Abstract
BACKGROUND: Postoperative bleeding is a serious complication following abdominal tumor surgery, but it is often not clearly diagnosed and documented in clinical practice in China. Previous studies have relied on manual interpretation of medical records to determine the presence of postoperative bleeding in patients, which is time-consuming and laborious. More critically, this manual approach severely hinders the efficient analysis of large volumes of medical data, impeding in-depth research into the incidence patterns and risk factors of postoperative bleeding. It remains unclear whether machine learning can play a role in processing large volumes of medical text to identify postoperative bleeding effectively. OBJECTIVE: This study aimed to develop a machine learning model tool for identifying postoperative patients with major bleeding based on the electronic medical record system. METHODS: This study used data from the available information in the National Health and Medical Big Data (Eastern) Center in Jiangsu Province of China. We randomly selected the medical records of 2,000 patients who underwent in-hospital tumor resection surgery between January 2018 and December 2021 from the database. Physicians manually classified each note as present or absent for a major bleeding event during the postoperative hospital stay. Feature engineering involved bleeding expressions, high-frequency related expressions, and quantitative logical judgment, resulting in 270 features. Logistic regression (LR), K-nearest neighbor (KNN), and convolutional neural network (CNN) models were developed and trained using the 1600-note training set. The main outcomes were accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) for each model. RESULTS: Major bleeding was present in 4.31% (69/1600) of the training set and 4.75% (19/400) of the test set. In the test set, the LR method achieved an accuracy of 0.8275, a sensitivity of 0.8947, a specificity of 0.8241, a PPV of 0.2024, an NPV of 0.9937, and an F1-score of 0.3301. The CNN method demonstrated an accuracy of 0.8900, sensitivity of 0.8421, specificity of 0.8924, PPV of 0.2807, NPV of 0.9913, and an F1-score of 0.4211. While the KNN method showed a high specificity of 0.9948 and an accuracy of 0.9575 in the test set, its sensitivity was notably low at 0.2105. The C-statistic for the LR method was 0.9018 and for the CNN method was 0.8830. CONCLUSIONS: Both the LR and CNN methods demonstrate good performance in identifying major bleeding in patients with postoperative malignant tumors from electronic medical records, exhibiting high sensitivity and specificity. Given the higher sensitivity of the LR method (89.47%) and the higher specificity of the CNN method (89.24%) in the test set, both models hold promise for practical application, depending on specific clinical priorities.