Tonsillectomy is one of the most common surgical procedures in otolaryngology-head and neck. Over 600,000 tonsil-related treatments are reportedly carried out annually across the globe (1). However, postoperative complications remain the greatest challenge for patient recovery, occurring in about 5%–15% of cases, the most common of which is related to post-tonsillectomy hemorrhage (PTH) (1, 2). The incidence of PTH is approximately 1%–19%. PTH is regarded as a primary cause of death related with tonsillectomy and has been suggested to be a potentially lethal complication, particularly in children (3, 4). Anemia and hemodynamic abnormalities, which can hinder the patient’s recovery and even necessitate repeat surgery, are possible PTH hazards that could have a negative psychological impact on the patient as well as physical harm. Sarny et al. found that the reoperation rate after tonsillectomy was 4.6 percent in a multicenter cohort of 9,405 adult and pediatric patients in Austria, with PTH being the most significant complication (5, 6). Clinical doctors should conduct studies and be concerned with determining how to find the risk factors and reduce PTH.

Previous investigations have recorded the characteristics and the prevalence of PTH (6). Age, gender, surgical indication, surgical method, perioperative drugs (including NSAIDs and glucocorticoids), and seasonal fluctuations are only a few of the many factors associated with PTH (7–10). The causes of PTH risk, however, are currently very highly contentious. Therefore, accurate risk prediction for PTH and excellent identification of high-risk individuals will give clinicians the required direction for prevention and treatment (11).

In recent years, artificial intelligence has been widely used to explore early warning predictions for many disease complications. Machine learning models and advanced statistical tools can be used to predict the outcome of complex datasets based on iterative learning, thereby making these models more accurate and stable through the selection of variable features. Therefore, more researchers advocate the use of new predictive models based on machine learning to support personalized treatment for patients (11–13). To the best of our knowledge, there is no predictive model for PTH based on machine learning.

This study used machine learning method to develop and validate prediction models, to provide individualized management for the postoperative prognosis of patients.

Patient studies were conducted according to the guiding principles of the Helsinki Declaration and were approved by the Ethics Committee of the 940th Hospital of Joint Logistics Support Force of Chinese People’s Liberation Army (2022KYLL165).

According to previous studies, the risk factors associated with PTH were more controversial. At this stage, despite the understanding of PTH and the implementation of the concept of post-tonsillectomy management in many centers, there are still some unknowns and regional differences that cause hemorrhage. Due to the increasing need for cost-effectiveness in healthcare systems, patterns of postoperative care and clinical decisions making also vary greatly in different countries and institutions. To the authors' knowledge, there has been little development of PTH prediction models. Accurate prediction of prognosis is crucial for patient-centered care, whether to guide inclusive decision-making or select treatment strategies. The application of machine learning to medical and clinical conditions forms a major emerging research trend (18). In this study, predictive risk factors for PTH were identified using the mutual information method and risk prediction models for PTH were created. The performance of five machine learning models was compared. The results showed that the XGBoost model had the highest AUC, positive predictive value, and negative predictive value. The SHAP method further explained the predictors and model prediction performance. It also provided a simple and robust method for personalized prediction of PTH, which can provide important information for medical decision support.

For an overview of the XGBoost logic, this non-linear machine learning model used the input (independent) variables in the training dataset to construct an array of decision trees in every possible combination to establish a series of thresholds that split variables to maximize the information gain (19). In our study, the XGBoost model predicted patient risk with excellent calibration and good validation. Previous reports have described machine learning techniques as “black boxes” that provide little information on how predictions are made. This reduces clinician acceptance because they are reluctant to make medical diagnoses based on opaque decisions (20). We used the SHAP method to explain the decision-making process of the XGBoost model. In clinical practice, visualization of SHAP can be translated into clinical insight. It can show the accuracy and complexity of the prediction model better and provide some reference for postoperative patient care and personalized decision-making treatment (14).

Among the several important factors explained by SHAP, age has been confirmed to be a risk factor for PTH. Compared with children, adult patients have a higher risk of PTH. Some studies suggest that the postoperative bleeding rate of adults is as high as 8.6%–14.5%, while that of children is about 2.1%-5% (4, 20). A reasonable explanation is that adults with long-lasting chronic tonsillitis are prone to fibrosis of the tonsillar bed and that dissection of the tonsils can make surgery more difficult (21). The results of a previous retrospective study involving 897 patients showed that INR ≥ 1.2 as a risk factor helped to identify patients at higher risk of hemorrhage (4). INR was also identified as an important predictor in our study. This is consistent with previous studies. In addition, many past studies have shown a correlation between season and increased risk of PTH (22). It has been suggested that higher altitude and lower atmospheric moisture enhance the incidence of mucosal surface hemorrhage (23). In this study, the bleeding rate of patients increased in the spring. Our hospital is located in a high-altitude area in northwest China, where spring is relatively dry. It is regrettable that more detailed atmospheric data information was not further obtained in this study. However, it is necessary for the season to be used as an important factor to predict the PTH. In addition, some studies suggested that obesity, smoking, blood type, and perioperative hormone use were all important factors affecting the incidence of PTH (24–27). It had even been suggested that patients undergoing tonsillectomy must quit smoking to avoid bleeding after tonsillectomy. It is worth mentioning that in this study, we believed that the wound of the tonsillar fossa after surgery played a role. Our study found an increased rate of postoperative hemorrhage in patients with thicker and yellowing tonsillar fossa wounds after tonsillectomy. Few studies have investigated trauma as a potential factor. Therefore, we believe that attention should be paid to patients with yellowing of the tonsillar fossa and thickened white membranes after tonsillectomy.

This study has several strengths. Firstly, our study created and validated the first PTH risk prediction model by machine learning methods. Second, we included 10 risk factors, including surgical wound, and season, which were easy to be ignored. Third, our model can be used by clinicians and nurses as an intuitive way to predict PTH risk for the appropriate management of high-risk patients. Finally, this model can provide a new idea for future PTH research.

Our study has some limitations. The sample size in this study was small. Although we performed internal validation, external validation is needed to obtain more clinical evidence. The machine learning model in this study used the presence of combined secretory otitis media and adenoidectomy as predictor variables, but the number of adults with these diseases was small, so there was some confounding bias, and further studies are still needed to develop prediction models for different age groups. In addition, some scholars believe that primary bleeding is usually attributed to the type of surgical technique, and primary and secondary bleeding needs to be carefully distinguished and studied separately (28). The coblation technique is used for all tonsillectomies in our hospital. Therefore, the study focuses on the secondary hemorrhage. Furthermore, although the SHAP model can determine the influence of patient characteristics on prediction, this is only a mapping relationship between predictor variables and prediction outcomes, rather than a causal relationship (29). Despite these limitations, we developed the XGBoost model as a tool that has the potential to facilitate individualized treatment strategies for patients.

In conclusion, we identified baseline demographic and clinical variables as predictive risk factors for PTH and used machine learning to construct a PTH risk prediction model for patients after tonsillectomy. The XGBoost model could efficiently improve the predictive accuracy of PTH. With increasing use in medicine, machine learning models could be considered for inclusion in clinical pathway decisions for clinical use and further validation. Future research will explore additional machine learning models applicable to predicting PTH risk.