AI Models for Depressive Disorder Detection and Diagnosis: A Review

1. Problem Statement

Major Depressive Disorder (MDD) is a leading cause of global disability, yet its diagnosis remains heavily reliant on subjective clinical assessments, which can be inconsistent and delayed. While Artificial Intelligence (AI) offers the potential for objective, scalable, and timely diagnostic tools, the field suffers from fragmentation. Existing literature reviews often focus on single data modalities (e.g., only text or only EEG), specific model families, or lack a structured framework for comparative analysis. This makes it difficult for researchers to navigate the landscape, compare methodologies, and identify gaps in areas such as multimodal fusion, explainability, and algorithmic fairness.

2. Methodology

The authors conducted a systematic review of 55 key studies published between 2021 and 2025. The methodology involved:

• Search Strategy: Queries on Google Scholar using combinations of keywords (e.g., "depression detection," "MDD," "GNN," "LLM," "AI").
• Selection Criteria: A 10-point scoring system evaluated papers on Relevance, Venue Quality, and Recency. Papers scoring below 5 on any criterion were excluded.
• Taxonomy Construction: The authors developed a novel hierarchical taxonomy to categorize the selected literature across three distinct levels:
  1. Clinical Task: Diagnosis (identifying current state) vs. Prediction (forecasting future onset/risk).
  2. Data Modality: Text, Voice, Neuroimaging (EEG, fMRI, fNIRS), and Multimodal.
  3. Computational Model: Traditional Machine Learning, Deep Learning (CNN, RNN), Hybrid Models, Large Language Models (LLMs/Transformers), and Graph Neural Networks (GNNs).

3. Key Contributions

The paper makes four primary contributions to the field of computational psychiatry:

1. Novel Hierarchical Taxonomy: A structured, three-level framework that organizes the field by task, data, and model, enabling rapid navigation and structural understanding of the research landscape.
2. Comprehensive Structured Review: An in-depth analysis of state-of-the-art methods within each sub-domain, accompanied by synthesis paragraphs that extract methodological patterns, contradictions, and areas of consensus.
3. Practical Resource Guide: A consolidated guide to prominent public datasets (DAIC-WOZ, MODMA, RSDD) and standard evaluation metrics (Accuracy, F1, AUC-ROC, MAE, RMSE, CCC), including mathematical definitions.
4. Future Roadmap: Identification of emerging trends and open challenges, specifically the shift toward causal inference, the need for improved interpretability, privacy-preserving learning (e.g., federated learning), and cross-cultural generalization.

4. Key Results and Findings

The review reveals distinct trends and performance characteristics across different modalities and models:

A. By Clinical Task

• Diagnosis: The dominant task in the literature. Models focus on classifying current depressive states or assessing severity (e.g., PHQ-9 scores).
• Prediction: A significantly smaller but critical niche focused on forecasting onset or treatment response. This area suffers from a lack of longitudinal datasets.

B. By Data Modality and Model Performance

• Text-Based Analysis:
  • Trend: Evolution from Traditional ML (SVM, Naive Bayes) to Deep Learning (LSTM, CNN) and currently dominated by Large Language Models (LLMs) and Transformers (BERT, GPT).
  • Insight: Hybrid architectures (e.g., FCL, DECEN) combining local and global feature extraction often outperform single models. There is a strong consensus on using clinical grounding (e.g., PHQ-9 alignment) to improve interpretability.
  • Challenge: Generalizability across platforms (Reddit vs. Twitter vs. Clinical Interviews) remains an issue.
• Voice-Based Analysis:
  • Trend: Surprisingly sparse literature. The primary finding is the high efficacy of transfer learning using pre-trained speech models like wav2vec 2.0.
  • Result: Achieved high accuracy (up to 96.49%) on the DAIC-WOZ dataset, suggesting voice is a promising but under-explored modality.
• Neuroimaging (EEG/fMRI/fNIRS):
  • Trend: Graph Neural Networks (GNNs) are the dominant architecture, naturally modeling brain connectivity as graphs.
  • Insight: There is a shift from static functional connectivity to dynamic or hybrid static-dynamic representations, which generally yield higher accuracy.
  • Key Models: BrainIB (Information Bottleneck), DSFGNN (Dynamic-Static Fusion), and ensemble GNNs.
  • Biomarkers: Consistent identification of temporal-parietal and frontal regions as key biomarkers.
• Multimodal Fusion:
  • Trend: The fastest-growing area. Combining text, audio, and physiological signals consistently outperforms unimodal approaches.
  • Architecture: GNNs and Transformers serve as the primary fusion backbones. Intermediate fusion (e.g., cross-attention) generally performs better than early or late fusion.
  • Fairness: Emerging focus on uncertainty-based multitask learning (e.g., U-Fair) to mitigate demographic biases.

C. Evaluation Metrics

The paper standardizes the evaluation landscape:

• Classification: Accuracy, Precision, Recall, F1-Score, Specificity, and AUC-ROC.
• Regression (Severity Scoring): Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Concordance Correlation Coefficient (CCC).

5. Significance and Future Directions

This survey serves as a critical roadmap for the next generation of AI in mental health. Its significance lies in:

• Standardization: Providing a unified taxonomy and metric guide to reduce fragmentation in the field.
• Clinical Trust: Highlighting the shift from "black-box" models to explainable AI (e.g., Retrieval-Augmented Generation, symptom capsules) which is essential for clinical adoption.
• Addressing Bias: Emphasizing the need for fairness-aware models and diverse, multilingual datasets to ensure generalizability across cultures.

Future Research Directions Identified:

1. Causal Inference: Moving beyond correlation to understand causal links between modalities (e.g., how brain connectivity changes causally influence speech patterns).
2. Longitudinal Prediction: Shifting focus from static diagnosis to predicting disease trajectory and treatment response over time.
3. Privacy-Preserving Learning: Utilizing federated learning to train models on decentralized, sensitive patient data without compromising privacy.
4. Cross-Cultural Generalization: Developing models that account for linguistic and cultural variations in symptom expression.

In conclusion, the paper asserts that while AI has made significant strides in depression detection, particularly through GNNs and LLMs, the field must now prioritize interpretability, fairness, and causal reasoning to transition from research prototypes to reliable clinical tools.