Alzheimer’s disease is considered the epidemic of the twentieth century, particularly the radical behavior produced by the patients, which can easily lead to an increase in mortality. This paper focuses on optimizing the search path algorithm to forecast the likelihood of aggressive behavior, enabling the implementation of targeted preventive measures. Additionally, it performs preprocessing procedures like data cleansing on the health data of the elderly. We propose a feature extraction optimization model to calculate the feature contribution of elderly radical behavior data, filter the feature data associated with radical behavior based on this contribution, and establish an interpolation model for association rule learning. Using Logistic Regression, Simple Bayes, and Support Vector Machine classification models, the risk warning model for aggravated behavior is constructed. After one month’s intervention with the risk warning model, the aggressive behavior of the elderly decreased by 58.83%, 43.06%, and 67.94%, respectively, compared to the pre-intervention period, and the intervention effect of the model was good.