This systematic review examines the current landscape of artificial intelligence (AI)-based Clinical Decision Support Systems (CDSS) designed to aid in the treatment of depression. With depression recognized as a complex and multifactorial condition, often differing across contexts such as postpartum, adolescence, and elderly populations, the review aims to evaluate the feasibility, acceptability, and potential of AI-based CDSS in supporting personalized care. The primary research question explores how these systems are implemented and evaluated in clinical settings, as well as the main challenges to their broader adoption.

The review analyzed studies focused on AI-based CDSS for depression, identifying variations in algorithms, performance metrics, and implementations. Studies were assessed based on system purposes, categorized into treatment selection, prediction and risk assessment, and clinical support and data review, providing a comprehensive overview of each study's approach and outcomes.

11 unique algorithms, 9 performance metrics, and 10 CDSS implementations were identified across the studies. Findings indicate positive results for CDSS in supporting clinical decision-making, yet highlight challenges in applying these tools in routine practice. Limitations include small, non-diverse sample sizes, immature methodologies, and difficulties in workflow integration, along with concerns about data privacy and the ethical use of AI.

AI-based CDSS for depression shows potential in patient care; however, foundational issues such as more diverse samples and robust model development remain critical. Addressing these aspects will set the stage for future research into advanced applications, such as virtual-reality integration and multi-condition prediction, thus advancing AI's role in mental health treatment.Clinical Relevance-This work illustrates the state of art of AI adoption in the domain of mental health, specifically in the context of depression.

This study classifies and estimates the intensity of multiple emotional states using physiological signals. We employed a jigsaw puzzle task to elicit both positive and negative emotions in participants. Mood states were assessed using the profile of mood states 2nd Edition (POMS2), while electroencephalogram (EEG) and heart rate variability (HRV) signals were recorded simultaneously. Support vector machines (SVMs) were used for emotion classification. Feature extraction techniques were applied to enhance classification accuracy, including principal component analysis (PCA) and autoencoders (AE). Recursive feature elimination (RFE) was utilized to identify key physiological indicators. When PCA or AE preprocessing was applied, the classification model achieved a κ coefficient of over 0.9 for all emotions. The key features for emotion classification were identified as mean RR interval (MRRI), low-frequency power (LF), high-frequency power (HF), ratio, and prefrontal alpha asymmetry (Fp1α-Fp2α), whereas HF, standard deviation of RR intervals, LF, and F7α-F8α showed lower importance. The findings suggest that EEG and HRV signals can classify and estimate multiple emotional states simultaneously. These results contribute to developing objective emotion recognition systems for applications in mental health monitoring and affective computing.Clinical Relevance- Accurately assessing emotional states is crucial for mental health care, stress management, and affective computing applications. The proposed emotion classification model utilizing EEG and HRV signals provides an objective and quantitative approach to evaluating mood states. This study demonstrates the feasibility of non-invasive physiological monitoring for mental well-being assessment, offering potential applications in workplace stress management, early detection of mood disorders, and human-computer interaction systems.

Depression is one of the leading causes of disability worldwide, posing a severe burden on individuals, healthcare systems, and society at large. Recent advancements in Large Language Models (LLMs) have shown promise in addressing mental health challenges, including the detection of depression through text-based analysis. However, current LLM-based methods often struggle with nuanced symptom identification and lack a transparent, step-by-step reasoning process, making it difficult to accurately classify and explain mental health conditions. To address these challenges, we propose a Chain-of-Thought Prompting approach that enhances both the performance and interpretability of LLM-based depression detection. Our method breaks down the detection process into four stages: (1) sentiment analysis, (2) binary depression classification, (3) identification of underlying causes, and (4) assessment of severity. By guiding the model through these structured reasoning steps, we improve interpretability and reduce the risk of overlooking subtle clinical indicators. We validate our method on the E-DAIC dataset, where we test multiple state-of-the-art large language models. Experimental results indicate that our Chain-of-Thought Prompting technique yields superior performance in both classification accuracy and the granularity of diagnostic insights, compared to baseline approaches.

Depressive disorder is one of the most common mental health conditions worldwide, with a high risk of suicide and a significant likelihood of becoming chronic. Currently, the diagnosis of depressive disorder relies on clinical interviews and self-report questionnaires, highlighting the need for an objective, digital biomarker-based time-series depression diagnosis model. This study aims to develop a deep learning-based multivariate time-series depression classification model using sleep data collected from wearable devices. The model architecture employs MLSTM-FCN, InceptionTime, and Time-series Transformer, extracting features from a total of ten sleep biomarker candidates to classify depression status. The performance of the models was evaluated, yielding AUC scores of 0.91, 0.82, and 0.78, respectively, with MLSTM-FCN demonstrating the highest performance. Among the sleep biomarker candidates, total time spent in bed, REM sleep latency, and light sleep duration were identified as significant indicators. The proposed model offers a cost-effective and objective method for depression diagnosis and is expected to be applicable to depression patients in community settings in the future.

Emotion recognition using eye movement data is a rapidly developing field with significant applications in affective computing and human-computer interaction. This study explores the potential of eye movement features alone for classifying emotions, leveraging the Deep Generalized Canonical Correlation Analysis with Attention Mechanism (DGCCA-AM) framework. Our experiments were conducted on the SEED-IV dataset, comprising eye movement recordings with 31 features from 15 subjects across three sessions, totaling approximately 37,600 samples. We evaluated the model in subject-dependent (intra-subject) as well as subject-independent (inter-subject) manner across all the three sessions. We obtained a high classification accuracy of 99.92% for distinguishing four emotions namely happy, sad, fear and neutral, in the third session; demonstrating the strong discriminative power of eye movement features in emotion classification. In inter-subject evaluation, we considered leave-one-subject-out strategy and evaluated the model's performance on all the three sessions of unseen subjects. Under this setting, the proposed approach provided an average accuracy of 63.14%, indicating challenges in generalization across different individuals. Further, we also analyzed the contribution of each of the three feature groups' subnetwork (namely pupil, event-statistics, and fixation/saccade/dispersion group) to the overall performance and found that pupil features are the most contributing feature groups, without which the model's accuracy drops by about 6%. Overall, these findings suggest that eye movement data alone is highly effective for within-subject emotion recognition but poses challenges in cross-subject generalization. Our study reinforces the importance of subject-specific modeling while opening new avenues for improving cross-subject adaptability in emotion recognition systems.Clinical relevance- Accurate emotion recognition is crucial in mental health assessment and human-computer interaction. This study leverages eye movement data and the DGCCA-AM framework to improve emotion classification without relying on EEG signals. The findings could aid clinicians in detecting emotional states for mental health diagnosis, enhancing interventions for conditions like anxiety, depression and affective disorders.

Depression is a debilitating condition, posing a significant challenge to public mental health globally. Despite advancements in diagnostic techniques, identifying depression remains challenging due to reliance on subjective assessments and diagnostic inaccuracies. These limitations underscore the need for exploring neural biomarkers to enhance diagnosis and treatment strategies. This research addresses the issue by investigating the effective connectivity (EC) patterns in the resting-state Electroencephalography (rsEEG) signals of individuals with Major Depressive Disorder (MDD) compared to healthy controls (HC). Existing studies primarily focus on functional connectivity (FC), often neglecting the causal interactions be-tween brain regions. To overcome this limitation, we apply the Frequency-Domain Convergent Cross Mapping (FD-CCM) technique, a model-free, nonlinear method capable of capturing EC patterns in the frequency domain. The study reveals that MDD subjects exhibit reduced EC across four major brain regions-frontal, parietal, temporal, and occipital-compared to HC participants. Notably, the findings indicate diminished frontal connectivity and altered power density in the delta and alpha frequency bands. Classification results demonstrate that FD-CCM features consistently outperform classical CCM across multiple classifiers, achieving an accuracy of 92.32% with the best-performing ANN classifier. These findings suggest that altered EC patterns are significant biomarkers for MDD, contributing to deficits in cognitive processing, emotional regulation, and sensory integration. The superior performance of the FD-CCM approach highlights its potential for clinical applications in mental health diagnostics.

Music therapy, especially Tibetan Singing Bowls (TSB) rooted in ancient practices, holds promise for enhancing mental well-being and autonomic regulation through their unique acoustic properties and therapeutic vibrations. Previous research indicated that TSB could influence physiological responses, including heart rate variability (HRV), a significant indicator of autonomic nervous system activity (ANS) and stress. This study aimed to study the impact of TSB on HRV. The Polar Ignite Series 1 wristwatch continuously measured Heart Rate (HR). Eight participants, aged 20-22, were recorded throughout a 10-minute relaxation period followed by a 30-minute TSB music meditation. Data was analyzed with the HRV Matlab toolbox. Results revealed a significant improvement in overall HRV and autonomic function (SDNN p = 0.02344). NN50 also increased significantly (p < 0.05), which suggests enhanced parasympathetic activity and relaxation response. These results highlight the potential of TSB music as a therapeutic intervention to improve stress generally and HRV in particular.Clinical Relevance- This study demonstrates the efficacy of Tibetan Singing Bowl music meditation in enhancing autonomic balance, evidenced by significant increases in SDNN and NN50. As a noninvasive, low-cost intervention, TSB therapy can be readily integrated into clinical stress-reduction programs, anxiety management strategies, and cardiac rehabilitation protocols to boost parasympathetic tone and improve heart rate variability, ultimately supporting patient resilience and cardiovascular health.

Speech-based AI models are emerging as powerful tools for detecting depression and the presence of Post-traumatic stress disorder (PTSD), offering a non-invasive and cost-effective way to assess mental health. However, these models often struggle with gender bias, which can lead to unfair and inaccurate predictions. In this study, our study addresses this issue by introducing a domain adversarial training approach that explicitly considers gender differences in speech-based depression and PTSD detection. Specifically, we treat different genders as distinct domains and integrate this information into a pretrained speech foundation model. We then validate its effectiveness on the E-DAIC dataset to assess its impact on performance. Experimental results show that our method notably improves detection performance, increasing the F1-score by up to 13.29 percentage points compared to the baseline. This highlights the importance of addressing demographic disparities in AI-driven mental health assessment.Clinical relevance Our proposed domain adversarial method can improve the fairness of AI models by mitigating gender bias, thereby improving their clinical applicability and trustworthiness in speech-based mental health detection.

High rates of alcohol-related problems have been reported among US service members (SMs). Screening questions on drinking and related behaviors can help identify individuals at-risk for alcohol-related problems. However, brief alcohol screeners, such as the alcohol use disorders identification test-consumption (AUDIT-C) and the 4-item rapid alcohol problems screening (RAPS-4), have not been adequately and concurrently validated among active-duty SMs.

From October to December 2021, 19,465 active-duty soldiers (including activated reserve soldiers) completed anonymous command-directed e-surveys (response rate= 31%); two random samples were drawn and sex-stratified. The AUDIT-C, RAPS-4, depression (PHQ2), anxiety (GAD2), and suicidal thoughts (2-item CSSRS) were analyzed to assess convergent validity and clinical utility of the AUDIT-C versus RAPS-4.

Findings indicate fair-to-moderate (φ = 0.310-0.399) convergence between screeners among males and weak-to-fair (φ = 0.227-0.391) convergence among female soldiers. Among male soldiers, the best level of agreement between screeners, albeit fair in concordance, was AUDIT-C ≥ 6 (weighted kappa = 0.381-0.399). Among female soldiers, AUDIT-C ≥ 4 or 5 demonstrated the best concordance with RAPS-4 (weighted kappa = 0.384-0.380, respectively). Importantly, however, less than one-third of soldiers screened positive by both AUDIT-C and RAPS-4; over two-thirds had discordant screening results. Although both screeners were independently and positively associated with risk for suicidal thoughts, depression, and/or anxiety, the RAPS-4 demonstrated stronger association with suicidal thoughts than AUDIT-C.

The AUDIT-C and RAPS-4 each capture unique but interrelated aspects of drinking behaviors. The RAPS-4 appears advantageous by including clinically oriented questions that have shown to strongly correlate with AUD risk, and in this study demonstrated strong correlations with risk for other mental health conditions. In contrast, the AUDIT-C is only limited to consumption-focused items. While the AUDIT-C is currently mandated primary alcohol screener in military settings, the stronger correlation of RAPS-4 with related behavioral health outcomes warrants further research and consideration as a preferable primary screener among SMs.