Deficient prefrontal-amygdalar connectivity underlies inefficient face processing in adolescent major depressive disorder

Major depressive disorder (MDD) often emerges during adolescence and is linked to cognitive negative biases in emotion processing. Converging evidence implicates dysregulation within the prefrontal-amygdala network—including aberrant activity in amygdala, subgenual anterior cingulate cortex (sgACC), and lateral prefrontal cortex (LPFC)—and altered functional coupling among these regions in adolescent MDD. The study investigates whether adolescents with MDD show aberrant emotion processing reflected in differences in evidence accumulation during dynamic facial affect matching, and whether such behavioral inefficiency relates to altered effective connectivity within the prefrontal-amygdala circuitry. Hypotheses: (a) adolescents with MDD exhibit reduced processing efficiency (drift rate) especially under ambiguity; (b) disrupted connectivity among sgACC, LPFC, and fusiform face area (FFA); (c) reduced top-down prefrontal influence on the amygdala; and (d) differential engagement of the network by positive versus negative valence.

Prior work shows adolescents at risk for and with MDD exhibit heightened amygdala reactivity during emotional face processing and altered activation in sgACC and LPFC. Functional coupling within the prefrontal-amygdala network appears disrupted in youth MDD, including altered sgACC–amygdala interactions and decreased sgACC–FFA connectivity during face processing. Longitudinal evidence suggests prefrontal-amygdala interactions can predict treatment response. In adults, disrupted prefrontal-amygdala effective connectivity during emotional processing is well documented. Collectively, literature suggests a network-level dysfunction involving perceptual (FFA), cognitive control (LPFC), and affective (amygdala/sgACC) nodes underpinning biased emotion processing in MDD.

Design: Cross-sectional case-control study combining computational modeling of decision processes with fMRI and dynamic causal modeling (DCM) during a dynamic face- and shape-matching task. Participants: 30 adolescents with MDD (mean age 16.1 ± 1.5 years) and 33 healthy controls (mean age 16.2 ± 1.9), matched on age, IQ, sex, and handedness. Clinical assessment via Kiddie-SADS or MINI-KID. Exclusion for controls: current psychiatric disorder, major medical illness, drug abuse, MRI contraindications, pregnancy, neurological injury. Most patients received SSRIs (n = 18) and psychoeducation as needed. Informed consent obtained; ethics compliant. Task: Dynamic matching with 4 blocks per condition (positive, negative, neutral, shapes), 80 trials total. Targets gradually morphed from neutral to target emotion. Neutral: neutral or contempt expressions; Positive: happy and positively rated surprise; Negative: sad and disgust. Shape control: match number of vertices of evolving shapes. Responses via two-button box; eye-tracking for gaze; stimuli via goggles (600×600). Post-scan continuous ratings of valence and arousal. Computational modeling: Hierarchical Linear Ballistic Accumulator (LBA) with two accumulators (left/right), parameters: drift rate (processing efficiency), threshold β (response caution), non-decision time γ, starting point α. Compared models where drift rate or threshold varied by condition vs a null model. Fitted separately in groups using DE-MCMC (36 chains), weak priors; convergence monitored; posterior inference at individual and group levels. Behavioral analysis: Linear mixed-effects models on log RT, accuracy, omissions; also on valence and arousal ratings. Fixed factors: group, condition (valence/type); random effects: subjects. Correlated ratings with model parameters (two-tailed, p < 0.05). MRI acquisition: 3T Philips Achieva, 32-channel head coil. EPI: TR 1600 ms, TE 35 ms, voxel ~2.24×2.24×2.7 mm3, flip 75°, gap 0.35 mm, SENSE 2, multiband 2, tilted FOV to enhance ventral signal; 233 vols/session; discard first 5. T1 structural: 1.05×1.05×1.2 mm3. Preprocessing (SPM12): slice-time correction, realignment, normalization to MNI-152, 6 mm FWHM smoothing; motion scrubbing for framewise displacement >1 mm. fMRI analysis: First-level GLM with regressors for conditions convolved with HRF; included motion and DCM-related nuisance regressors; high-pass 128 s; AR(1). Second-level t-tests for task activations; voxel-wise p < 0.001 with cluster-extent FWE via Monte Carlo (10,000 iterations; minimum cluster ~240 mm3/55 voxels) and AAL-based labeling. Regressions assessed association between LBA drift rate and whole-brain activity. DCM ROI selection: Guided by prior adolescent MDD literature and model-based analyses. ROIs from faces > shapes: amygdala, FFA, LPFC, plus sgACC; 6 mm spheres around participant-specific maxima within group-defined peaks; constrained by Neurosynth coactivation masks; extracted first eigenvariate/time-derivative from active voxels (p < 0.05). DCM estimated full connectivity (A-matrix), valence-dependent modulations (B-matrix), and inputs; group comparisons via empirical Bayes; leave-one-out cross-validation (LOOCV) for predictive validity; SSRI intake modeled as covariate.

Behavioral performance: RTs showed a strong valence effect, F(3,4666.9) = 5107.0, p < 10^-15 (fastest for positive, then negative, then neutral/ambiguous); no main effect of group (F(1,611) = 0.260, p = 0.613) or interaction (F(3,4666.9) = 0.91, p = 0.424). Accuracy: no group effect (F(1,610) = 0.246, p = 0.620) or interaction (F(3,183) = 0.04, p = 0.989); main effect of condition (F(3,183) = 129.9, p < 10^-15) with worse performance for neutral faces. Omissions: no group effect (F(1,611) = 1.23, p = 0.272) or interaction (F(3,183) = 0.83, p = 0.766); more omissions for neutral faces and shapes (F(3,183) = 38.26, p < 10^-15). Ratings: Arousal showed condition-by-group interaction, F(2,7313) = 22.70, p < 10^-15; patients rated negative faces as more arousing. Valence ratings showed condition-by-group interaction, F(2,7313) = 434, p < 10^-15; patients rated positive faces less positively. Neutral-face valence rating correlated negatively with neutral drift rate (β = -0.395, p = 0.002). LBA modeling: Winning model indicated condition-dependent drift rate. MDD showed reduced drift rate for neutral faces versus controls (M_d_neutral = -0.14, 95% HPDI [0.03, 0.24]). Trends (ns) toward lower drift for positive (A_positive = 0.22 [-0.02, 0.3]) and negative faces (A_negative = -0.19 [-0.01, 0.40]). No group differences in shapes drift (A_shapes = 0.05 [-0.13, 0.21]), decision threshold (B_A = -0.20 [-0.67, 0.24]; B_L = 0.26 [-0.30, 0.88]) or non-decision time (A_D = 0.03 [-0.06, 0.13]). Posterior predictive checks reproduced empirical effects. Task fMRI: Faces > shapes activated amygdala, fusiform gyrus, vmPFC, and STS/TPJ. Drift rate was positively associated with sgACC activity; reduced behavioral efficiency linked to greater sgACC deactivation. DCM connectivity (group differences): Decreased bidirectional FFA–LPFC connectivity in MDD (expected value ≈ 0.033 Hz, PP = 1.00 for difference). Decreased sgACC→LPFC coupling (−0.023 Hz, PP = 1.00) with increased LPFC→sgACC coupling (0.054 Hz, PP = 1.00). Decreased sgACC–amygdala connectivity in patients during face processing (−0.053 Hz, PP = 1.00). LOOCV predicted group (AUC = 0.71, 95% CI [0.56, 0.84]); most predictive single edges: FFA→sgACC and FFA→LPFC. Medication effect: SSRI intake associated with decreased sgACC self-connection (−0.145 Hz, PP = 1.00), indicating increased sgACC input sensitivity. Valence-dependent modulation (both groups): During positive and negative faces, amygdala→FFA (pos: −0.434 Hz; neg: −0.26 Hz) and amygdala→LPFC (pos: −0.121 Hz; neg: −0.174 Hz) couplings became more negative; amygdala→sgACC decreased (pos: −0.308 Hz; neg: −0.127 Hz); LPFC→sgACC increased (pos: 0.228 Hz; neg: 0.095 Hz). For positive faces specifically, LPFC–sgACC coupling decreased (−0.186 Hz) and sgACC→amygdala increased (0.291 Hz). None of these context modulations differed between groups.

Findings support that adolescent MDD is characterized by inefficient evidence accumulation during facial affect processing, most evident under ambiguity (neutral expressions), aligning with a cognitive negative bias. Neural results indicate that this behavioral inefficiency is linked to altered prefrontal-amygdala circuitry: disrupted cortical FFA–LPFC–sgACC pathway and reduced sgACC–amygdala coupling. The sgACC, a putative gatekeeper between prefrontal cognitive control and limbic affect, showed deactivation associated with lower processing efficiency, suggesting impaired integration of perceptual (FFA) and regulatory (LPFC/sgACC) signals and diminished top-down influence on amygdala. These distributed coupling abnormalities may lead to inefficient sampling of facial evidence and inappropriate emotional responses, providing a mechanistic account of biased emotion evaluation in adolescent MDD. Valence modulations of connectivity were present across participants and did not differ by group, suggesting a general network alteration rather than valence-specific deficits. SSRI-associated increases in sgACC input sensitivity may reflect treatment-related normalization mechanisms and could inform biomarker development for treatment response. Developmentally, observed connectivity patterns may evolve with maturation, indicating shifting balances between affective reactivity and cognitive control during adolescence.

This study integrates computational modeling with effective connectivity analyses to reveal that adolescents with MDD exhibit diminished information processing efficiency during face matching, particularly for ambiguous expressions, underpinned by disrupted prefrontal-amygdala coupling involving the FFA–LPFC–sgACC pathway and sgACC–amygdala link. These findings advance mechanistic understanding of emotion processing deficits in youth MDD and highlight sgACC-centered network dysfunction as a potential target for intervention. Future work should use longitudinal designs to establish causal trajectories, examine emotion-specific cues beyond coarse valence categories, evaluate developmental changes, and assess whether connectivity metrics can guide personalized treatments, including psychotherapy, pharmacotherapy, and neurofeedback aimed at cognitive bias modification.

Cross-sectional design limits causal inference; modest sample size typical for adolescent clinical neuroimaging may reduce power and generalizability; task yielded high accuracy for positive and negative faces, potentially constraining modeling sensitivity for these conditions; the paradigm dichotomized positive versus negative valence and may not capture emotion-specific nuances (e.g., fear vs sadness), warranting finer-grained affect categories in future studies.