Sleep apnea ( SA) affects the quality of sleep and increases the risk of cerebrovascular and cardiovascular diseases. It is advantageous to implement the accurate classification for the timely treatment at the early stage of SA. In this paper, one novel SA classification method utilizing heterogeneous ensemble learning and heterogeneous feature fusion is proposed. Firstly, the SE-ResNet is used as primary classifier of the extracted wavelet time-frequency spectrum from raw electrocardiogram (ECG). Then the 1D CNN-LSTM is used as primary classifier of the extracted R-peak to R-peak interval(RRI) sequence and R-peak amplitude (RAMP) sequence. And the SVM is used as primary classifier of extracted heart rate variability features. Finally, the stacking method is adopted as fusion strategy for heterogeneous ensemble learning, and then another SVM is used as the secondary classifier to implement SA classification. The proposed SA classification method is evaluated on Apnea-ECG dataset, whose accuracy is 89. 12% . Experimental results show that the proposed method utilizes the diversity of primary classifiers and complementarity of heterogeneous features efficiently, which outperforms the conventional SA classification method.