Clinical Scorecard: A Transparent TimeMIL Approach for fNIRS: Distinguishing Schizophrenia from Bipolar Disorder

At a Glance

Key Highlights

• TimeMIL achieved 0.928 ± 0.016 accuracy and a macro-averaged AUC of 0.984 ± 0.007 for three-class classification.
• The model outperformed traditional deep learning models such as 1D-CNNs, Transformers, and TCNs.
• Attribution analyses indicated significant differences in the orbitofrontal cortex (OFC) among HC, SCZ, and BD.
• The study introduced an interpretability framework combining GradientSHAP and Integrated Gradients.
• TimeMIL represents the first application of this model to fNIRS-based psychiatric disease classification.

Guideline-Based Recommendations

Diagnosis

• Utilize fNIRS during verbal fluency tasks to differentiate between SCZ and BD.

Management

• Incorporate objective neuroimaging tools for early screening and differential diagnosis.

Monitoring & Follow-up

• Assess prefrontal dysfunction through fNIRS-based evaluations.

Risks

• Misdiagnosis due to symptom overlap between SCZ and BD.

Patient & Prescribing Data

Healthy controls (HC), individuals with schizophrenia (SCZ), and individuals with bipolar disorder (BD).

The study emphasizes the need for objective biomarkers to guide personalized care.

Clinical Best Practices

• Implement high-dimensional time-series analysis for neuroimaging data.
• Ensure interpretability of models to validate clinical relevance.
• Use machine learning to capture subtle hemodynamic and connectivity differences.

Related Resources & Content

• Source Article