Frequency Feature Fusion Graph Network For Depression Diagnosis Via fNIRS

TL;DR

This paper tackles depression diagnosis from fNIRS by introducing a $DFT$-based frequency-domain temporal biomarker and a phased $TGCN$ that processes three task phases. It trains on a large real-world dataset of 1,086 subjects and a propensity-score matched subset, using a Temporal Fusion Module to fuse $OTF$ and top-$k$ $DTF$ features with a Frequency Point Biserial Correlation Attention (FAM) mechanism that weights informative frequency bands. SHAP-based interpretability demonstrates alignment between learned biomarkers and neurophysiological patterns, while ablations show the frequency-domain approach and phase-specific modeling boost metrics over baselines. The result is a scalable, interpretable framework for practical depression diagnosis using portable fNIRS data.

Abstract

Data-driven approaches for depression diagnosis have emerged as a significant research focus in neuromedicine, driven by the development of relevant datasets. Recently, graph neural network (GNN)-based models have gained widespread adoption due to their ability to capture brain channel functional connectivity from both spatial and temporal perspectives. However, their effectiveness is hindered by the absence of a robust temporal biomarker. In this paper, we introduce a novel and effective biomarker for depression diagnosis by leveraging the discrete Fourier transform (DFT) and propose a customized graph network architecture based on Temporal Graph Convolutional Network (TGCN). Our model was trained on a dataset comprising 1,086 subjects, which is over 10 times larger than previous datasets in the field of depression diagnosis. Furthermore, to align with medical requirements, we performed propensity score matching (PSM) to create a refined subset, referred to as the PSM dataset. Experimental results demonstrate that incorporating our newly designed biomarker enhances the representation of temporal characteristics in brain channels, leading to improved F1 scores in both the real-world dataset and the PSM dataset. This advancement has the potential to contribute to the development of more effective depression diagnostic tools. In addition, we used SHapley Additive exPlaination (SHAP) to validate the interpretability of our model, ensuring its practical applicability in medical settings.

Figures (11)

• Figure 1: An overview of the paper. Our methodology which consists a robust biomarker and a customized graph-based neural network manages to resolve shortcomings in depression diagnosis task and got a competitive F1 score on our larger dataset.
• Figure 2: Diagram of DFT and iDFT in our work. After amplitude selection and iDFT, new temporal brain channel series reserve basic trends of origin temporal series. The middle topk frequency domain compoments can be used to extract our biomarkers.
• Figure 3: (a) An overview of our TGCN based model. Three independent TGCN modules deal with HbO, HbR and their total concentration representively. (b) Details of GCN block which was used to aggregate information from neighbor brain channels in this work.
• Figure 4: Consistency vertification of our novel biomarker. Y-axis represents brain channels and x-axis represents frequencies' index, where a larger x coordinate represents sinusoid with higher frequency. Color means relative size of $|r|$.We found that frequency component of specific brain regions was strongly associated with depression.
• Figure 5: Temporal fusion model. Green block means Original Temporal Feature which contains six statistics of FNIRS. Blue block means our Dynamic Temporal Feature under frequency domain. Red dot block represents alternative FAM module to enhance model performance.