Optimizing depression diagnosis: fNIRS and machine learning differentiate unipolar, bipolar, and healthy states.

BACKGROUND

Differentiating Unipolar Depression (UD) from Bipolar Depression (BD) remains challenging due to overlapping symptoms, which can lead to misdiagnosis and inappropriate treatment. fNIRS provides a portable, objective measure of prefrontal cortical activity, yet its utility for clinically relevant ternary classification (UD, BD, HC) remains underexplored.

METHODS

Prefrontal hemodynamics signals were recorded using 22-channel fNIRS during a verbal fluency task in 171 participants (UD = 72, BD = 59, HC = 40). Six classical machine learning models (Support Vector Machine, Random Forest, XGBoost, Logistic Regression, Naive Bayes, Decision Tree) and two deep learning models (1D-CNN, LSTM) were evaluated for ternary classification using whole-brain (FULL) features and eight prefrontal subregions. All models underwent hyperparameter optimization, stratified 5-fold cross-validation with class-weight adjustment, and data augmentation. Performance metrics included accuracy, sensitivity, specificity, and area under the ROC curve (AUC).

RESULTS

SVM achieved the highest FULL-region accuracy (80.19 %). FULL integration outperformed single-region models by 14.51 % average accuracy. Among subregions, the orbitofrontal cortex (OFC) contributed most to classification performance (average left OFC accuracy: 67.40 %). UD was most reliably identified, whereas BD frequently overlapped with HC, reflecting intrinsic neurophysiological similarity. Deep learning models achieved moderate performance, consistent with limited sample size and high feature dimensionality.

CONCLUSIONS

fNIRS-based machine learning reliably differentiates UD, BD, and HC at the group level. Whole-brain integration yields superior performance, with OFC as a key discriminative region. SVM and Random Forest offer the best balance of accuracy and clinical feasibility. Persistent BD-HC overlap remains a challenge, highlighting the need for larger BD cohorts and complementary multimodal assessments.