Neural activity during inhibitory control predicts suicidal ideation with machine learning

Suicide is a significant public health concern, with rates increasing globally. Developing accessible and effective biomarkers for suicidal ideation (SI) is crucial for prevention and intervention. While neuroimaging techniques like fMRI have been explored, their cost and accessibility limit widespread application. This study investigates the potential of EEG, a more affordable and scalable technology, for SI prediction using ML. Previous attempts at using ML for SI prediction have often faced challenges, including small sample sizes, inadequate control for confounding variables (like depression and anxiety), and reliance on resting-state data which may not capture the nuances of cognitive processing related to SI. Neurocognitive deficits, particularly in inhibitory control, interference processing, working memory, and emotional bias, are linked to suicidality. This study hypothesizes that analyzing EEG signals during these cognitive tasks will enhance SI classification accuracy compared to resting-state analysis. The study will employ a carefully matched cohort to control for age, sex, and symptoms of depression and anxiety, addressing previous methodological limitations.

Existing research on biological predictors of suicide has utilized various methods, including neuroimaging (MRI/fMRI). fMRI studies have identified alterations in prefrontal cortical areas and connectivity in individuals with SI. Machine learning (ML) approaches have further explored alterations in default mode and sensorimotor networks. However, the high cost of fMRI limits its accessibility for widespread screening. While some studies have explored EEG biomarkers of suicidality, results have been mixed and often confounded by lack of control for depression and anxiety severity. Furthermore, most studies have focused on resting-state data, neglecting the potential insights from task-related neural activity. This study aims to address these gaps by employing a rigorous methodology, including a matched cohort and focus on task-related EEG data during cognitive tasks relevant to SI.

Seventy-six subjects (38 with SI, 38 without) participated in the main study, matched for age, sex, depression, and anxiety symptoms. An independent validation set of 35 clinically depressed subjects (12 with SI) was also included. Participants completed four neurocognitive tasks (inhibitory control, interference processing, working memory, and emotion bias) and a resting-state EEG recording using a 24-channel wireless EEG system. EEG data were processed to extract power in theta, alpha, and beta frequency bands. Cortical source imaging was used to localize neural signals. Various ML models (logistic regression, decision tree, multilayer perceptron) were trained on different datasets (task-related source network power, task-related source network power + task performance, event-related source network power, and task-related scalp power) using a five-fold nested cross-validation approach. Model performance was assessed using Matthews correlation coefficient (MCC), sensitivity, and specificity. Shapley Additive exPlanations (SHAP) were used to determine feature importance. The best-performing model was then applied to the independent validation set.

The best-performing ML model used beta band power during the inhibitory control (IC) task, demonstrating high sensitivity (89%) and specificity (98%). Shapley explainer plots revealed that feedback-related beta power in the visual and posterior default mode networks, and response-related beta power in the ventral attention, fronto-parietal, and sensory-motor networks, were top predictors of SI. External validation in a clinically depressed sample yielded moderate sensitivity (50%) and specificity (61%). No behavioral differences in task performance were observed between SI+ and SI- groups. The superior performance of source-localized data compared to scalp EEG data highlights the importance of accurate spatial localization of neural activity for SI prediction.

The findings support the feasibility of using EEG-based biomarkers for predicting SI. The high accuracy of the model, particularly its specificity, suggests its potential for identifying individuals at high risk of SI. The involvement of beta oscillations in key brain networks related to attention, cognitive control, and self-agency aligns with existing literature on the neural underpinnings of suicidality. The increased beta power observed in SI+ individuals during the IC task may reflect compensatory mechanisms to maintain behavioral performance rather than impaired cognitive function. This highlights the importance of considering not only behavioral performance but also underlying neural activity when assessing SI risk. The moderate performance in the external validation dataset may be attributed to differences in task administration (adaptive vs. fixed response windows) and the smaller sample size. However, the overall results suggest promise for the use of EEG in suicide risk assessment.

This study demonstrates that EEG data from a readily-administered inhibitory control task, analyzed with machine learning, can effectively predict suicidal ideation. The identification of specific neural biomarkers strengthens the potential of EEG as a clinically relevant and scalable tool for suicide risk assessment and intervention. Future research with larger, more diverse samples and further investigation into other EEG features and more complex ML models could further enhance the predictive power and generalizability of this approach.

Despite its promising results, this study has limitations. The relatively small sample size, especially in the external validation set, could affect the generalizability of the findings. The use of a single cognitive task and a limited set of EEG features might not capture the full complexity of neural activity related to SI. Further research with larger sample sizes is required to validate these findings across diverse populations and clinical settings. The differences in task administration between the main study and the validation set might have impacted the generalizability of the findings. The reliance on self-reported SI also presents a potential limitation.