LNM and reduce the rate of unnecessary additional surgery.

Methods

We retrospectively analyzed 651 patients with T1 CRCs. The patient cohort was randomly divided into a training set (546 patients) and a test set (105 patients) (ratio 5:1), and a classification and regression tree (CART) algorithm was trained on the training set to develop a predictive AI model for LNM. The model used 12 clinicopathological factors to predict positivity or negativity for LNM. To compare the performance of the AI model with the conventional guidelines, the test set was evaluated according to the Japanese Society for Cancer of the Colon and Rectum (JSCCR) and National Comprehensive Cancer Network (NCCN) guidelines. Finally, we tested the performance of the AI model using the test set and compared it with the JSCCR and NCCN guidelines.

Results

The AI model had better predictive performance (AUC=0.960) than the JSCCR (AUC=0.588) and NCCN guidelines (AUC=0.850). The specificity (85.8% vs. 17.5%, p<0.001), balanced accuracy (92.9% vs. 58.7%, p=0.001), and the positive predictive value (36.3% vs. 9.0%, p=0.001) of the AI model were significantly better than those of the JSCCR guidelines and reduced the percentage of the high-risk group for LNM from 83.8% (JSCCR) to 20.9%. The specificity of the AI model was higher than that of the NCCN guidelines (85.8% vs. 82.4%, p=0.557), but there was no significant difference between the two. The sensitivity of the NCCN guidelines was lower than that of our AI model (87.5% vs. 100%, p=0.301), and according to the NCCN guidelines, 1.2% of the 105 test set patients had missed diagnoses.

Conclusion

The AI model has better performance than conventional guidelines for predicting LNM in T1 CRCs and therefore could significantly reduce unnecessary additional surgery.