Artificial Intelligence in Psychiatry: A Review of Biological and Behavioral Data Analyses

The paper addresses how AI can enhance psychiatric diagnosis and treatment by providing objective, data-driven analyses in a field traditionally reliant on subjective assessments. It frames the research question around the extent to which modern AI methods—spanning EEG, ECG, speech/NLP, biomarkers, and social media—can improve diagnostic accuracy, prediction, and personalization in psychiatry. The introduction outlines the historical trajectory of AI in healthcare from statistical methods to machine learning and deep learning, noting psychiatry’s complexity and the need for tools that reduce diagnostic error and delay. It highlights a surge in publications since 2020 across major databases (Web of Science, PubMed, ScienceDirect, MDPI), indicating growing academic and clinical interest and significant research potential in AI for psychiatry.

The review situates AI in psychiatry within a rapidly expanding literature, especially post-2020, with machine learning being the most studied topic and notable growth in explainable AI by 2024. It synthesizes trends across EEG-based diagnostics (depression, schizophrenia), ECG/HRV for stress and emotion, speech/NLP for cognitive and affective assessment, biomarker-driven precision psychiatry, and social media/EHR mining for surveillance and risk prediction. It emphasizes methodological diversity (CNNs, ensemble learning, connectivity analysis, wavelet/scattering transforms, LLMs) and recurring issues: small sample sizes, data heterogeneity, and interpretability challenges.

This narrative review synthesizes diverse AI applications in psychiatry. Databases searched included Web of Science (SCIE), PubMed, and ScienceDirect. Keywords combined domain and modality terms: ‘artificial intelligence AND psychiatry’, ‘EEG AND AI’, ‘ECG AND mental health’, ‘NLP AND psychiatry’, and ‘social media AND mental health surveillance’. Inclusion criteria: peer-reviewed, English-language studies providing empirical data or substantive theoretical analysis of AI in psychiatric settings. Exclusions: non–peer-reviewed items (e.g., conference abstracts, opinion pieces) and studies not explicitly focusing on AI in psychiatric contexts. Extracted data encompassed authors, year, aims, AI technologies, primary outcomes, and implications for psychiatric practice. Articles were categorized by AI technology and its clinical application to enable thematic analysis of technological advances, challenges, and trajectories. Methodological rigor was assessed by soundness, robustness of findings, and publication outlet prestige. The approach ensured a comprehensive, methodologically grounded synthesis of current innovations and future directions for integrating AI into psychiatric practice.

Across modalities, AI demonstrates strong potential:

• EEG: CNNs and advanced feature methods achieved high accuracies. Example results include TRD vs. non-TRD classification at 90.05% (external validation 73.33%), sleep EEG analysis at 92.85% with data augmentation, MDD detection up to 98.8% accuracy (F1: 99.81%) using wavelet scattering and BiLSTM, and ADHD ensemble models at 97.4%. Functional connectivity and microstate analyses revealed condition-specific signatures but mixed findings (e.g., no PLV difference in MDD with vs. without self-harm). Common limitations: small samples, retrospective designs, and variable preprocessing.
• ECG/Physiological: End-to-end and transfer learning approaches achieved very high accuracies (e.g., 99.35% for mental state classification without preprocessing; mental fatigue detection at 98.44%). Multimodal fusion (EEG, ECG, ACC/GSR) reached 94.58–98.2% accuracy with high sensitivity/specificity. Inter-subject variability and dataset biases remain challenges.
• Speech/NLP: LLMs and feature-based models effectively assessed thought disorders and emotions. PPA classification reached 97.9%; schizophrenia thought disorder evaluation achieved F1 ≈ 92%; emotion recognition exceeded 98–99% on benchmark datasets; depression detection via hybrid features reached 94.63%. Loneliness detection in older adults achieved 88.9% accuracy using XAI. Limitations include dataset representativeness, reliance on short samples, and interpretability trade-offs.
• Blood biomarkers: ML models integrating immunologic, metabolic, genetic/epigenetic markers improved differentiation among psychiatric conditions and identified molecular subtypes, supporting precision psychiatry. Generalization requires large, diverse cohorts and external validation.
• Social media/EHR: NLP/LLMs enhanced risk prediction and information extraction (e.g., suicide attempt AUC up to 0.932; veterans’ suicide risk AUC improved by ~19%; violence classification precision/recall 89–98%). Findings show feasibility for real-time surveillance and clinical decision support, with caveats about data bias, privacy, and generalizability.

The findings support the central premise that AI can provide more objective, scalable, and precise assessments in psychiatry. EEG and ECG models capture neurophysiological and autonomic correlates of mental states, improving diagnosis and monitoring of depression, stress, and fatigue. Speech/NLP, including LLMs, objectively quantify thought disorder, emotion, and cognitive decline, enabling remote assessments. Biomarker-based AI advances the biological understanding and subtyping of psychiatric conditions, aligning with precision psychiatry. Social media/EHR analytics extend surveillance and risk prediction outside traditional clinical encounters. Collectively, these advances address the need for faster, more reliable, and personalized psychiatric care. However, clinical impact depends on overcoming interpretability barriers, standardizing preprocessing, expanding multimodal datasets, and ensuring ethical, privacy-preserving deployment. The synthesis highlights that multi-center, diverse cohorts and explainable models are key to translating promising accuracies into generalizable clinical tools.

AI is reshaping psychiatry by leveraging EEG, ECG, speech/NLP, biomarkers, and social media/EHR data to enhance diagnosis, monitoring, and personalized treatment. Notable successes include high-accuracy EEG/ECG classifiers, speech-based emotion and thought disorder detection, biomarker integration for precision psychiatry, and robust risk prediction from unstructured text. Persistent challenges—model interpretability, dataset bias, heterogeneity, and privacy—limit clinical adoption. Future work should expand diverse, high-quality datasets; develop explainable, validated models; and establish regulatory frameworks. Interdisciplinary collaboration among clinicians, data scientists, and policymakers will be critical to safely realize AI’s potential for precision psychiatry and improved mental health outcomes.

Key limitations include heterogeneous and often small datasets; retrospective or single-center designs; variability and lack of standardization in signal preprocessing (EEG/ECG) and feature extraction; limited demographic diversity and language/cultural generalizability; risk of overfitting and reduced external validity; interpretability concerns for deep learning models; privacy, security, and ethical issues in handling sensitive psychiatric and social media/EHR data; and challenges integrating AI tools into clinical workflows. Regulatory validation and multi-center prospective trials are needed to confirm efficacy and generalizability.