Deep learning‐based diagnostic model for predicting complications after gastrectomy

BACKGROUND: Gastric cancer is one of the leading causes of cancer deaths, and gastrectomy with lymph node dissection is the mainstay of treatment. Despite clinician efforts and advances in surgical methods, the incidence of complications after gastrectomy remains 10%-20% including fatalities. To the best of our knowledge, this is the first report on utilization of a deep learning method to build a new artificial intelligence model that could help surgeons diagnose these complications. METHODS: A neural network was constructed with a total of 4000 variables. Clinical, surgical, and pathological data of patients who underwent radical gastrectomy at our institute were collected to maintain a deep learning model. We optimized the parameters of the neural network to diagnose whether these patients would develop complications after gastrectomy or not. RESULTS: Seventy percent of the data was used to optimize the neural network parameters, and the rest was used to validate the model. A model that maximized the receiver operating characteristics (ROC) area under the curve (AUC) for validation of the data was extracted. The ROC-AUC, sensitivity, and specificity of the model to diagnose all complications were 0.8 vs 0.7, 81% vs 50%, and 69% vs 75%, for the teaching and validation data, respectively. CONCLUSIONS: A predictive model for postoperative complications after radical gastrectomy was successfully constructed using the deep learning method. This model can help surgeons accurately predict the incidence of complications on postoperative day 3.