Category imbalance is an important challenge in many domains, especially in medical data analysis. Misclassification of a few categories is more costly and may weaken the overall classification efficacy compared to the majority category because conventional machine learning models focus more on learning the characteristics of the majority category, thus ignoring the critical information of the few categories. This study used a machine learning algorithm to develop a new prognostic model specifically targeting the imbalance in sepsis. The model employs an adaptive strategy that balances the number of samples by determining the nearest neighbours of each minority category sample, ensuring that the importance of the minority categories is well represented in the prediction process. At the same time, we optimise the loss function of the XGBoost model and dynamically adjust the weights so that the model focuses more on the difficult-to-classify samples in the prediction process. Experimental results indicate that our model successfully learns the complete information of a few categories while providing high classification accuracy (the AUROC value reached 87.69% and the F1 score was 86.11%). To enhance the transparency of the model, we used the SHAP method to visualise the importance of different features on the prognostic impact of sepsis. Overall, this study provides a novel machine learning algorithm for handling imbalanced medical datasets and accurately predicting sepsis outcomes.