Using the functional Near-Infrared Spectroscopy (fNIRS) for prenatal depression recognition can improve accuracy compared to relying solely on rating scales. However, due to the low signal-to-noise ratio (SNR) and limited data in fNIRS, deep learning based classification methods still face challenges. Previous works primarily exhibit two main limitations: they lack manually extracted statistical features to guide temporal models, and they only use one single task fNIRS as classification features. Researchers have found that the emotional responses of individuals with depression differ from those of healthy individuals, and that women become more emotionally sensitive during pregnancy. Based on these observations, we design three different tasks for the collection of fNIRS signals: happy stimulation, sad stimulation, and resting status. Both the temporal and statistical features of fNIRS are utilized as inputs for the model. We propose an Attention-Based-Features-Fusion Emotion-guided network, which integrates distinct features from various tasks to obtain the classification results. The proposed model achieves the best results on a dataset with 27 subjects, verifying its effectiveness. The code has been open-sourced ( https://github.com/SCUT-Xinlab/fNIRS-prenatal-depression ).