Evaluating AI Models for Mental Health Monitoring and Diagnosis: A Comprehensive Systematic Review across Text, Speech and Physiological Signals

Abstract

The scale of the global mental health crisis — affecting roughly 970 million people worldwide stands in sharp con-trast to the tools we currently have for detecting and monitoring it. Questionnaires are subjective. Clinical interviews are re-sourceintensive. And in much of the world, there simply are not enough trained professionals to meet demand. Over the past decade, artificial intelligence has been proposed as a way to address some of these gaps, and a substantial body of research has accumulated around three main data sources: physiological signals, text, and speech. This paper reports findings from a PRISMA-guided systematic review in which we searched eight major databases including PubMed, IEEE Xplore, ACM Dig-ital Library, and ScienceDirect for peer-reviewed studies published between 2015 and 2025. After two independent review-ers screened 120 records, 63 met our inclusion criteria. We grouped these into four modality categories: physiological signal studies (EEG, HRV, ECG, electrodermal activity), NLP-based text analysis, speech and acoustic analysis, and multimodal fusion approaches. EEG-based deep learning models reported the highest accuracy figures sometimes above 97%, though we argue these numbers are difficult to interpret without knowing the sample size and experimental conditions behind them, which in most cases were quite limited. Transformer models such as BERT and RoBERTa did well on depression and suici-dality detection tasks, speech models showed useful diagnostic signals via acoustic and prosodic features, and multimodal systems outperformed single-source approaches across the board. What the review also surfaces, though, is a consistent set of problems: small and unrepresentative samples, almost no external validation, very little attention to fairness or interpretabili-ty, and reporting practices that make replication difficult. These gaps matter. They are what currently stands between techni-cally interesting research and tools that could actually be used in clinical practice.