Prediction Model of Anastomotic Leakage Among Esophageal Cancer Patients After Receiving an Esophagectomy: Machine Learning Approach

Ziran Zhao; Xi Cheng; Xiao Sun; Shanrui Ma; Hao Feng; Liang Zhao

doi:10.2196/27110

Prediction Model of Anastomotic Leakage Among Esophageal Cancer Patients After Receiving an Esophagectomy: Machine Learning Approach

JMIR Med Inform. 2021 Jul 27;9(7):e27110. doi: 10.2196/27110.

Authors

Ziran Zhao¹, Xi Cheng², Xiao Sun³, Shanrui Ma¹, Hao Feng¹, Liang Zhao¹

Affiliations

¹ Thoracic Surgery Department, National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China.
² Department of Global Health Management, School of Public Health and Tropical Medicine, Tulane University, New Orleans, LA, United States.
³ Department of Epidemiology, School of Public Health and Tropical Medicine, Tulane University, New Orleans, LA, United States.

PMID: 34313597
PMCID: PMC8367102
DOI: 10.2196/27110

Abstract

Background: Anastomotic leakage (AL) is one of the severe postoperative adverse events (5%-30%), and it is related to increased medical costs in cancer patients who undergo esophagectomies. Machine learning (ML) methods show good performance at predicting risk for AL. However, AL risk prediction based on ML models among the Chinese population is unavailable.

Objective: This study uses ML techniques to develop and validate a risk prediction model to screen patients with emerging AL risk factors.

Methods: Analyses were performed using medical records from 710 patients who underwent esophagectomies at the National Clinical Research Center for Cancer between January 2010 and May 2015. We randomly split (9:1) the data set into a training data set of 639 patients and a testing data set of 71 patients using a computer algorithm. We assessed multiple classification tools to create a multivariate risk prediction model. Our ML algorithms contained decision tree, random forest, naive Bayes, and logistic regression with least absolute shrinkage and selection operator. The optimal AL prediction model was selected based on model evaluation metrics.

Results: The final risk panel included 36 independent risk features. Of those, 10 features were significantly identified by the logistic model, including aortic calcification (OR 2.77, 95% CI 1.32-5.81), celiac trunk calcification (OR 2.79, 95% CI 1.20-6.48), forced expiratory volume 1% (OR 0.51, 95% CI 0.30-0.89); TLco (OR 0.56, 95% CI 0.27-1.18), peripheral vascular disease (OR 4.97, 95% CI 1.44-17.07), laparoscope (OR 3.92, 95% CI 1.23-12.51), postoperative length of hospital stay (OR 1.17, 95% CI 1.13-1.21), vascular permeability activity (OR 0.46, 95% CI 0.14-1.48), and fat liquefaction of incisions (OR 4.36, 95% CI 1.86-10.21). Logistic regression with least absolute shrinkage and selection operator offered the highest prediction quality with an area under the receiver operator characteristic of 72% in the training data set. The testing model also achieved similar high performance.

Conclusions: Our model offered a prediction of AL with high accuracy, assisting in AL prevention and treatment. A personalized ML prediction model with a purely data-driven selection of features is feasible and effective in predicting AL in patients who underwent esophagectomy.

Keywords: anastomotic leakage; esophageal cancer; esophagectomy; machine learning; risk factors.

©Ziran Zhao, Xi Cheng, Xiao Sun, Shanrui Ma, Hao Feng, Liang Zhao. Originally published in JMIR Medical Informatics (https://medinform.jmir.org), 27.07.2021.