Research on hospital patient classification and disease prediction based on data mining

This study focuses on hospital patient classification and disease prediction based on data mining, aiming at building an efficient and interpretable prediction model to solve key problems in medical decision-making. Firstly, multi-modal data fusion is realized by constructing heterogeneous data mapping model and using tensor decomposition method. Then, the dynamic feature selection algorithm is developed, the mutual information screening mechanism constrained by medical knowledge map is introduced, and the feature space dimension is constrained by LASSO regression. On this basis, a dual-channel model integrating static classification and dynamic prediction is constructed, in which the static classification channel is based on the improved XGBoost classification model, and the dynamic prediction channel adopts the Long Short-Term Memory (LSTM) and introduces the attention mechanism of medical events. In addition, a multi-level interpretability enhancement mechanism is introduced, including the attribution analysis with SHAP value and the development of decision tree proxy model to transform complex model predictions into intuitive and understandable rules. The incremental learning strategy based on sliding window is also constructed, and the Page-Hinkley statistical test method is introduced to monitor the real-time concept drift to ensure the dynamic adaptability of the model. The experiment is based on the cardiovascular disease data set of a 3A hospital. The results show that the multi-modal data fusion significantly improves the prediction performance, and the full-modal fusion method achieves the sensitivity of 0.854 AUC-ROC and 81.3% in the test. The performance of the two-channel prediction model is better than that of the single model, with AUC of 0.906, The importance analysis of SHAP features reveals the influence degree of each feature in the model on the prediction results, and the mapping of clinical rules further verifies the consistency between the decision logic of the model and clinical practice. Dynamic incremental updating method performs better in model degradation period, updating efficiency and resource consumption. The results of clinical utility verification show that doctors give positive comments on all aspects of the system.