The advancement of AI technologies within the medical domain has spurred significant interest in enhancing the quality and efficiency of healthcare services. However, current methodologies for topic information extraction from extensive medical review datasets present several challenges. These include topic overlap and redundancy during the topic clustering phase, the inability of models to dynamically adjust feature weights, and difficulties in simultaneously capturing linear and non-linear characteristics during prediction. These limitations hinder the comprehensive capture of topic popularity trends, thereby affecting predictive accuracy. To address these issues, this paper introduces an innovative solution, AIMedTop, which integrates an enhanced BERTopic model with a hybrid time series model. Initially, review data from multiple medical platforms were collected and preprocessed. Subsequently, an improved BERTopic model was employed for topic modeling of both physician and patient datasets, extracting topic distributions specific to each group. The topic popularity formula was then constructed, incorporating features such as like counts, comment volumes, collection numbers, and reading metrics, with feature weights dynamically optimized using the PPO algorithm. Finally, a combined Multi-LSTM and ARIMA model was utilized, employing a 7-year training and a 3-year testing dataset to forecast popularity trends over the subsequent four years. Experimental results demonstrate that the enhanced BERTopic model mitigates topic overlap and redundancy, yielding clearer, more precise, and coherent topic and keyword distributions. The integration of Multi-LSTM and ARIMA significantly improves prediction accuracy and stability compared to other predictive models, offering a novel approach and technical framework for topic popularity forecasting.