Early prediction of dementia, a long-term progressive disease, has always been a challenge. In recent years, advances in artificial intelligence have led to new computer aided diagnostic tools. However, these methods often offer limited interpretability due to their simplistic binary outputs and black-box algorithms, restricting their use. In this work, we addressed aforementioned shortcomings by assigning clinically meaningful categories to a longitudinal cohort dataset and using an interpretable Random Forest algorithm to train the prediction model. Our results show that the model predicts various categories effectively. We further applied an advanced machine learning explanation framework to analyse the predictions, revealing the impact of some key risk factors on the prediction and varying interaction patterns between these factors when predicting different development stages of dementia.