Reduced prefrontal activation and region-specific laterality reversal in bipolar depression during verbal fluency: a 52-channel fNIRS study with leakage-free machine-learning discrimination

Abstract Background Bipolar (BPAD) and unipolar (MDD) depression are difficult to distinguish yet diverge in treatment. Functional near-infrared spectroscopy (fNIRS) during a verbal fluency task (VFT) probes prefrontal function, but three-group comparisons under leakage-free machine learning are scarce. Methods This single-center cross-sectional study recorded 52-channel fNIRS during a phonemic VFT in 173 adults (64 healthy controls [HC], 61 MDD, 48 BPAD). Oxygenated-hemoglobin activation (VFT > Counting) was compared across eight Brodmann-based regions and four laterality indices with false-discovery-rate (FDR) correction. Three classification tasks used leakage-free nested cross-validation with permutation testing and Shapley additive explanation (SHAP) attribution. Results Both patient groups showed less prefrontal activation than HC (HC vs. BPAD Cohen’s d = 0.58, p = 0.003; HC vs. MDD Cohen’s d = 0.49, p = 0.010), most focally in the left inferior frontal gyrus/Broca’s area (HC vs. BPAD FDR p = 0.005, Cohen’s d = 0.60), persisting after age adjustment. BPAD showed a region-specific laterality reversal: a leftward dorsolateral prefrontal bias (laterality index + 0.22, p = 0.021) and loss of leftward Broca asymmetry seen in HC (+ 0.18, p = 0.032), though between-group laterality was not FDR-significant. Region-level features discriminated patients from HC (AUC 0.654, 95% CI 0.608–0.700, permutation p = 0.010), with SHAP converging on left Broca; BPAD versus MDD did not exceed chance (0.587, p = 0.31). Conclusions Prefrontal hypoactivation during verbal fluency, focal to left Broca, characterizes depression and separates patients from controls. Bipolar depression differs from controls in a region-specific laterality reversal rather than global magnitude, potentially informing laterality-guided brain stimulation.