Sleep apnea syndrome (SAS), which can lead to a range of Cardiopulmonary diseases, is a common chronic sleep disorder. The unobtrusive detection based on wearable devices is helpful for early diagnosis and treatment of SAS. To this end, this paper presents a method based on a one-dimensional multi-scale bidirectional temporal convolutional neural network (1D-MsBiTCNet) and two model performance optimization techniques, i.e., regularized dropout (RD) and logit adjustment (LA). Among them, 1D-MsBiTCNet has outstanding capabilities in both feature extraction and temporal dependence representation. RD and LA play an effective role in solving the overfitting problem of model training and the class imbalance problem of the dataset, respectively. The proposed model was trained and tested on a photoplethysmography (PPG) dataset (including data from 92 subjects) collected from commercial wearable bracelets. On this dataset, our method achieved accuracy, sensitivity and specificity of 82.76%, 71.58%, 86.74% for per-segment detection, and 97.83%, 88.89%, 100.00% for per-recording severe SAS detection. For the precise quantification of apnea-hypopnea index (AHI), our method achieved a mean absolute error of 5.44 between the predicted AHI and the ground truth AHI. The experimental results show that our proposed method has an outstanding performance and can provide a methodological reference for large-scale SAS automatic detection.