Advancing mental health diagnostics: a review on the role of smartphones, wearable devices, and artificial intelligence in depression and anxiety detection

Abstract

The integration of smartphones, wearable devices, and artificial intelligence (AI) has revolutionized mental health diagnostics, particularly for depression and anxiety, by enabling real-time data collection and early intervention. This review synthesizes the findings from recent studies on the use of these technologies for diagnostic precision and predictive modeling. Following the for Systematic Reviews and Preferred Reporting Items Meta-Analyses guidelines, a systematic search of PubMed, Scopus, and Web of Science was conducted for publications up to April 2025, resulting in the inclusion of 62 relevant studies. Our critical analysis revealed that, while artificial intelligence demonstrates high accuracy in detecting mental health symptoms, its performance is highly context-dependent. We examined significant challenges, including the lack of generalizability owing to disparate datasets, the critical yet often unstandardized role of feature engineering, and the “black box” nature of complex algorithms that hinder clinical trust. Addressing these limitations requires interdisciplinary collaboration, robust ethical and regulatory frameworks (e.g., GDPR and HIPAA), and scalable interpretable solutions. Future research must prioritize long-term validation, inclusivity across diverse populations, and development of explainable AI to bridge the gap between technological potential and clinical reality.

Introduction

The mental healthcare landscape is undergoing a transformative shift with advancements in digital technologies and artificial intelligence (AI), paving the way for digital mental health [1]. Mental health is vital for individual well-being and socioeconomic progress [2]. AI and machine learning (ML) enhance the accessibility, personalization, and effectiveness of mental health interventions [3–6]. Smartphones, wearable devices, and ML algorithms provide real-time objective insights into patient behavior, physiology, and interactions, enabling early intervention and personalized care [4, 7, 8]. Unlike traditional subjective diagnostic methods, digital tools analyze large-scale data to identify patterns linked to mental health conditions such as depression and anxiety [7, 9]. The increase in mental health conditions exacerbated by the COVID-19 pandemic has emphasized the need for effective diagnostic tools and treatments. This review examined digital technologies for diagnosing depression and anxiety, focusing on their benefits, limitations, and future research directions [1, 10–12].

The role of machine learning in mental health

Machine learning offers a data-driven approach that analyzes vast datasets to advance mental health diagnostics [5, 6, 13, 14]. Deep learning algorithms identify patterns and predict outcomes, aiding clinicians in making better decisions [6, 14]. These technologies redefine mental health diagnoses by considering an individual’s unique biopsychosocial profile [3, 4, 15].

Digital phenotyping, a key innovation, uses data from smartphones and wearable devices to monitor behavioral and physiological patterns [16, 17]. Smartphones passively collect contextual data through sensors, such as accelerometers, and microphones, such as location, activity, and usage. Wearables complement these devices by providing physiological data, such as heart rate and sleep patterns, contributing to a “digital phenotype”—a holistic representation of an individual’s mental state [10, 11, 18, 19]. Machine learning identifies correlations in these data streams and detects indicators of conditions, such as depression (e.g., reduced activity) and anxiety (e.g., altered sleep patterns) [4, 5, 9].

Advancements in mental health diagnostics

ML and big data analytics have enabled advanced risk models to identify individuals predisposed to mental illnesses [8, 14]. Techniques such as transfer learning address challenges such as limited data availability and enhanced model robustness [4, 5, 14, 16, 20–22]. Smartphones and wearables enable passive and active monitoring, bridging the gap between mental healthcare needs and limited resources, particularly in underserved regions [1, 13, 17].

Wearable devices further support personalized interventions, offering real-time feedback based on an individual’s state and context [10, 23, 24]. For example, heart rate variability and GPS data correlate with depression severity and social isolation, whereas sleep disturbances indicate early signs of mental disorders [2, 10, 24–28].

However, challenges remain, including data quality, algorithmic transparency, and ethical concerns such as privacy and bias. Despite this, digital tools provide nuanced, continuous monitoring of mental health, enabling personalized care and early intervention [5, 11, 18].

As AI-driven technologies advance, they promise to revolutionize mental health diagnostics and care, raise awareness, encourage help-seeking behaviors, and address global mental health challenges [5, 29–31].

Methodology

A systematic literature review was conducted to investigate the roles of smartphones, wearable devices, and machine learning in the diagnosis of depression and anxiety. The review was structured in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines.

• Search strategy and time frame: We performed a comprehensive search of PubMed, Scopus, and Web of Science electronic databases. We conducted a search for all relevant articles published up to April 2025. The search query combined keywords related to mental health (“mental health,” “depression,” “anxiety”), technology (“smartphones,” “wearable devices,” “digital phenotyping,” “mobile health”), and analytics (“machine learning,” “artificial intelligence,” “prediction” prediction).

• Inclusion and Exclusion criteria: Articles were selected based on the following predefined criteria.

  • The inclusion criteria were as follows: (1) study design: original empirical studies, systematic reviews, or meta-analyses; (2) focus on the use of smartphones, wearables, and/or ML algorithms for the detection, prediction, or management of depression and anxiety; (3) language: articles published in English.
  • The exclusion Criteria: We conference abstracts, dissertations, editorials, study protocols, and articles that did not specifically address the diagnostic applications of these technologies.

• Study selection: The initial search yielded a total of 845 records. After removing 231 duplicates, titles and abstracts of 614 articles were screened, resulting in the exclusion of 537 articles that did not meet the scope of this review. The full texts of the remaining 77 articles were assessed for eligibility, of which 15 were excluded. This resulted in a final set of 62 studies that were included in the qualitative synthesis. This process is described in detail in the PRISMA flowchart (Fig. 1).

Fig. 1.

Study PRISMA flowchart

Results

Search results

The initial systematic search across PubMed, Scopus, and Web of Science yielded 845 records. After the removal of 231 duplicates, the titles and abstracts of the remaining 614 articles were screened for relevance. This screening process led to the exclusion of 537 articles that did not meet the inclusion criteria. The full texts of the remaining 77 articles were then assessed for eligibility, from which a further 15 were excluded. Ultimately, a final set of 62 studies was included in the qualitative synthesis. This selection process is detailed in the PRISMA flowchart (Fig. 1).

Characteristics of the included studies

The included studies reflect a field experiencing rapid growth, with the majority of publications appearing between 2019 and 2024. Geographically, the research is predominantly concentrated in high-income countries, particularly the United States, with limited representation from low- and middle-income countries (LMICs), a gap noted in several reviews [32, 33].

The target populations varied significantly across studies, including university students [13, 34], clinically diagnosed patient cohorts [35, 36], and specific demographic groups such as adolescents [17, 37, 38]. Sample sizes were highly heterogeneous, ranging from small-scale exploratory studies with as few as 12 patients [35] to larger cohorts of over 200 participants [39, 40].

A scoping review by Abd-alrazaq et al. (2023) found that the primary application of AI in this domain is for diagnosis or screening (59% of studies), followed by monitoring of symptoms (22%) and prediction of future mental states (19%). Notably, there is a significant lack of studies focusing on AI-driven treatment interventions, with most research centered on assessment [16].

Features of wearable devices

Wearable devices were a central technology in a majority of the reviewed literature, with a clear preference for specific form factors and data types.

• Device type and placement: Consistent with broad market trends, wrist-worn devices are the most common form factor. A comprehensive scoping review found that smart bands and smartwatches were used in 83% of studies investigating wearable AI for depression and anxiety [16]. Commercial devices such as the Fitbit series and clinical-grade actigraphy watches like the Actiwatch AW4 were frequently cited [16, 41].

• Measured biosignals and sensors: The data collected from these devices primarily focus on physiological and behavioral proxies for mental states. The most commonly measured biosignals were physical activity (captured in 90% of studies), sleep metrics (77%), and heart rate data (46%) [16]. This corresponds directly to the underlying sensor technology, with accelerometers being the most prevalent sensor (91% of studies), used for tracking activity and sleep, followed by photoplethysmography (PPG) sensors for heart rate monitoring (45%) [16, 19, 42] .

Features of smartphones

Smartphones serve as powerful, ubiquitous sensors for capturing behavioral data, often referred to as digital phenotyping.

• Data sources: Studies predominantly utilize data from built-in sensors to infer behavior. Key data streams include:

  • Mobility patterns: GPS data is used to derive features like location variance, entropy, and time spent at home, which have been correlated with depressive symptoms [37, 43].
  • Social interaction: Metadata from call and SMS logs (e.g., frequency, duration, reciprocity) are used as proxies for social engagement [17, 36].
  • General usage patterns: Screen-on time, frequency of app launches, and patterns of app category usage are analyzed to detect changes in routine and engagement [13].

• Operating system and data access: A significant practical consideration is the mobile operating system. The majority of research applications are developed for Android, as it provides more open access to raw sensor and usage data compared to the more restrictive iOS platform [32, 38]. This reliance on a single OS has implications for the generalizability of findings to the entire smartphone-using population.

Features of AI algorithms

The analysis of the complex, multimodal data streams from these devices is driven by a range of AI and ML algorithms.

• Algorithm types: While over 50 different algorithms were identified across the literature, a clear trend towards certain classes of models is evident. Traditional ML algorithms such as Random Forest (used in 52% of studies) and Support Vector Machines (38%) are the most common, likely due to their robustness and relative interpretability [16]. More recently, gradient boosting models (e.g., LightGBM, XGBoost) have shown high performance, particularly in studies with structured, tabular data [7, 13]. Deep learning models, including Deep Neural Networks (DNNs), are increasingly being used to predict treatment response and diagnostic status from complex sensor data [36].

• Problem formulation: The overwhelming majority of studies (91%) frame the task as a classification problem (e.g., depressed vs. healthy) [16]. A smaller subset of studies uses regression to predict the severity of symptoms on a continuous scale, such as a PHQ-9 score [37, 44].

• Validation and ground truth: The most common validation method is k-fold cross-validation (48%), though a significant portion of studies use a simple hold-out set (36%) [16]. A critical limitation across the field is the lack of external validation on independent datasets. The “ground truth” used for model training is most often derived from self-report scales like the Montgomery-Asberg Depression Rating Scale (MADRS) or the Patient Health Questionnaire-9 (PHQ-9), rather than formal clinical diagnoses [16, 45] (Tables 1 and 2; Fig. 2).

Table 1.

Literature review summary

Author (Year)	Study type	Technique/Algorithm	Sample size	Duration/Data	Key findings (Quantitative)
Grunerbl et al. (2014) [38]	Empirical study	Machine learning	12 patients with bipolar disorder	Several weeks of phone sensor data	81% accuracy (state recognition).
Saeb et al. (2015) [46]	Exploratory	Logistic regression	28 participants	2 weeks of GPS & phone usage data	GPS data correlated with depressive symptoms. 64% sensitivity.
Fitzpatrick, Darcy, and Vierhile (2017) [47]	Randomized controlled trial (RCT)	CBT Chatbot (Woebot)	70 students	2 weeks of app usage	Significant reduction in depression symptoms (PHQ-9).
Wang et al. (2018) [37]	Longitudinal	Machine learning	83 students	9 weeks of Microsoft Band & phone sensor data	Tracked depression dynamics. AUC = 0.809.
Doryab et al. (2019) [48]	Empirical study	Machine learning	79 students	Several weeks of Fitbit & phone data	Identified loneliness with 80.2% accuracy.
Walch et al. (2019) [49]	Validation study	Gradient boosting	31 participants	Accelerometer & PPG data	69% accuracy (4-stage sleep classification).
Cao et al. (2020) [17]	Development & usability	Machine learning (SVR)	13 families	12 weeks of sensor data, self-reports, parental evaluations	Significantly improved prediction accuracy with parental input (90%).
Dogrucu et al. (2020) [43]	Empirical study	Machine learning (SVM, RF, etc.)	294 participants	14 days of multimodal data	58.8% accuracy (PHQ-9 cutoff = 10), 77.1% (PHQ-9 cutoff = 20).
Ware et al. (2020) [50]	Empirical study	NLP	Data from Reddit & Twitter	Social media text data	> 90% accuracy (classifying depressive language).
Altini and Kinnunen (2021) [7]	Validation study	LightGBM	106 participants, 440 nights	3444 h of PSG & Oura ring data	79% accuracy (4-stage sleep classification).
Chikersal et al. (2021) [42]	Longitudinal	Machine learning, feature selection	212 participants	> 100 days of Fitbit & phone data	Predicted depression with F1-score > 80%.
Jacobson et al. (2021) [51]	Longitudinal	Deep learning	133 participants	17–18 years of data	Predicted deterioration in anxiety symptoms with high accuracy.
Moshe et al. (2021) [52]	Empirical study	Elastic net regression	54 participants	8 weeks of phone & wearable data	Significantly predicted depressive symptoms (R2 = 0.24).
Rykov et al. (2021) [44]	Cross-sectional	Random forest	69 participants	Fitbit data & questionnaires	AUC = 0.82 (depression screening).
Mullick et al. (2022) [40]	Exploratory study	Multimodal machine learning	37 adolescents	Mobile & wearable sensor data	Predicted depression severity with RMSE of 4.31.
Ahmed and Ahmed (2023) [25]	Empirical study	LightGBM, Stacking model	100 students	7 days of app usage data	82.4% sensitivity (LGBM), 77.9% balanced accuracy (Stacking).
Kadirvelu et al. (2023) [41]	Pilot study	“Mindcraft” platform	39 adolescents	Mobile app data	High engagement rate (80%).
Kim et al. (2023) [39]	Empirical study	Deep learning (DNN)	24 adolescents with MDD, 10 controls	5 weeks of smartphone data	77% validation accuracy for diagnosis; 76% for treatment response.
Allayla and Ayvaz (2024) [21]	Empirical study	MLP, Big data	Data from Reddit & Twitter	Social media text data	93.47% accuracy (classifying suicidal ideation).
Haque et al. (2023) [53]	Empirical study	Machine learning	73 students	High-stress exam period	HRV data predicted anxiety spikes with 78% accuracy.

Table 2.

Timeline of AI-driven mental health tools

Period	Focus	Signals collected (Data)	ML/DL algorithms used	Detailed methods/Approach	Key studies
2014–2015	Feasibility & Mobile sensing	GPS location traces; Accelerometer (movement/activity); Call and SMS logs (frequency, timing); App usage duration; Screen on/off events	Logistic Regression, Random Forest, SVM	Feature engineering on raw mobile sensor data; Statistical correlation with self-reported mood/PHQ-9; Supervised machine learning; Cross-validation within small samples; Exploratory data analysis for feasibility assessment	Grunerbl, Saeb [38, 46]
2017–2019	Intervention & Wearable data	Heart rate, step count, sleep stages from wearables (Fitbit, Apple Watch, Garmin); Text logs from chatbot conversations; In-app behavioral data (message frequency, sentiment); Device usage patterns	Random Forest, Gradient Boosting, KNN, Basic LSTM (for time-series), Rule-based classifiers	AI conversational agents (chatbots) for intervention delivery; Predictive modeling using time-series wearable data; NLP on chatbot/text data; Multimodal data fusion; Randomized controlled trials for chatbot effectiveness	Fitzpatrick, Wang [37, 47]
2020–2022	Validation & Big data approaches	Multi-week or multi-site wearable data (heart rate, activity, sleep); Clinical gold-standard data (polysomnography, structured symptom scales); Electronic Health Records; Survey data at scale	Deep Learning: CNN, LSTM, Transformer (BERT, etc.), XGBoost, Random Forest, Logistic Regression	Deep learning on timeseries and text data; NLP for symptom extraction from clinical notes; Large-scale external validation; Benchmarking against clinical labels; Data harmonization from multiple centers	Ware, Altini, Jacobson, Mullick [7, 40, 50, 51]
2023-Present	Personalization & Real-time feedback	Real-time HRV from smartwatches; Continuous physiological data (EDA, skin temperature, SpO2); Longitudinal patient-reported outcomes (EMA via app); Passive smartphone and IoT data	Personalized/Adaptive ML, Online Learning, Deep LSTM, GRU, Transformer, Ensemble Models, Bayesian ML	Streaming analytics for real-time prediction; Individual-level ML models; Online learning algorithms; Predictive modeling of treatment response; Just-in-time adaptive interventions (JITAI) based on live data	Kim, Haque, Allayla and Ayvaz [21, 53, 54]

Fig. 2.

Taxonomy of technologies

Discussions

Principal findings

This review reveals a rapidly advancing but methodologically fragmented field at the intersection of AI and mental health diagnostics. Our synthesis of 62 studies identifies several key trends.

First, the primary application of these technologies is overwhelmingly focused on the diagnosis, screening, and monitoring of mental health conditions. A comprehensive scoping review by Abd-alrazaq et al. (2023) found that 59% of studies on wearable AI aimed at diagnosis or screening, with only 22% focused on monitoring [16]. Critically, a significant gap exists in the literature regarding AI-driven treatment interventions, with most research efforts concentrated on assessment rather than therapeutic application.

Second, the field is characterized by extreme heterogeneity in data sources and methodologies, making direct comparisons of model performance challenging. Studies leverage a wide array of data, from minimalist smartphone metadata and large-scale social media text to granular, multi-sensor data from clinically diagnosed patients [8, 13, 35]. This diversity, while innovative, means that a reported accuracy of 93% in one context is not comparable to 81% in another.

Third, a fundamental limitation is the inconsistency in defining the “ground truth” for model validation. The majority of studies do not use formal clinical diagnoses (e.g., via structured interviews like the SCID) as their benchmark. Instead, they rely on self-report questionnaire scores, most commonly the Patient Health Questionnaire-9 (PHQ-9). As demonstrated by Dogrucu et al. (2020), the performance of an algorithm can change dramatically based on the chosen PHQ-9 cutoff score, highlighting the instability of this validation approach [16, 40, 45].

Finally, there is a clear trade-off between model performance and interpretability. While traditional machine learning algorithms like Random Forest and Support Vector Machines remain the most common due to their robustness and relative transparency (Abd-alrazaq et al., 2023), there is a growing use of deep learning models. These “black box” architectures may offer marginal gains in accuracy but at the cost of explainability, a critical barrier to clinical trust and adoption [11, 14, 16, 45].

Research and practical implications

The findings of this review have significant implications for the future trajectory of research and the practical deployment of these technologies in clinical settings.

First, there is an urgent need for methodological standardization. To move from a collection of isolated successes to a body of cumulative science, the field requires standardized benchmarks and reporting guidelines, similar to the CONSORT-AI statement. This includes establishing common datasets for testing, clear definitions of clinical endpoints, and mandatory external validation of models to ensure their generalizability beyond the initial training data [3, 45, 50].

Second, future research must prioritize model robustness. The performance of algorithms on clean, curated datasets often does not translate to real-world “in the wild” scenarios, where data is noisy, incomplete, and subject to sensor error. Studies must explicitly test and report on the resilience of their models to such data perturbations to ensure they are reliable enough for clinical use [11, 12, 50].

Third, a shift towards human-centered and hybrid systems is crucial. The goal of AI should be to augment, not replace, human clinicians. This involves a co-design process where tools are developed in close collaboration with both clinicians and patients to ensure they are usable and meet real-world needs. Furthermore, models should be designed to integrate human intelligence. The work by Cao et al. (2020), which demonstrated that incorporating parental evaluations as a “human sensor” significantly improved prediction accuracy for adolescent depression, provides a powerful template for this hybrid approach [14, 16, 17].

Finally, the development of explainable AI (XAI) is not merely a technical exercise but a prerequisite for clinical adoption. For a clinician to trust and act upon an AI-generated recommendation, the model must be able to provide a clear rationale for its prediction. Moving from opaque “black box” models to interpretable “glass box” systems is essential for building trust, ensuring accountability, and debugging biased or flawed algorithms [14, 39, 52].

Ethical considerations

The deployment of AI in mental healthcare is not solely a technical challenge but is also laden with profound ethical imperatives. Mental health data is exceptionally sensitive, and the continuous, passive collection of behavioral and physiological information raises significant privacy concerns. Existing regulatory frameworks like HIPAA and GDPR were not designed for the era of continuous, user-generated data from consumer devices, creating a legal and ethical gray area that puts user privacy at risk [51, 53, 54].

Furthermore, algorithmic bias represents a systemic risk to health equity. AI models are a product of their training data; if this data is not representative of the broader population (e.g., being sourced primarily from a single demographic or socioeconomic group), the resulting model will be biased. This can lead to the amplification of existing health disparities, where models systematically fail to detect signs of distress in marginalized communities [21, 55, 56].

Finally, the opacity of many AI models creates an accountability vacuum. When a “black box” model makes an incorrect prediction that leads to a negative health outcome, it is unclear who is responsible: the developer, the clinician who used the tool, or the healthcare system that deployed it. This lack of transparency undermines clinician trust and complicates the process of obtaining truly informed patient consent, as the full implications of how one’s data will be used are difficult to convey [54, 57–59].

Limitations

This review has several limitations that should be considered. First, our search was restricted to articles published in English, which may have led to the omission of relevant studies published in other languages. Second, due to the heterogeneity in study populations, data sources, algorithms, and outcome measures, we were unable to perform a meta-analysis to statistically synthesize the results. The findings are therefore based on a narrative synthesis of the included literature. Third, this review focused primarily on the role of AI in the detection and diagnosis of depression and anxiety. A detailed analysis of AI-driven therapeutic interventions, such as chatbots or digital therapeutics, was beyond the scope of this paper. Finally, as with any literature review, there is a potential for publication bias, where studies with positive or statistically significant results are more likely to be published than those with null findings [47, 50, 58].

Conclusions

The integration of smartphones, wearable devices, and machine learning (ML) is transforming mental health diagnostics and treatment, offering new opportunities to enhance care [47]. These technologies enable continuous and objective monitoring of physiological and behavioral parameters, providing a deeper understanding of an individual’s mental state [16, 58, 59]. Real-time data collection and analysis facilitate the early detection of mental health conditions, such as depression and anxiety, supporting timely interventions [50, 53, 56] .

However, several challenges remain. Data collected from different devices and sensors are often heterogeneous, with varying levels of accuracy that may not meet professional standards [47, 48, 56]. The lack of standardized protocols for data collection and analysis further complicates reliability and validity. Additionally, some ML models lack interpretability, making it difficult for clinicians to understand how predictions are generated, which is crucial for informed decision making [25–27, 31].

Ethical considerations such as data privacy, algorithmic transparency, and biases also need to be addressed to ensure responsible implementation [51, 54]. Despite these challenges, digital interventions have already improved access to mental health care, particularly during crises such as the COVID-19 pandemic, when traditional services are overwhelmed [9, 33, 53].

With standardized protocols, interpretable models, and rigorous validation, these technologies have the potential to revolutionize mental health care and improve accessibility, early detection, and personalized treatment for individuals worldwide.

Declarations

Conflict of interest

The authors declare no conflicts of interest.

Ethical approval

Ethical approval was not required for this study, as it was a review paper.

Consent to participate

Not applicable as the study did not involve any human participants or personal data.

Consent for publication

Not applicable, as this study did not include any individual data or identifiable images.