During the initial stages, atrial fibrillation (AF) typically presents as paroxysmal atrial fibrillation (PAF), which may further progress into persistent atrial fibrillation, leading to high-risk diseases such as ischemic stroke and heart failure. Given that the current machine learning algorithms used for predicting AF involve time-consuming and labor-intensive processes of feature extraction and labeling electrocardiogram data, this study proposes a novel two-stage semi-supervised AF attack prediction algorithm. The first stage is designed as unsupervised learning based on convolutional autoencoder (CAE) network when inputting RR interval time series signal, while the second stage is designed as supervised learning using a Long Short-Term Memory (LSTM) model. A training set consisting of 20 segments of PAF and 20 normal heart rates was used to evaluate the performance of the CAE-LSTM combination model. The results showed that the average accuracy and root mean square error of ten-fold cross-validation were 93.56% and 0.004, respectively, with an F1 parameter of 0.9345. In summary, the preliminary results suggest that the combination of unsupervised CAE model and supervised LSTM model can reduce the dimensionality of the input data while using a small amount of labeled data as input for subsequent classification. Furthermore, the proposed algorithm can be used for predicting atrial fibrillation when the sample size is limited.Clinical Relevance- Compared with common supervised methods, our proposed method only requires a small number of tagged ECG signals, which can reduce the workload of clinicians to complete the task of atrial fibrillation attack prediction.