FPT-Former: A Flexible Parallel Transformer of Recognizing Depression by Using Audiovisual Expert-Knowledge-Based Multimodal Measures

Table 3

The performance of depression recognition on E-DAIC databases.

Study	Model name	Modality	RMSE	MAE
Al Hanai et al. 2018 [43]	Long-short term memory (LSTM) neural network	Audio and text features	6.50	5.13
Zhang et al. 2020 [44]	An autoencoder model based on a bidirectional gated recurrent unit (BiGRU)	Speech signals	5.68	4.64
Yang et al. 2020 [45]	Deep convolutional generative adversarial network (DCGAN)	Speech, text, and face data	5.52	4.63
Han et al. 2023 [46]	Spatial-temporal feature network (STFN)	Speech data	6.29	5.38
Fang et al. 2023 [47]	A multimodal fusion model with a multilevel attention mechanism (MFM-Att)	Audiovisual and text data	5.17	—
Ours	A flexible parallel transformer model (FPT-Former)	Audiovisual expert-knowledge-based measures	4.80	4.58

The bold font indicates the lowest value among the compared studies.