Neural activity during inhibitory control predicts suicidal ideation with machine learning

The study addresses the need for scalable, affordable, and biologically grounded biomarkers to identify individuals with suicidal ideation (SI). Suicide rates have risen markedly, and existing prediction models often rely on self-report, social factors, or costly neuroimaging that limit clinical scalability. Prior neuroimaging work suggests alterations in prefrontal, default mode, and sensorimotor networks in SI, but many studies did not control for confounders such as depression/anxiety severity, age, or sex. EEG offers a cost-effective alternative, yet prior EEG findings for SI have been mixed and frequently confounded. The authors hypothesized that EEG signals recorded during cognition—particularly tasks taxing inhibitory control, interference processing, working memory, and emotion bias—would improve the identification of neural biomarkers of SI relative to resting-state signals. They specifically posited that cortical source-localized power in canonical frequency bands during cognitive events would enable accurate ML-based classification of SI, even when groups are matched on demographics and depression/anxiety symptoms.

The paper reviews ML efforts using social media and self-report features for SI risk detection, which show limited accuracy. Neurobiological predictors from MRI/fMRI have linked SI to altered activity/connectivity in prefrontal regions and default mode/sensorimotor networks, but many studies lacked control for key confounders. EEG studies of suicidality report mixed results: some found no alpha-band connectivity differences post-attempt, others noted reduced theta or elevated frontal gamma in SI, often without controlling for depression/anxiety differences. Other EEG biomarkers (e.g., alpha asymmetry, microstate differences) have yielded limited or confounded associations. Most prior ML approaches used resting-state data, providing less insight into task-evoked neural processing differences. Cognitive domains implicated in suicidality include inhibitory control deficits, emotional dysregulation, impaired interference processing, working memory deficits, and negative emotional bias. This motivates task-based EEG analyses to uncover more specific neural markers of SI.

Design and participants: A matched case-control design included 76 community participants (38 SI+, 38 SI−; median age 23 ± 7.6 MAD; 33 males), recruited 2018–2020 under UCSD IRB approval (protocol #180140). SI status was determined by C-SSRS (SI− = 0; SI+ = 1–3). Groups were matched on age, sex, anxiety (GAD-7), depression (PHQ-9), socio-economic status, and ethnicity (no significant differences). No suicide attempts were reported. Antidepressant use: 7 SI+, 5 SI−. Exclusions relied on self-report; no structured clinical interview was performed for the main sample. External validation cohort: 35 clinically depressed participants from two clinics (median age 57 ± 10 MAD; 16 males; 12 SI+), 2022–2023 (UCSD and VA IRBs). SCID-confirmed depression; common antidepressant use (9/12 SI+; 18/23 SI−). They completed adaptive versions of tasks for repeated-assessment aims; baseline data were used for validation. Neurocognitive tasks and EEG acquisition: Participants completed four BrainE tasks in one ~40-minute session: Inhibitory Control (IC; Go/Wait), Interference Processing (IP; flanker-like middle fish), Working Memory (WM; delayed target probe), and Emotion Bias (EB; arrow-over-face), plus a 3-minute eyes-closed resting state. Task trial structure included cue, stimulus (100 ms), response window (task-dependent), feedback (200 ms), and inter-trial interval (500 ms). Behavioral metrics: d' (scaled to 1), response time (RT), and response time consistency (1/CV); WM also included item span. EEG: 24-channel Smarting wireless EEG, 10–20 layout, 250 Hz sampling, 24-bit resolution; event timing via Lab Streaming Layer. Data were analyzed at scalp and in cortical source space (source imaging per prior methods), with spectral power in theta (4–8 Hz), alpha (8–13 Hz), and beta (13–30 Hz). Source regions were grouped into 8 networks: Fronto-Parietal (FPN), Cingulo-Opercular (CON), anterior DMN (aDMN), posterior DMN (pDMN), medial temporal lobe DMN (mtIDMN), Visual, Sensory-Motor (SM), and Ventral Attention (VAN). Scalp electrodes were grouped into 7 regions (frontal medial/left/right, central, posterior occipital medial/left/right). Event windows: Cue activity averaged 0–500 ms post-cue; stimulus processing 100–500 ms post-stimulus; response activity 50–150 ms post-response; feedback activity 100–400 ms post-feedback. Cue and stimulus analyses used all clean trials; response/feedback used correct trials (mean incorrect rate 9.36% ± 10.38% SD). Rest: time-averaged epochs. Some EEG task data were missing/corrupted for a few participants (IC missing in 2; IP/WM/EB missing in 3). Model datasets (a priori): Primary category: task-related source network power for each task (IC, IP, WM, EB) and rest, in each frequency band (theta/alpha/beta), yielding 15 datasets. For each task-band dataset, 32 features (4 events × 8 networks); rest had 8 network features per band. Secondary categories: (1) Task-related source power + performance (add d', RT, consistency; WM adds span; 35–36 features); (2) Event-related source power across tasks (4 events × 8 networks across IC/IP/WM/EB; 32 features per band); (3) Task-related scalp power (same structure as primary but 7 scalp regions). No dimensionality reduction (e.g., PCA) was used to preserve interpretability; datasets were pre-specified to avoid leakage/overfitting. ML pipeline: Implemented in Python 3 with scikit-learn using a pipeline wrapper. Preprocessing: iterative imputation for missing values (about 2.3% imputed in event-related datasets), standard scaling. Models: Logistic Regression (LR), Decision Tree (DT), and Multilayer Perceptron (MLP). Evaluation: repeated five-fold nested cross-validation with grid-search hyperparameter tuning; outer folds provided validation on data unseen in training. Optimization metric: Matthews correlation coefficient (MCC); sensitivity (SEN) and specificity (SPE) also computed; naive baseline MCC=0, SEN=0.5, SPE=0.5. Feature importance: SHAP was applied to the best-performing model to rank features and visualize directionality. Top features were further examined with group-wise tests and raincloud plots. External validation: The final best model from the main sample was applied without retraining to the independent clinical sample that completed adaptive versions of the tasks.

- Behavioral performance: No significant SI+/SI− differences across tasks for d', RT, consistency, or WM span after FDR correction (Table 2). - Best overall model: Task-related source power, beta band during the Inhibitory Control (IC) task using Logistic Regression achieved MCC 88% (±8), sensitivity 89% (±11), specificity 98% (±5), and overall accuracy ~93% across nested CV (Fig. 3; Table 3). - Frequency/task specificity: Beta-band IC task source power outperformed corresponding theta/alpha models and other tasks and rest. Separating IC trial types (Go vs. Wait) did not improve performance. - Source vs. scalp: Source-space models substantially outperformed scalp-space models (e.g., best scalp model: LR beta WM, SEN 78% ±15, SPE 55% ±17, MCC 35% ±23; Table 3). - Alternative dataset organization: The best event-related source model (MLP; alpha band; response period aggregating all tasks) reached SEN 87% ±16, SPE 100% ±0, MCC 88% ±15 (Table 3), but the primary best model remained superior on average considering robustness and category comparisons. - SHAP-derived top predictors in best model (LR beta IC): Higher beta power predicted SI+ for (1) feedback-period Visual network, (2) response-period VAN, (3) response-period FPN, (4) response-period SM, and (5) feedback-period pDMN (Figs. 4–5). Group differences for these features were significant after FDR correction; SI+ showed greater power than SI−. - External validation (independent depressed clinical sample, adaptive tasks): Applying the primary best model yielded SEN 50%, SPE 61%, MCC ≈ 0.10, suggesting limited generalization under methodological differences and/or sample shift.

Task-evoked, source-localized EEG beta power during inhibitory control robustly classified SI status in a sample matched on demographics and depression/anxiety, addressing confounds common in prior work. Findings support the hypothesis that cognitive control processes, particularly response execution/inhibition and feedback processing, reveal neural biomarkers of SI beyond what is captured at rest or in behavioral performance alone. Elevated beta power in ventral/fronto-parietal attention and sensorimotor networks during responses and in visual/posterior DMN during feedback predicted SI+, potentially reflecting compensatory neural engagement to maintain similar behavioral performance or altered cognitive control/feedback integration mechanisms. Source localization proved critical, likely by enhancing spatial specificity and reducing volume conduction effects. While an event-related alpha response model also showed strong internal CV performance, the IC beta source model provided the most consistent classifier across task-constrained datasets. The reduced performance in an external depressed clinical sample, obtained with adaptive tasks and fewer trials, underscores challenges in out-of-sample generalization due to methodological differences, smaller training data, and clinical heterogeneity. Nonetheless, the approach demonstrates a scalable, interpretable neurophysiological biomarker framework that can inform risk identification and potential intervention targets for suicidality.

This study demonstrates that source-localized EEG beta power during an inhibitory control task accurately classifies suicidal ideation in a matched community sample, achieving high sensitivity, specificity, and MCC. SHAP analyses identified key networks—ventral attention, fronto-parietal, sensorimotor, visual, and posterior DMN—during response and feedback epochs as principal contributors. The findings highlight the utility of task-based, physiologically interpretable EEG biomarkers over resting-state or scalp-level features. Future work should: (1) expand sample sizes and include multi-site datasets; (2) harmonize task protocols (avoid adaptive vs. fixed-task mismatches) to improve external validity; (3) incorporate richer EEG features (e.g., temporal dynamics, connectivity) and advanced ML methods; (4) explore channel optimization for clinical scalability; and (5) evaluate prediction of other suicidal behaviors and translation into clinical workflows.

- Sample size and generalizability: Although larger than many prior EEG studies, the N=76 main sample is modest; external validation (n=35) used adaptive tasks, limiting direct comparability and reducing generalization. Single-site data may not generalize to other settings. - Methodological differences: External validation tasks were adaptive with fewer trials, potentially increasing noise and reducing model transferability. - Feature/model scope: Only spectral power features were used; more complex temporal/connectivity features and advanced ML architectures were not explored and could improve performance/generalizability. - EEG spatial resolution: 24-channel recordings may limit spatial specificity; channel count optimization could affect performance and scalability. - Demographic considerations: Although groups did not differ significantly in ethnicity, SI prevalence varies across minoritized groups; both groups were majority female, which may limit generalizability across sexes. - Clinical scope: Main model differentiates SI presence in a community sample rather than predicting attempts or broader suicidal behaviors; structured clinical interviews were not used in the main sample.