Quality-aware mask-propagation graph attention for fNIRS-based bipolar disorder classification without channel interpolation

Objective. Bipolar disorder (BD) remains difficult to characterise objectively, motivating a neural marker to complement clinical assessment. Functional near-infrared spectroscopy (fNIRS) offers portable cortical haemodynamic measurement, but channel availability can vary across recordings because of bad channels. Channel deletion changes the graph input, whereas spatial interpolation reconstructs unavailable signals. We therefore propose a quality-aware mask-propagation graph attention framework that retains all channels and propagates a continuous channel-quality mask through graph attention and a masked readout. Approach. The framework was instantiated as a SoftMaskGAT classifier with a continuous quality mask q∈[0,1], an operational quality proxy. It was evaluated on 197 participants (93 with BD and 104 healthy controls (HC)) recorded with a 48-channel continuous-wave fNIRS system during a verbal fluency task, using leave-one-subject-out (LOSO) cross-validation across three seeds. Two pluggable temporal-aggregation variants — global average pooling (GAP) and attention-based aggregation (Attn) — were combined by a parameter-free 50/50 probability fusion. Ten-fold comparisons examined channel deletion, spatial interpolation and binary weighting; descriptive encoder, channel-loss and regional-attribution analyses characterised model behaviour. Main results. At the aggregate level, GAP attained an area under the receiver operating characteristic curve (AUC) of 0.807, Attn 0.812 and the 50/50 fusion 0.825 [95% confidence interval (CI) 0.760, 0.885]. Descriptive analyses showed a reversal in the GAP–Attn AUC ordering between the lowest and highest quartiles of participant-mean q. The right perisylvian region ranked among the top three across surrogate feature attribution, graph attention and regional ablation. Significance. SoftMaskGAT supports interpolation-free fNIRS-based BD classification across different channel-availability patterns while preserving a common spatial graph and the correspondence between observed inputs and measurement locations. By linking regional responses to observed channel coverage, quality weights and direct readout share, the framework makes the measurement basis of regional model interpretation explicit.