Error-Related Negativity (ERN)

• ERN is a negative-going event-related potential occurring 0–100 ms post-error, indicating rapid error detection and performance monitoring in the ACC.
• Controlled paradigms such as Go/No-Go and Stroop tasks yield quantifiable behavioral metrics (e.g., 30% commission errors) essential for extracting ERN features.
• Analysis of ERN features, including peak amplitude and theta power, shows promising 80–90% accuracy for classifying anxiety and other internalizing disorders.

Error-related negativity (ERN) is a prominent event-related potential (ERP) component that appears as a sharp, negative-going deflection in the scalp-recorded EEG, peaking approximately 0–100 ms after an erroneous response. Most prominent at fronto-central electrodes (notably Fz, FCz, and Cz), the ERN is widely considered an index of rapid error detection and performance monitoring. Converging evidence implicates the anterior cingulate cortex (ACC), particularly its dorsal aspect (dACC), as the principal cortical generator underlying the ERN. The amplitude and dynamics of the ERN have been foundational in characterizing neural responses to errors, with extensive application to the study of psychopathology, cognitive control, and, more recently, machine-learning–driven diagnostic paradigms (Perera et al., 13 Apr 2025, Chandrasekar et al., 2024).

1. Neurophysiological Basis and Theoretical Context

The ERN emerges in response to an action that deviates from task instructions, predominantly in speeded choice paradigms such as the Flanker, Go/No-Go, and Stroop tasks. Theoretical accounts posited by Holroyd & Coles (2002) and Botvinick et al. (2001) interpret the ERN as a neural signature of performance monitoring, signaling a mismatch between intended and actual actions to facilitate the adjustment of cognitive control (Perera et al., 13 Apr 2025). Source localization, including intracranial recordings, consistently attributes its origin to the dACC and adjacent medial prefrontal regions (Chandrasekar et al., 2024).

2. Experimental Paradigms and Behavioral Metrics

Typical ERN research employs a controlled error-eliciting task. An example is the Go/No-Go response-inhibition paradigm, with 350 experimental trials (80 % Go, 20 % No-Go) and a well-defined sequence: fixation cross (500 ms), stimulus (200 ms), blank (600 ms), and feedback (500 ms). Participants respond rapidly to Go trials and withhold responses on No-Go trials, producing quantifiable error rates (e.g., No-Go commission ≈ 30 %, Go omission ≈ 0.02 %) (Perera et al., 13 Apr 2025). These tightly controlled paradigms ensure that error-related brain responses are both frequent and temporally aligned for ERP extraction.

3. EEG Acquisition, ERP Processing, and ERN Quantification

ERN extraction adheres to high-fidelity EEG protocols. A representative workflow involves:

• Hardware: 32-channel dry, wireless EEG (e.g., Cognionics HD-72), configured according to the international 10–20 system, initially referenced to the left mastoid and re-referenced to a common average offline (Perera et al., 13 Apr 2025).
• Acquisition: 500 Hz sampling, 0.1–250 Hz hardware bandwidth.
• Preprocessing steps:
  1. Band-pass filtering (1–30 Hz).
  2. Epoching around each response (−500 ms to +500 ms).
  3. Baseline correction (−400 ms to −200 ms).
  4. Independent Component Analysis (Infomax runica) for artifact reduction.
  5. Epoch rejection for amplitudes exceeding 100 µV; exclusion of participants with high artifact rates.
  6. On average, 13.6 % of trials are removed, yielding mean artifact-free counts (e.g., 167 correct Go, 66 incorrect No-Go trials).

ERN amplitude is quantified at an electrode such as Fz over the window −80 ms to +100 ms relative to response onset. The canonical formula for mean amplitude over $N_\text{error}$ trials is:

$$\mathrm{ERN} = \frac{1}{N_\text{error}} \sum_{i=1}^{N_\text{error}} \left( \frac{1}{t_\text{off} - t_\text{on}} \int_{t_\text{on}}^{t_\text{off}} EEG_i(t)\,dt \right)$$

The difference waveform, termed ΔERN or AERN, captures error-specific activity by subtracting the mean amplitude for correct responses (CRN):

$\Delta\mathrm{ERN} = \mathrm{ERN} - \mathrm{CRN}$

4. ERN Features in Psychopathology and Individual Differences

ERN amplitude serves as a transdiagnostic biomarker for internalizing disorders. Studies consistently report larger (more negative) ERN in anxiety, obsessive–compulsive symptoms, and depression (Perera et al., 13 Apr 2025, Chandrasekar et al., 2024). However, the ERN–depression association is not uniform, likely due to moderating variables. Socioeconomic status (SES) has been identified as a crucial moderator: low-SES individuals exhibit larger (more negative) ERN amplitudes, with SES fully mediating the relation between depressive symptoms and ERN. For instance, in a cohort from Kuala Lumpur (Perera et al., 13 Apr 2025):

• Low-SES: AERN mean ≈ –1.8 µV (SD = 2.9)

• High-SES: AERN mean ≈ –0.18 µV (SD = 2.4)
• SES vs. AERN: $r = .40$, $p < .001$
• Depression vs. AERN: $r = –.27$, $p < .001$
• Regression form: $$\mathrm{AERN} = \beta_0 + \beta_1\, SES + \beta_2\, \mathrm{CESD} + \beta_3\, \mathrm{Gender} + \varepsilon$$

SES, rather than depression per se, appears to predict individual variation in ERN amplitude. This suggests environmental adversity or punitive contexts may sensitize low-SES individuals to errors, reflected in enlarged ERN responses even with equivalent behavioral performance (Perera et al., 13 Apr 2025). A plausible implication is that ERN findings in psychopathology studies may conflate SES effects with neural correlates of clinical risk.

5. Feature Extraction, Machine-Learning Applications, and Metrics

ERN features have been systematically utilized in both conventional statistics and machine-learning approaches, especially for anxiety detection. The two principal feature sets are:

Time-domain:

• Peak amplitude: $$A_\text{peak} = \min_{t\in [t_1, t_2]} A_\text{error}(t)$$
• Mean amplitude: $$A_\text{mean} = \frac{1}{t_2 - t_1} \int_{t_1}^{t_2} A_\text{error}(t)\,dt$$
• Area under the curve: $$\mathrm{AUC} = \int_{t_1}^{t_2} |A_\text{error}(t)|\,dt$$
• Latency: $$t_\text{latency} = \arg\min_{t\in [t_1, t_2]} A_\text{error}(t)$$

Time–frequency:

• Wavelet/Fourier transform to compute $W(f,t)$
• Instantaneous power: $P(f,t)=|W(f,t)|^2$
• Average theta-band ($4$–$8$ Hz) power in the ERN window: $$\mathrm{Power}_\theta = \frac{1}{F_2-F_1} \int_{F_1}^{F_2} \int_{t_1}^{t_2} P(f,t)\,dt\,df$$ (Chandrasekar et al., 2024)

These features are concatenated into vector representations for classification via support vector machines (SVM), random forests, logistic regression, and—more recently—deep architectures (CNNs, RNNs). Empirical studies report classification accuracies of 80–90% for ERN-based group separation tasks, typically validated using k-fold cross-validation or leave-one-subject-out approaches. However, most research still centers on statistical group differences rather than comprehensive classifier evaluation (e.g., precision, recall, F1) for clinical prediction (Chandrasekar et al., 2024).

6. Methodological and Clinical Caveats

The use of ERN as a biomarker is constrained by several limitations:

• Task specificity: ERN requires controlled error-elicitation tasks, challenging ecological validity and routine deployment (Chandrasekar et al., 2024).
• Feature variability: No consensus exists regarding the most robust extractive feature (peak amplitude, ΔERN, time–frequency measures), impeding cross-study synthesis.
• Demographic diversity: The generalizability of ERN findings is hampered by underrepresentation of low-income and non-college samples. SES is a key confound to consider.
• Limited ML adoption: Most studies rely on univariate statistics rather than fully integrated, generalizable multivariate or deep-learning pipelines.

A critical clinical implication is that ERN’s diagnostic value for internalizing disorders can be compromised if SES confounds are unaccounted for. This suggests that future studies should stratify or adjust for SES to enhance ERN’s specificity and utility, particularly in low-resource settings (Perera et al., 13 Apr 2025).

7. Future Directions and Open Challenges

Several avenues are identified to advance the field:

1. Multimodal fusion: Combining ERN with resting-state EEG and peripheral autonomic signals (e.g., heart rate, skin conductance).
2. Feature selection: Systematic algorithms to distill the most discriminative ERN-derived features for anxiety classification.
3. Large-scale, diverse cohorts: Multi-center studies to assess ERN test–retest reliability and moderation by SES dimensions (income, education, occupation).
4. Ecological validity: Development of naturalistic error-elicitation tasks (e.g., in virtual reality) to capture real-world cognitive–emotional responses.
5. Clinical deployment: Addressing population diversity and model generalizability for robust, non-invasive, objective diagnostics (Chandrasekar et al., 2024, Perera et al., 13 Apr 2025).

Replication in geographically and culturally distinct samples is essential. Disentangling which SES components most strongly influence ERN remains a high-priority research goal. Consideration of environmental, developmental, and methodological moderators will be central to the translation of ERN into reliable clinical and translational tools.