An interpretable TimeMIL framework for fNIRS: differential diagnosis between schizophrenia and bipolar disorder

Background Schizophrenia (SCZ) and bipolar disorder (BD) exhibit substantial overlap in clinical symptoms, making differential diagnosis a persistent challenge and leading to a high risk of misdiagnosis. Recently, functional near-infrared spectroscopy (fNIRS) during the verbal fluency task (VFT) has provided more objective neurofunctional markers for distinguishing between the two disorders than traditional behavioral scales. However, most existing studies have predominantly relied on conventional, hand-crafted feature analyses. A significant limitation of this approach is the scarcity of deep learning models tailored for high-dimensional fNIRS time-series data, coupled with a lack of methods for their interpretation. Methods In this study, a time-aware multiple instance learning model (TimeMIL) was applied to fNIRS data (healthy controls [HC]: 52 participants, SCZ: 50 participants, BD: 67 participants) collected during VFT. This model captures global dependencies in multivariate, high-dimensional time series of fNIRS, which represents each fNIRS recording as a bag and automatically locates informative temporal segments via a learnable wavelet encoder and self-attention. Furthermore, an interpretability framework integrating GradientSHAP and Integrated Gradients was proposed to quantify segment-and region-level attributions. Model performance was evaluated using 5×5-fold cross-validation and an independent test set. Results TimeMIL achieved 0.928 ± 0.016 accuracy and a macro-averaged AUC of 0.984 ± 0.007 for three-class classification, significantly outperforming other representative models, including one-dimensional convolutional neural networks (1D-CNNs), Transformers, and temporal convolutional networks (TCNs). Attribution analyses highlighted robust differences in the orbitofrontal cortex (OFC) across HC, SCZ, and BD, suggesting its potential as a candidate discriminative feature. The frontal pole cortex (FPC) and ventrolateral prefrontal cortex (VLPFC) exhibited SCZ-specific attribution patterns. Conclusion This study represents the first application of TimeMIL to fNIRS-based psychiatric disease classification, accompanied by a dedicated interpretability framework that illuminates prefrontal dysfunction. By integrating high accuracy, robustness, and interpretability, the proposed interpretable TimeMIL framework facilitates development of objective, neuroimaging-based tools for early screening and differential diagnosis of major psychiatric disorders, pending further validation in clinical settings.