AI-powered wearables show promise in mental health monitoring

The reviewed studies examined the use of wearable devices - including smartwatches, fitness bands, and sensor-equipped clothing - to passively collect physiological and behavioral data. These biosignals, such as heart rate variability (HRV), galvanic skin response (GSR), electrodermal activity (EDA), and activity data from accelerometers, are then analyzed using machine learning models to assess mental states including stress, anxiety, and depression.

A systematic review of 48 studies has found that artificial intelligence integrated with wearable biosensors can detect mental health conditions such as stress, anxiety, and depression with high accuracy, but faces major barriers to real-world implementation. Published this week in the journal Biosensors, the review titled "Fusing Wearable Biosensors with Artificial Intelligence for Mental Health Monitoring: A Systematic Review" highlights rapid growth in the field, while warning that small sample sizes, lack of clinical integration, and limited real-world testing threaten its scalability.

The findings arrive as mental health systems globally face unprecedented demand. With an estimated economic burden of over $16 trillion projected by 2030, scalable and remote mental health tools are increasingly seen as essential.

Wearable biosensors take center stage in AI monitoring systems

The reviewed studies examined the use of wearable devices - including smartwatches, fitness bands, and sensor-equipped clothing - to passively collect physiological and behavioral data. These biosignals, such as heart rate variability (HRV), galvanic skin response (GSR), electrodermal activity (EDA), and activity data from accelerometers, are then analyzed using machine learning models to assess mental states including stress, anxiety, and depression.

The review found that stress was the most frequently studied condition, appearing in 60% of the studies. Depression and anxiety followed at 31% and 9%, respectively. Researchers observed a steady rise in publications after 2019, coinciding with the pandemic-driven surge in demand for mental health technologies.

Unlike traditional mental health assessments, which rely on self-reports or clinical evaluations, AI-powered biosensor systems promise real-time monitoring with minimal user input. Several studies reviewed in the paper demonstrated that machine learning algorithms could detect elevated stress levels or depressive symptoms using only physiological markers and behavioral signals.

Among the most frequently applied algorithms were convolutional neural networks (CNNs), support vector machines (SVMs), and ensemble methods such as random forests. Accuracy rates in lab settings reached up to 97% in some studies, although performance often declined in real-world environments due to noise, variability, and smaller training datasets.

Ecological validity and sample size emerge as key challenges

Despite the promise of wearable-based mental health monitoring, the review flagged several limitations that hinder generalizability. Many studies used small, homogenous samples, often restricted to university students or short-term lab participants. This raises concerns about the reliability of the results across diverse populations.

Moreover, a lack of ecological validity - the ability to translate lab findings to real-world settings - was a recurring problem. Only a minority of studies implemented monitoring “in the wild,” where uncontrolled environmental factors and inconsistent sensor compliance can compromise data quality.

Battery life, data loss from connectivity issues, and the absence of standardized benchmarks for stress or emotional states further complicate large-scale deployment.

The review also noted that while detection of mental health signals has improved, few studies focused on integrating these systems into clinical care pathways. Most models focused on classification or prediction tasks, but there was limited discussion on how this data could support interventions, therapy, or medication management.

Researchers called for a shift toward personalized models that adapt to individual baselines and provide actionable insights for both patients and care providers. Some promising examples included systems that combined wearable data with ecological momentary assessments (EMAs), real-time prompts asking users to rate their emotional state or environment.

Opportunities for advancement

The review identified several avenues for future research and development:

• Multimodal sensor fusion: Combining heart rate, skin conductivity, voice, and movement data can increase accuracy and robustness.
• Explainable AI models: Enhancing interpretability to increase clinician and patient trust in predictions.
• Standardized data protocols: Establishing common benchmarks and open datasets to improve reproducibility.
• Longitudinal studies: Tracking mental health over weeks or months to account for fluctuations and individual differences.

The paper also emphasized the need for privacy-preserving machine learning techniques, as the sensitive nature of mental health data raises ethical concerns around data ownership and algorithmic bias.

Regulatory oversight and evidence-based validation will be essential to ensure that these technologies do not outpace clinical safeguards. The authors conclude that continued interdisciplinary collaboration is needed to turn biosensor-based AI tools from experimental systems into mainstream clinical infrastructure.