Predicting Treatment Response in Female Adolescents With Non-Suicidal Self-Injury Using Neurophysiological Biomarkers and Machine Learning
Article information
Abstract
Objective
This study investigated whether quantitative electroencephalography (qEEG) features, combined with clinical data, could predict treatment outcomes in female adolescents with non-suicidal self-injury (NSSI).
Methods
We analyzed clinical and EEG data from 104 female adolescent inpatients with repetitive NSSI. Resting-state EEG was recorded, and various brain activity patterns across frequency bands were extracted. Clinical outcomes were assessed using pre- and post-admission scores on the Health of the Nation Outcome Scales (HoNOS), Clinical Global Impression-Severity (CGI-S), World Health Organization Disability Assessment Schedule (WHODAS), and Global Assessment of Functioning (GAF). Machine learning models were trained to predict outcomes using EEG and medication data. Model performance was evaluated using cross-validation, and feature importance was interpreted using SHapley Additive exPlanations (SHAP) analysis.
Results
All predictive models demonstrated excellent predictive performance (R2≥0.96, mean squared error [MSE] as low as 0.02). The HoNOS model showed the highest performance (R2=0.99, MSE=0.32), followed by the WHODAS (R2=0.98, MSE=1.32), GAF (R2=0.97, MSE=0.76), and CGI-S (R2=0.96, MSE=0.02) models. Key qEEG predictors included relative low-beta power at Pz, absolute theta power at Fp1, and the delta-to-beta ratio at Cz. Pre-admission clinical severity, particularly CGI-S and HoNOS, also significantly contributed to prediction accuracy.
Conclusion
Our findings suggest that qEEG features, combined with machine learning, can effectively predict treatment response in adolescents with NSSI, supporting their use as neurophysiological biomarkers for individualized care.
INTRODUCTION
Non-suicidal self-injury (NSSI) is defined as the deliberate, direct, and repetitive damage to one’s own body tissue without suicidal intent, carried out in socially unacceptable ways [1]. This behavior often serves as a way to relieve negative emotions such as tension, anxiety, or self-blame, and may also have interpersonal or self-punitive functions [2].
The global prevalence of NSSI has been rising significantly in recent years. International studies report that the prevalence of NSSI in adolescents ranges from 7.5% to 46.5% [3], with approximately 38.9% among college students and 5.5% to 13.4% among adults [4]. In South Korea, a national mental health screening conducted by the Ministry of Education in 2018 found that 7.9% of middle school students and 6.4% of high school students reported a history of self-injury [5]. However, many adolescents engaging in NSSI refrain from seeking medical help due to fear of stigma or a desire to maintain secrecy.
In the past, NSSI was often seen as either a temporary expression of adolescent rebellion, a symptom of personality disorder, or a precursor to suicide. Even without suicidal intent, NSSI remains dangerous because it is linked to both physical and psychological complications, as well as increased risk of suicidal thoughts and attempts. Recognizing its clinical importance, the Diagnostic and Statistical Manual of Mental Disorders-Fifth Edition proposed NSSI as a condition for further study in 2013 [6]. Nevertheless, research exploring NSSI as an independent clinical entity—rather than a symptom of other psychiatric disorders—remains scarce, especially when considering its high prevalence and societal burden [7].
The causes of NSSI are believed to involve individual, environmental, and neurobiological factors. However, due to the population’s heterogeneity, clinical phenotyping alone is often insufficient to understand its pathophysiology [8]. Consequently, recent research has focused on identifying neurophysiological markers, with quantitative electroencephalography (qEEG) emerging as a promising tool due to its high temporal resolution, noninvasive nature, and sensitivity to dynamic brain function.
Previous studies have suggested that individuals engaging in NSSI may exhibit altered frontal cortical activity, including frontal dysfunction and increased gamma power in frontal and temporal regions [9]. Some studies have also reported correlations between NSSI severity and alpha band asymmetry, as well as associations between right hemispheric hypoactivation and depressive symptoms or auto-aggressive behaviors [10]. In a recent event-related potential study, Liu et al. [11] demonstrated that greater prefrontal P300 activity predicted subsequent reductions in NSSI frequency among adolescents with major depressive disorder, highlighting the potential of frontal neurophysiological markers in forecasting treatment outcomes in this population. Despite these findings, most existing studies have primarily focused on depressive symptoms, and there remains a notable lack of research investigating biological predictors of treatment response in adolescents with NSSI.
Given that electrophysiological biomarkers derived from qEEG may reflect early functional impairments associated with NSSI and capture temporal changes in brain dynamics, they hold potential for both early detection and treatment planning [12]. Therefore, this study aims to explore the utility of qEEG features as neurophysiological predictors of treatment response in female adolescents with repetitive NSSI. The ultimate goal is to establish a foundation for personalized and timely interventions in this high-risk population.
METHODS
Participants
This retrospective study reviewed medical records of female adolescent inpatients who were admitted to the Department of Psychiatry at Daegu Catholic University Medical Center between July 1, 2018, and December 31, 2023. A total of 104 participants were included based on the following criteria: female adolescents aged between 9 and 24 years at the time of EEG recording, and a history of at least five or more days of NSSI in the preceding year, as per the Korean Adolescent Protection Law definition of adolescence. Patients with intellectual disability, congenital genetic disorders, schizophrenia or other psychotic disorders, or medical or neurological conditions that may affect brain function (e.g., epilepsy, brain tumors, metabolic encephalopathies) were excluded. The study protocol was reviewed and approved by the Institutional Review Board (IRB) of Daegu Catholic University Medical Center (DCUMC IRB approval No. 2025-09-017). All procedures were conducted in accordance with the Declaration of Helsinki. Additional participant-level information, including clinical diagnoses, age of onset, and pre- and post-admission medication details, is provided in Supplementary Table 1 to enhance clarity and transparency regarding the sample characteristics.
EEG data acquisition and preprocessing
Resting-state EEG was recorded under eyes-closed conditions from 19 scalp electrodes placed according to the international 10–20 system. The raw EEG data were bandpass filtered (1–100 Hz), re-referenced to the average, and cleaned of ocular and muscle artifacts using Independent Component Analysis (ICA) [13]. Preprocessing was implemented using Python-based libraries, including MNE-Python and SciPy. Artifact components were identified using the MNE-ICALabel plugin [14].
qEEG feature extraction
qEEG features were extracted from the preprocessed data, including absolute and relative power values across standard frequency bands—delta (1–4 Hz), theta (4–8 Hz), alpha (8–12 Hz), beta (12–30 Hz), and gamma (30–50 Hz)—at each electrode site. In addition, power ratios such as theta/beta and delta/beta were computed to capture frequency balance indicators. To assess cross-frequency dynamics, phase-amplitude coupling (PAC) metrics were calculated using the modulation index, which quantifies the degree to which the phase of a lower-frequency oscillation modulates the amplitude of a higher-frequency signal [15]. All qEEG analyses were performed using Python-based MNE-Python software, and PAC computations were conducted using the Tensorpac toolbox [16].
Clinical outcome measures
Treatment response was evaluated using the change in scores between admission and discharge on the following validated clinical instruments: the Health of the Nation Outcome Scales (HoNOS) [17], Clinical Global Impression-Severity (CGI-S) [18], the World Health Organization Disability Assessment Schedule (WHODAS) [19], and the Global Assessment of Functioning (GAF) [20]. Additional clinical data collected included psychiatric diagnoses, age of onset, history of suicide attempts, handedness, education level, and presence of physical illness. Medication information was documented and converted into fluoxetine-equivalent doses for antidepressants and chlorpromazine-equivalent doses for antipsychotics, based on published equivalency guidelines. These standardized dosages were included as covariates in subsequent predictive models.
Machine learning analysis
To evaluate the predictive capacity of qEEG features for treatment response, we employed Extreme Gradient Boosting (XGBoost) regression models. Due to the limited sample size, we augmented the dataset by segmenting each subject’s EEG recordings into overlapping 30-second epochs, moving in 15-second steps, across the three EEG conditions. This approach was adopted primarily to increase the number of training samples for machine-learning analysis, thereby enhancing model stability and allowing for better capture of intra-individual variability. For each 30-second segment, qEEG features were extracted across 19 channels, yielding a total of 570 features per segment. In addition, four medication-related features—pre-and post-admission equivalent doses of antidepressants and antipsychotics—were included to account for potential pharmacological effects.
Prior to modeling, all features were normalized using minmax scaling. The dataset was split into training (80%) and test (20%) sets, ensuring that segments from each subject were confined to either set to prevent data leakage. For cross-validation, Leave-One-Subject-Out (LOSO) validation was performed, a robust method commonly used in biomedical signal analysis to assess generalization across individuals.
Model training was conducted using the XGBoost algorithm, a scalable tree-boosting framework known for its efficiency and high predictive accuracy [21]. Hyperparameter optimization was carried out using five-fold cross-validation with the macro-averaged F1 score as the objective function to balance precision and recall across classes. Prediction performance was evaluated by computing the Pearson correlation coefficient (r) between actual and predicted clinical outcome scores. To improve model interpretability, SHapley Additive exPlanations (SHAP) analysis was performed [22]. SHAP values were used to quantify the contribution of individual features to model output, and the top 20 most influential features were visualized to facilitate neurophysiological interpretation.
RESULTS
We developed XGBoost-based regression models to predict changes in four clinical outcomes during hospitalization: HoNOS, CGI-S, WHODAS, and GAF. Pre-admission scores for each outcome were included as features to improve model performance. Across all models, predictive accuracy was high, with strong coefficients of determination (R²) and low mean squared error (MSE) values (Table 1).
Predictive model performance summary for clinical outcome measures
HoNOS change prediction
The model predicting change in HoNOS scores (pre- minus post-admission) achieved an MSE of 0.32, R² of 0.99, and r=0.99. SHAP analysis indicated that the baseline HoNOS score was the most influential predictor. Among qEEG features, lower delta-to-beta ratio at Fz (Fz-ratio-delta-beta), higher relative high beta power at O1 (O1-rel-highbeta), and higher absolute high beta power at Cz (Cz-abs-highbeta) were associated with poorer response. In contrast, higher relative delta power at Fp1 (Fp1-rel-delta) was linked to greater symptom improvement (Figure 1A).
SHAP summary and mean SHAP value plots of EEG features predicting changes in clinical outcomes. A: HoNOS. B: CGI-S. C: WHODAS. D: GAF. Each panel displays SHAP summary plots for the top 20 EEG features contributing to the respective prediction model. Each dot represents a 30-second EEG segment. The color gradient indicates the feature value (red=high, blue=low), while the SHAP value (x-axis) reflects the feature’s impact on the predicted outcome. Positive SHAP values indicate that the feature increases the predicted clinical improvement, whereas negative values indicate a decrease in predicted improvement. Features are ranked by their average absolute SHAP values (mean impact on model output magnitude). SHAP, SHapley Additive exPlanations; EEG, electroencephalography; HoNOS, Health of the Nation Outcome Scales; CGI-S, Clinical Global Impression-Severity; WHODAS, World Health Organization Disability Assessment Schedule; GAF, Global Assessment of Functioning.
CGI-S change prediction
The CGI-S model showed an MSE of 0.02, R² of 0.96, and r=0.98. The pre-admission CGI-S score was the strongest predictor. Among qEEG features, lower absolute gamma power at T7 (T7-abs-gamma) predicted poorer response, while higher relative slow alpha power at O2 (O2-rel-slowalpha) and higher absolute theta power at O2 (O2-abs-theta) were associated with greater improvement (Figure 1B).
WHODAS change prediction
The model for WHODAS improvement yielded an MSE of 1.32, R² of 0.98, and r=0.99. Key predictors included the baseline WHODAS score, higher absolute slow alpha power at C4 (C4-abs-slowalpha), and higher absolute alpha power at Fp1 (Fp1-abs-alpha). Better outcomes were also associated with lower absolute low beta power at Fz (Fz-abs-lowbeta) and a lower theta-to-alpha ratio at T7 (T7-ratio-theta-alpha) (Figure 1C).
GAF change prediction
The GAF model achieved an MSE of 0.76, R² of 0.97, and r=0.98. Important predictors included the baseline GAF score and relative theta power at T8 (T8-rel-theta). Improved functioning was further associated with lower relative low beta power at O1 (O1-rel-lowbeta) and C3 (C3-rel-lowbeta), and higher absolute gamma power at O1 (O1-abs-gamma) (Figure 1D).
DISCUSSION
This study investigated the potential of qEEG features to predict treatment response in female adolescents with repetitive NSSI using a machine learning-based approach. By utilizing multiple clinical outcome measures—namely the Ho-NOS, CGI-S, WHODAS, and GAF—we investigated how EEG-derived electrophysiological signals relate to treatment-related symptom changes during inpatient care.
Previous studies have reported atypical cortical activity among adolescents engaging in NSSI, including increased gamma activity in frontal and temporal regions, and asymmetric alpha power, which have been linked to emotional dysregulation and impulsivity [9,10]. In particular, Liu et al. [11] identified greater prefrontal P300 amplitudes as predictors of subsequent reductions in NSSI among adolescents with major depressive disorder, suggesting a role for attention-related neural mechanisms in treatment responsiveness.
Consistent with previous findings, our results demonstrated that in addition to baseline clinical scores, specific qEEG features—namely Fz-ratio-delta-beta, O1-rel-highbeta, and T7-abs-gamma—significantly contributed to treatment outcome prediction. These features may reflect neurophysiological dysfunctions relevant to NSSI. Fz-ratio-delta-beta, indicative of frontal cross-frequency coupling, has been associated with impaired cognitive control and emotion regulation [23]. O1-relhighbeta suggests heightened occipital arousal, often linked to internalizing symptoms [24], while T7-abs-gamma may reflect disruptions in emotional memory and social-affective processing in the left temporal lobe [25]. Such region- and frequency-specific alterations offer additional explanatory value beyond traditional symptom-based assessment and may help identify individuals less responsive to conventional treatments. Given their non-invasive and scalable nature, EEG biomarkers also hold promise for early intervention and longitudinal monitoring in clinical settings.
Unlike previous qEEG studies, our approach segmented EEG recordings into 30-second epochs, effectively increasing training data samples and enhancing model generalizability—an important advantage for potential clinical application of machine learning in psychiatry. Meanwhile, functional neuroimaging studies have shown that neurobiological markers can predict treatment response. For example, Huang et al. [26] demonstrated that in adolescents with depression and NSSI, altered resting-state functional connectivity within frontolimbic circuits—and its association with elevated pro-inflammatory markers—predicted self-injury severity and treatment trajectories. Our findings extend this line of research into electrophysiology, underscoring that qEEG-derived features may serve as non-invasive biomarkers for personalized intervention strategies in youth with NSSI.
Importantly, we also found that pre-treatment symptom severity—particularly admission scores on the respective clinical scales—was the most influential predictor of outcome across all models. This supports longstanding clinical observations that initial symptom burden may serve as a proxy for both illness severity and the capacity for improvement [27]. Prior studies in adolescent depression and psychotherapy have similarly shown that higher baseline severity often predicts greater absolute symptom reduction [28]. While baseline symptom severity was consistently the strongest predictor, the inclusion of qEEG features provided incremental predictive value across clinical domains. In our SHAP analysis, several neurophysiological features ranked immediately after baseline scores in importance, suggesting that they capture unique aspects of brain function not fully reflected by clinical ratings. Furthermore, supplementary models excluding qEEG features showed slight but consistent reductions in predictive accuracy, highlighting the additive value of electrophysiological data in individualized prediction models.
However, several limitations should be considered when interpreting these findings. First, the sample was restricted to female adolescents recruited from a single clinical center, which may limit the generalizability of the results. In particular, prior studies suggest that sex-specific neurophysiological patterns may influence NSSI-related behaviors. For instance, male adolescents with NSSI have been reported to show higher impulsivity and externalizing symptoms, which may be underpinned by distinct EEG markers such as reduced frontal theta activity or altered sensorimotor rhythms [29]. These findings highlight the importance of future studies including both sexes to clarify potential sex-based neural mechanisms of NSSI and to enhance the generalizability of EEG-based predictive models. Second, the absence of potentially relevant clinical variables—such as the frequency of NSSI episodes, presence of suicidal ideation, and co-occurring psychiatric diagnoses—may have influenced treatment outcomes. NSSI is also characterized by substantial heterogeneity in both clinical presentation and etiological mechanisms, which may not be fully captured by our current analytic framework [30]. Additionally, the retrospective design of the study limits causal inference and restricts the temporal interpretation of neurophysiological changes. The EEG data were based on a single resting-state recording at baseline, precluding the evaluation of dynamic changes in brain activity over the course of treatment. Future prospective studies employing repeated qEEG measurements would help validate the temporal stability of EEG-derived predictors and clarify the neurophysiological trajectories underlying treatment response. Finally, statistical considerations should also be acknowledged. Although our outcome measures were continuous rather than binary classification labels, individual differences in symptom change may have resulted in uneven response distributions. While the LOSO cross-validation strategy helps mitigate overfitting, potential imbalances in response variability could still introduce bias into model training. Larger and more stratified cohorts will be essential for confirming the generalizability and robustness of these predictive models.
In conclusion, this study is among the first to apply machine learning algorithms to predict treatment outcomes using qEEG features in female adolescents with repetitive NSSI. By leveraging both neurophysiological and clinical data, we demonstrated that baseline clinical severity—as well as specific EEG features such as Fz-ratio-delta-beta, O1-rel-highbeta, and T7-abs-gamma—significantly contributed to the prediction of treatment response across multiple clinical domains. These findings underscore the potential clinical utility of EEG-derived functional brain markers in guiding individualized treatment strategies for adolescents at high risk of NSSI.
Supplementary Materials
The Supplement is available with this article at https://doi.org/10.30773/pi.2025.0256.
Demographic and clinical characteristics of participants
Notes
Availability of Data and Material
The datasets generated or analyzed during the study are available from the corresponding author on reasonable request.
Conflicts of Interest
The authors have no potential conflicts of interest to disclose.
Author Contributions
Conceptualization: Jun Won Kim. Data Curation: Seng Yoon Kim, Jun Won Kim. Formal analysis: Jun Won Kim. Funding acquisition: Jun Won Kim. Investigation: Seng Yoon Kim, Jun Won Kim. Methodology: Jun Won Kim. Project administration: Seng Yoon Kim, Jun Won Kim. Supervision: Jun Won Kim. Visualization: Jun Won Kim. Writing—original draft: Seng Yoon Kim, Jun Won Kim. Writing—review & editing: Jun Won Kim.
Funding Statement
This work was supported by the grant of Research Institute of Medical Science, Daegu Catholic University (2024).
Acknowledgments
None