Machine Learning Risk Prediction Model of 90-day Mortality After Gastrectomy for Cancer

Abstract

Objective: To develop and validate a risk prediction model of 90-day mortality (90DM) using machine learning in a large multicenter cohort of patients undergoing gastric cancer resection with curative intent. Background: The 90DM rate after gastrectomy for cancer is a quality of care indicator in surgical oncology. There is a lack of well-validated instruments for personalized prognosis of gastric cancer. Methods: Consecutive patients with gastric adenocarcinoma who underwent potentially curative gastrectomy between 2014 and 2021 registered in the Spanish EURECCA Esophagogastric Cancer Registry database were included. The 90DM for all causes was the study outcome. Preoperative clinical characteristics were tested in four 90DM predictive models: Cross Validated Elastic regularized logistic regression method (cv-Enet), boosting linear regression (glmboost), random forest, and an ensemble model. Performance was evaluated using the area under the curve by 10-fold cross-validation. Results: A total of 3182 and 260 patients from 39 institutions in 6 regions were included in the development and validation cohorts, respectively. The 90DM rate was 5.6% and 6.2%, respectively. The random forest model showed the best discrimination capacity with a validated area under the curve of 0.844 [95% confidence interval (CI): 0.841-0.848] as compared with cv-Enet (0.796, 95% CI: 0.784-0.808), glmboost (0.797, 95% CI: 0.785-0.809), and ensemble model (0.847, 95% CI: 0.836-0.858) in the development cohort. Similar discriminative capacity was observed in the validation cohort. Conclusions: A robust clinical model for predicting the risk of 90DM after surgery of gastric cancer was developed. Its use may aid patients and surgeons in making informed decisions.

Artificial Intelligence–enabled Decision Support in Surgery

“…Twenty-three articles9,19,21,23–25,27,29,30,32–34,36,37,39,41,42,44–46,48–50 (63.9%) reported precision metrics (area under the precision-recall curve, positive predictive value, or F1 score). Twenty-five articles9,20,21,23–28,31,33,34,36,38,40,42–50 (69.4%) included explainability mechanisms to convey the relative importance of input features in determining outputs. Thirteen articles9,16,17,20,25,27,29,30,35,38,45,46,50 (36.1%) presented a framework that could be used for clinical implementation; none of the articles assessed the efficacy of clinical implementation.…”

“…The average AUROC for the best model across all 36 articles was 0.83; of the 8 articles with sample sizes of less than 2000 (ie, less than 1000 samples per class), 7 22,[26][27][28]31,35,37 (87.5%) had below-average (ie, less than 0.83) AUROC or accuracy. Twenty-four 9,18,20,[24][25][26][27][28][29][32][33][34][35]37,[39][40][41][43][44][45][46][47][48][49] articles (66.7%) presented confidence intervals for performance metrics. Among the 8 articles with sample sizes of less than 2000, 5 [26][27][28]35,37 presented confidence intervals.…”

“…Twentythree articles 9,19,21,[23][24][25]27,29,30,[32][33][34]36,37,39,41,42,[44][45][46][48][49][50] (63.9%) reported precision metrics (area under the precision-recall curve, positive predictive value, or F1 score). Twenty-five articles 9,20,21,[23][24][25][26][27][28]31,33,34,36,38,40,[42][43][44][45][46][47][48][49][50] (69.4%) included explainability mechanisms to convey the relative importance of input features in determining outputs. Thirteen articles 9,16,17,…”

Machine learning models in clinical practice for the prediction of postoperative complications after major abdominal surgery

“…Several authors have tried to use ML models to optimize postoperative management by predicting postoperative complications. Unfortunately, many conclude that the implementation of these models is far from being clinically viable, even though most ML models achieve reasonable performance [5][6][7][8][9][10][11][12][13][14][15]. Cao et al [5], Weller et al [12], and Van den Bosch et al [15] concluded that no practical implementation could be achieved for ML models due to the predictive value being too low to clinically implement.…”

AutoScore-Ordinal: an interpretable machine learning framework for generating scoring models for ordinal outcomes

“…Most risk prediction models in the literature were developed using multivariable logistic regression models or ML techniques to predict a binary outcome. Aside from the AutoScore framework, ML applications include the use of Naive Bayes (NB), XGBoost, k-nearest neighbor (K-NN), multilayer perceptron, support vector machine (SVM) and CatBoost for predicting the risk of cardiovascular disease [ 21 ], random forest (RF), XGBoost, logistic regression, SVM and K-NN for the risk of incident diabetic retinopathy among patients with type 2 diabetes mellitus [ 22 ], a stroke risk prediction model using NB, decision tree and RF models [ 23 ], a XGBoost based cerebral infarction risk prediction model [ 24 ], and a developed risk model for 90-day mortality of patients undergoing gastric cancer resection with curative intent using cross validated elastic regularized logistic regression method, boosting linear regression, RF and an ensemble model [ 25 ].…”