Background

Depression and anxiety are among the leading causes of disability worldwide. The Global Burden of Disease Study 2019 placed depressive and anxiety disorders among the top mental disorders for disability-adjusted life years [1]. The COVID-19 pandemic added an estimated 53.2 million cases of major depressive disorder and 76.2 million cases of anxiety disorders in 2020 [2]. The World Health Organization now estimates that about 322 million people live with depression, and that in high-income countries only about one third of them receive treatment [3]. In Saudi Arabia, the Saudi National Mental Health Survey found that about one in three adults had a lifetime mental disorder, and anxiety disorders were the most common group [4].

Diagnosis still depends on clinical interviews and self-report scales such as the PHQ-9 and GAD-7. These tools are valid, but they need clinical time and honest reporting. Many people do not seek help because of stigma. For this reason, researchers have tried to detect depression and anxiety from data that people produce in daily life, such as written text, speech, facial behavior, physiological signals and health records [5,6].

Early studies used classical machine learning with hand-crafted features, and later studies trained convolutional and recurrent networks from the start [7,8]. Both approaches face the same limit. Labelled mental health data are small, expensive to collect and hard to share. The DAIC-WOZ interview corpus, which is used in many studies, contains fewer than 200 participants [9]. The MODMA dataset contains (Electroencephalogram) EEG and speech from about 50 participants [10]. Deep networks trained on such data tend to over fit.

Transfer learning and its variants

Transfer learning uses knowledge learned on a large source task to help a smaller target task [11,12]. In this review, five forms are separated, because they differ in cost and in risk.

Feature extraction keeps the pre-trained network frozen and uses its outputs as input to a new classifier. A typical example is a BERT sentence embedding passed to a support vector machine, or an ImageNet ResNet applied to EEG spectrogram images. Fine-tuning updates all or part of the pre-trained weights on the target data [13]. Parameter-efficient fine-tuning updates only a small set of new parameters, such as prefix vectors or Low-Rank Adapters (LoRA). Domain adaptation aligns the feature distributions of a labelled source domain and a target domain, for example across EEG subjects, MRI sites or languages. Self-supervised pre-training learns representations from unlabeled data, as in wav2vec 2.0 for speech [14]. Large Language Models (LLMs) add a further option. They can be used without any training through zero-shot or few-shot prompting, or they can be fine-tuned on mental health text. Domain-specific encoders such as Mental BERT were pre-trained on mental health forum posts for the same purpose [15].

Rationale

Several reviews have covered machine learning and deep learning for depression [5–8], social media data [16], and LLMs in mental health [17,18]. Recent reviews have looked at transformer language models [19], multimodal LLMs [20] and fusion methods [21]. These reviews treat transfer learning as one technique among many, or they focus on one modality. None has compared the transfer strategies across text, speech, video, EEG, imaging and health records, and none has related the reported performance to the validation design. This matters because many studies split data at the level of segments or posts. The same person can then appear in both training and test sets, and the accuracy is inflated [22]. The earlier version of this manuscript was a narrative review and did not report a search method. The present version is a systematic review conducted and reported according to PRISMA 2020 [23].

Objectives

The review addresses four questions:

RQ1. Which transfer learning strategies and pre-trained models have been used to detect, classify or estimate the severity of depression and anxiety, and for which data modalities?

RQ2. What performance do these studies report, and how does it vary with the validation design?

RQ3. How well do the studies report methods that matter for clinical use, such as participant-independent testing, external validation, fairness and privacy?

RQ4. Which gaps remain for future research?

Protocol and registration

The review was not registered, and no protocol was published. The eligibility criteria, information sources, and search strings were fixed before screening began. The review follows the PRISMA 2020 statement [23]. The completed PRISMA 2020 checklist is given in Appendix B.

Eligibility criteria

Table 1 lists the criteria. Primary studies were included if they (a) addressed individual-level detection, classification, risk prediction or severity estimation of depression or anxiety in humans; (b) used at least one deep model with a transfer learning component, as defined in Section 1.2, in the proposed method or in a reported comparison; (c) reported at least one quantitative result; and (d) were peer-reviewed journal articles or full conference papers in English, published between 1 January 2019 and 30 September 2026. Reviews, editorials, protocols, conference abstracts, dataset papers without a detection experiment, and preprints were excluded. Studies were also excluded if the outcome was treatment response, suicide risk alone, stress alone or bipolar disorder alone, or if an LLM was used only as a chatbot, for therapy or to answer patient questions. Studies that used only classical machine learning, networks trained from random initialization, or static word embedding’s (for example, GloVe) were excluded.

Information sources

Two sources were searched on 27 October 2026. PubMed covers the clinical and psychiatric literature. Open Alex is an open index of scholarly works that covers journals and conference proceedings from IEEE, ACM, Springer, Elsevier and MDPI, among others. Scopus, Web of Science, and IEEE Xplore were not available to the author through an application programming interface at the time of the search. To reduce the effect of this gap, the reference list of the earlier version of this manuscript, which contained 111 references, was screened for eligible studies (citation searching).

Search strategy

The search combined three concept blocks: the condition (depression or anxiety in the title), the transfer learning or pre-trained model (in the title or abstract), and the task (detection, prediction, classification, diagnosis, screening, recognition or identification). The PubMed string was:

(depress*[ti] OR anxiety[ti] OR anxious[ti]) AND ("transfer learning"[tiab] OR "pre-trained"[tiab] OR pretrained[tiab] OR "fine-tun*"[tiab] OR finetun*[tiab] OR "domain adaptation"[tiab] OR BERT[tiab] OR RoBERTa[tiab] OR "large language model*"[tiab] OR "foundation model*"[tiab] OR GPT*[tiab]) AND (detect*[tiab] OR predict*[tiab] OR classif*[tiab] OR diagnos*[tiab] OR screen*[tiab] OR recogni*[tiab] OR identif*[tiab]) AND ("2019/01/01"[dp] : "2026/09/30"[dp]) AND english[la]

The OpenAlex search used the same blocks with the title.search and title_and_abstract.search filters, restricted to the document type "article" (which includes conference papers) and to English. The full strings are given in Appendix A.

Selection process

Records were exported with their abstracts and merged. Duplicates were removed by DOI, PubMed identifier, and normalized title. Titles and abstracts were then screened against the criteria. Records that passed screening were assessed for eligibility on the full abstract and bibliographic record. Where neither an abstract nor an open full text could be obtained, the record was classed as not retrieved. Screening and eligibility assessment were done by one reviewer (the author). A large language model (Claude, Anthropic) was used as an assistant: it proposed a decision and a reason code for each record, and the author checked these decisions. The decision log is provided as Supplementary File S2.

Data collection process and data items

Data were extracted into a structured form (Supplementary File S1). The items were: year; condition (depression, anxiety or both); data modality; transfer learning strategy; pre-trained backbone; dataset and population; validation design; and the main reported result. Modality was coded as social media text, clinical or narrative text (interview transcripts, diaries, notes, open responses), speech, video or facial images, EEG, (Magnetic Resonance Imaging) MRI or (Functional Magnetic Resonance Imaging) fMRI, wearable or physiological signals, electronic health records or tabular data, or multimodal. The strategy was coded as feature extraction, fine-tuning, parameter-efficient fine-tuning, domain adaptation, self-supervised pre-training, cross-corpus transfer, prompting of an LLM (zero-shot, few-shot, chain-of-thought or retrieval-augmented), or LLM fine-tuning. A study could have more than one strategy. Validation was coded as random hold-out split, k-fold cross-validation, participant-independent (subject- or user-disjoint) evaluation, external validation on an independent dataset or site, or not reported. Where an abstract gave only qualitative results, the result was recorded as not reported.

Study risk of bias assessment

A full PROBAST assessment [24] needs the full text of each paper, and this was not possible for 318 studies. A simplified appraisal of the PROBAST analysis domain was used instead. Three signaling items were judged for each study: (1) whether the validation was participant-independent or external; (2) whether the abstract reported a quantitative result; and (3) whether the evaluation unit was a segment, post or frame rather than a person. Studies that reported only random hold-out splits of segments or posts were judged to be at high risk of optimistic bias in the analysis domain.

Effect measures and synthesis methods

The studies used different datasets, labels and metrics, so meta-analysis was not appropriate. The results were grouped by modality and by transfer strategy. Counts, proportions and medians were calculated. For studies that reported accuracy, the median accuracy was compared across validation designs. Other metrics (F1 score, area under the curve, mean absolute error and root mean square error) are reported as given by the authors. Representative studies were selected for each modality on the basis of sample size, validation design and direct comparisons between strategies.

Reporting bias and certainty assessment

Reporting bias was examined by looking at the distribution of reported accuracies and at the share of abstracts without numerical results. A formal GRADE assessment was not applied, because GRADE is designed for intervention effects. The certainty of the main findings is described in words in Section 3.8.

Study selection

The database searches returned 798 records: 257 from PubMed and 541 from Open Alex (Figure 1). After 212 duplicates were removed, 586 records were screened. Of these, 202 were excluded: 77 did not address individual-level detection of depression or anxiety, 46 were reviews, editorials, case reports, or dataset papers, 33 did not use a transfer learning component, 23 were outside the topic (for example, respiratory depression or animal studies), 22 were preprints, and one was a conference abstract. Of the 384 reports sought, 43 had no abstract and no accessible full text. The remaining 341 were assessed, and 27 were excluded (14 without a deep transfer learning component, 4 with a different outcome, 4 without quantitative results, 4 that were reviews or conference abstracts, and one duplicate report of the same study). This left 314 studies. Citation searching of the earlier version identified six further reports; two were excluded (one preprint and one treatment-response study), and four were included [25–28]. In total, 318 studies were included.

Figure 1. PRISMA 2020 flow diagram of study selection.

Study characteristics

Table 2 summarizes the included studies. The field has grown quickly. Three studies were published in 2019, 28 in 2023, 65 in 2024, 106 in 2025, and 87 in the first nine months of 2026. In total, 258 studies (81.1%) were published from 2024 onward (Figure 2).

Figure 2. Number of included studies by publication year and pre-trained model family. *2026 covers January to September only.

Depression was the focus of almost all studies. Only 10 studies (3.1%) addressed anxiety alone, and 20 (6.3%) addressed both conditions. Text was the main data source. Social media posts from platforms such as Reddit, X (Twitter) and Weibo were used in 111 studies (34.9%), and clinical or narrative text in 55 (17.3%). Multimodal designs that combined text, audio, and video were used in 60 studies (18.9%). Speech alone was used in 28 studies, EEG in 21, electronic health records or tabular data in 15, MRI or fMRI in 11, video or facial images in 8, and wearable or physiological signals in 8.

Public benchmark datasets were common. At least 99 studies used a public dataset, and 62 used the DAIC-WOZ, E-DAIC or AVEC interview corpora. Fourteen studies used MODMA. At least 45 studies worked with languages other than English, including Chinese, Bengali, Urdu and Roman Urdu, Korean, Thai, Indonesian, Spanish, Turkish, Hebrew, and Arabic. Only two studies on Arabic text were found [29,30].

Fine-tuning was the most common strategy (143 studies), followed by feature extraction (86), prompting of an LLM (59), domain adaptation (18), LLM fine-tuning (18), self-supervised pre-training (10), parameter-efficient fine-tuning (9) and cross-corpus transfer (2). By backbone family, 141 studies used pre-trained transformer encoders such as BERT, RoBERTa or MentalBERT, 87 used LLMs or other foundation models, 40 used ImageNet-type convolutional networks or vision transformers, and 17 used pre-trained speech models such as wav2vec 2.0, HuBERT, WavLM, or Whisper. The use of LLMs rose from one study in 2023 to 16 in 2024, 31 in 2025, and 39 in 2026. Table 3 gives examples of each strategy.

Methodological quality

Table 4 shows the validation designs. A random hold-out split was the most common design (165 studies, 51.9%). Cross-validation was reported in 21 studies (6.6%). Only 16 studies (5.0%) reported participant-independent evaluation and 29 (9.1%) reported external validation. Together, 45 studies (14.2%) met the first signaling item. Table 5 lists representative studies from this group. In 87 studies (27.4%), the abstract did not describe the validation design.

The second item was also weak. In 98 studies (30.8%), the abstract gave no numerical result, only statements such as "outperforms state-of-the-art methods". For the 103 studies that reported accuracy, the median was 91.3%, and 36 of them reported 95% or more. Ten studies reported 99% or higher [29,35,61–63]. Several of these very high values came from small EEG datasets, from evaluation on segments rather than people, or from datasets in which one class was generated or curated by the authors. For example, the normal class in one Arabic tweet study was generated with ChatGPT [29]. One speech study stated clearly that its 96.7% accuracy was a segment-level result and should not be read as participant-level performance [64]. Such statements were rare.

Accuracy fell as validation became stricter. The median was 95.0% in studies that did not report their design, 90.8% for hold-out splits, 90.0% for cross-validation, 90.8% for participant-independent testing and 88.0% for external validation. None of the five externally validated studies with an accuracy value reported 95% or more. The numbers in the stricter groups are small, so the pattern is descriptive.

Results by data modality

Social media text

Social media studies were the largest group. Most fine-tuned BERT-family encoders on labelled posts and reported high accuracy on in-domain test sets. Domain-specific pre-training helped with several comparisons. MentalBERT gave the best results among BERT, RoBERTa, and DistilBERT on 19,800 posts from three platforms (F1 0.918, AUC 0.962), but all models lost performance when tested on a different platform [32]. A large comparison of 14,983 tweets labelled with five depression subtypes found that fine-tuned encoders reached a macro-F1 of 0.94–0.96, while few-shot prompting of instruction-tuned LLMs reached 0.73–0.77 [31]. A DSM-5-guided BERT model on the WU3D Weibo dataset (32,570 users) reached an F1 of 0.918, compared with 0.869 for few-shot GPT-4 [77]. Studies that used user-level splits and an external dataset were few [73,74].

Prediction of future diagnosis was harder than classification of posts. When BERT and MentalBERT were used to identify Reddit users who later disclosed a depression diagnosis, the mean F1 was 0.64 [78]. In low-resource languages, transfer learning did not always win. A support vector machine outperformed BERT on a Roman Urdu and English dataset [79], a stacking ensemble of character n-grams outperformed AlephBERT on Hebrew health forum posts [80], and an LSTM outperformed a fine-tuned BERT in another comparison [81].

Clinical interviews, narratives and notes

Interview transcripts, diaries, sentence completion tests, and clinical notes were used in 55 studies. These data are closer to clinical practice, and the results were more modest. Two studies tested locally deployed open-source LLMs with fewer than 15 billion parameters on DAIC-WOZ. Accuracy ranged from 0.23–0.68 in one study [39] and reached 0.677 at best in the other [82]. The authors concluded that such models were not accurate enough for screening. In contrast, LLMs used as feature extractors performed better. Prompting an LLM to extract interpretable depression factors, followed by linear regression, gave a PHQ-8 mean absolute error of 2.91 on DAIC-WOZ and 2.86 on E-DAIC [40]. Prefix tuning of a pre-trained language model gave a root mean square error of 4.67 [55].

Fine-tuning improved LLM results in direct comparisons. In 3,536 Korean participants, supervised fine-tuning of five open-source LLMs raised the F1 score to between 0.818 and 0.865, compared with a best zero-shot F1 of 0.735, and the models were tested externally on E-DAIC [43]. In another comparison, fine-tuned GPT-3.5 reached an accuracy of 0.902 on 428 diary entries [44]. On psychotherapy transcripts, fine-tuned BERT, RoBERTa, Longformer and Mistral-7B did not improve on a support vector machine with engineered features [83].

Clinical notes gave mixed results. A GPT-4 prompt detected depression in notes of patients with multiple sclerosis with a sensitivity of 97.3% but a specificity of 68.3% [41]. GPT-4o estimates of PHQ-9 scores from 15,000 outpatient notes were only modestly correlated with the true scores (r² = 0.264), with a positive predictive value of 0.309 [84]. In 299 participants, four of eight LLMs trained on Hamilton interview transcripts outperformed the self-report HADS-D, and the gap was largest in people with alexithymia [85]. For anxiety, a fine-tuned transformer reached an AUROC of 0.64 for generalized anxiety disorder in 2,000 speech transcripts, compared with 0.58 for a LIWC baseline [86]. GPT-4 ratings of social anxiety from 51 interviews correlated with self-report scores at r = 0.79 [87].

Speech

Pre-trained speech models gave some of the best externally validated results. In a multicenter cohort of 1,816 participants, self-supervised WavLM and HuBERT representations reached an AUC of 0.932 internally and 0.879 in external validation [65]. In 1,838 Chinese adolescents, a fine-tuned Whisper model reached an AUROC of 0.95 in development and 0.88 in an external cohort, and performance was similar across age, sex, and severity subgroups [33]. In 865 adults from the general population, Whisper features gave AUCs of 0.76–0.78 for depression and 0.77 for anxiety, with fairness and uncertainty analyzed [88]. Earlier, transfer from a deep language model trained on 16,000 recorded interactions gave AUCs of 0.79–0.86 for depression, anxiety, and their co-occurrence [25], and transfer of encoder weights gave relative gains of up to 27% in PHQ-8 prediction [26].

Cross-corpus and cross-language transfer was possible but weaker. A speech emotion model trained on German and English corpora separated Danish patients with depression from controls with an AUC of 0.71 [59]. Transfer from English to Chinese speech reached 81.1% accuracy [49]. Under participant-independent evaluation on DAIC-WOZ, a wav2vec 2.0 model reached 70.2% accuracy but a sensitivity of only 0.36 [72]. This contrasts with studies on the same corpus that reported accuracies above 96% without a participant-independent design [89].

Multimodal data

Multimodal studies combined pre-trained text encoders (mostly BERT or RoBERTa), speech models and visual networks. Most reported gains over single modalities on DAIC-WOZ, E-DAIC, or AVEC. A model that fused auditory, visual, and textual cues from chatbot interviews reached an AUC above 0.95 internally and 0.978 in an external dataset of 100 participants [69]. Other studies were more cautious. On the DAIC-WOZ test set, a late-fusion model reached an F1 of only 0.50 for the depressed class, and visual features added noise [90]. A controlled study of 512 fusion conditions found that affective pre-training could help one modality and harm another, which shows that negative transfer occurs [91]. The AudiBERT model used audio and text representations for screening and reported higher F1 scores than earlier models [28].

EEG and neuroimaging

EEG studies often converted signals into time-frequency images and applied ImageNet networks such as ResNet, VGG, DenseNet, or EfficientNet. Within-dataset accuracies were high, often above 95% [35,61]. Ten of the 21 EEG studies reported cross-subject (participant-independent) evaluation, a higher rate than in any other modality. Cross-subject results were lower: 87.1% on MODMA with multi-source few-shot adaptation [47] and 77.0% with three-channel wearable EEG [71]. One study reported 88%–98% within each database but 75.1% when a model trained on one database was tested on another [34]. For anxiety, a model reached 95.7% on a public dataset and 86.6% after domain adaptation to seven new participants [92].

MRI studies used pre-training on large cohorts or other disorders, such as Alzheimer's disease. The largest study used 3,335 participants from 33 sites and reached an average accuracy of 70.1% with site-wise transfer [70]. A model pre-trained on polygenic scores and diffusion MRI in 4,741 adolescents reached an AUC of 0.62, and 0.67 in an independent Korean sample [53]. These values are more modest than those of EEG studies, but they come from much larger and more varied samples.

Health records, wearables and other signals

Health record studies used fine-tuned BERT on clinical text or LLMs on tabular data. Contrastive pre-training on 85,085 unlabeled discharge records improved prediction of depression and anxiety in type 2 diabetes, with an AUC of 0.91 internally and 0.75 externally [51]. In 16,159 patients with cancer, BERT models on notes and emails reached an AUROC of about 0.71, below a LASSO model on structured data (0.74), and the BERT risk estimates were poorly calibrated [93]. A fine-tuned 70-billion-parameter LLM on UK Biobank data reached an AUC of 0.892 [46]. Wearable studies were few. A smartwatch model pre-trained on electrocardiogram data tracked post-stroke anxiety with an AUC of 0.859 [52], and a wrist-worn system used domain adaptation to reach 90.8% cross-subject accuracy [50]. A tabular foundation model (TabPFN) did well on a very small student dataset but not on larger ones [94].

Large language models

The 87 studies that used LLMs or other foundation models fell into three groups. In the first group, a general model such as GPT-4, GPT-4o, Llama, Qwen, or Gemini was prompted without training. These studies were quick to run, but results were variable and often lower than those of fine-tuned encoders [31,77]. The performance of small local models was low [39,82]. In the second group, an LLM was fine-tuned, often with LoRA or quantisation, on depression data. These studies reported clear gains over prompting [43,44,67]. Some also trained the model to give DSM-5-based explanations, with reinforcement learning [95]. In the third group, the LLM served as a feature extractor or data generator. It produced symptom scores, summaries or synthetic samples that a smaller model then used [40,96,97]. One study found that clinician-selected LLM outputs gave better classifiers than outputs selected by the LLM itself [97]. Few LLM studies tested the stability of outputs across repeated runs, and agreement among different LLMs was poor in one study [39].

Comparison of transfer strategies

Direct comparisons within the same study allow some conclusions. Fine-tuning a pre-trained speech model usually outperformed a frozen model with a new head. Full fine-tuning of WavLM outperformed a head-only WavLM model and a CNN-RNN model trained from the start (ROC-AUC 0.90) [98]. LoRA fine-tuning of wav2vec 2.0 outperformed WavLM and HuBERT with far fewer trainable parameters [56]. Parameter-efficient methods were less prone to overfitting on DAIC-WOZ [55]. Domain adaptation was mainly used for EEG and MRI, where differences between subjects and sites are large [48,99–101]. Pre-training on emotion data was a common bridge to depression, but the effect depended on how close the source domain was to the target. Negative transfer was reported when the source audio domain did not match [91].

Anxiety and comorbidity

Anxiety was studied far less than depression. The 30 studies that included anxiety used text [57,86,87,102], speech [25,56,88], EEG [92] and wearable data [52]. Comorbidity was addressed in only a few studies. Multi-label classification of depression, anxiety, and substance use disorder in 1,661 Korean counselling sessions gave F1 scores of 0.90 for depression and 0.85 for anxiety [103]. In 130 Turkish women, comorbid cases were assessed more accurately from speech than depression-only cases [56], and the best results in one large speech corpus were seen for comorbid cases [25]. A stacked transformer model separated anxiety from depression posts with an accuracy of 93.4% under a user-wise split [74].

Reporting biases and certainty of evidence

The distribution of results suggests reporting bias. Over a third of studies that reported accuracy gave values of 95% or higher, while studies with stricter designs gave lower values. Almost one third of abstracts did not report any number. Negative findings were reported, but by a small group of studies [39,82–84,93]. On this basis, the certainty that deep transfer learning can detect depression from text and speech in research datasets is moderate. The certainty that the reported accuracies would hold in clinical populations is low. For anxiety, comorbidity, wearable data, and non-English populations, the certainty is very low, because the studies are few and small.

Summary of evidence

This review identified 318 studies of deep transfer learning for depression and anxiety. Transfer learning has become the standard approach in this field. Pre-trained transformer encoders dominate text studies, pre-trained speech models dominate speech studies, and LLMs have spread rapidly since 2024. The reported performance is high, but the evidence base has clear weaknesses. Only one study in seven tested models on new participants or new data sources, and these tests gave lower values. The strongest evidence comes from a small number of large speech and health record studies with external validation [33,51,65].

Interpretation

The results agree with earlier reviews that describe high reported accuracy and limited clinical validation [5,6,17]. This review adds three points. First, the strategy of transfer matters. Fine-tuned domain-specific encoders outperformed prompted general LLMs on most text classification tasks [31,77]. LLMs were most useful when they were fine-tuned or used to extract interpretable features. Second, transfer learning does not remove the need for representative data. In low-resource languages and in clinical text, simple models sometimes performed as well as or better than pre-trained ones [79,80,83]. Third, the validation design is the main source of the variation in results. Segment-level or post-level splits allow information from one person to leak into the test set [22]. A model can then learn speaker identity or writing style rather than signs of depression. One study showed that depression signals in speech overlap with the features used to recognize speaker identity [76], which explains why participant-independent results are lower.

The narrow focus on depression is also a problem. Depression and anxiety often occur together and share symptoms. Few studies modelled them jointly, although the available results suggest that comorbid cases can be detected well [25,56]. Social media labels are a further concern. Many datasets use self-disclosure, subreddit membership or keyword-based labels rather than clinical assessment. Studies that used clinical labels, such as Hamilton or PHQ-9 scores and psychiatrist diagnosis, gave a more realistic picture [33,65,85].

Ethics, privacy, and fairness

Privacy was addressed more often in recent studies. Several groups deployed open-source LLMs on local servers so that patient text did not leave the hospital [39,43,82]. One study used federated learning with audio and EEG data [104], and a recent review describes the wider use of federated learning in mental health [105]. Another study removed speaker identity from speech features to protect privacy [106]. Fairness was tested in only a few studies. BERT models for patient messages in cancer care showed larger performance gaps across sex, race, and ethnicity than a Reddit-pre-trained variant [107]. LLM accuracy varied with age and language in a multilingual corpus [108], whereas a speech model for adolescents showed no significant fairness gaps [33]. Social media screening also raises questions of consent, because posts are often collected without the users' knowledge. These issues need ethical review and clear consent procedures, and data protection should follow national rules.

Implications for research and practice

Five practical steps follow from this review. Studies should split data by participant and should report results per person, not per segment or post. External validation on a different site, population or language should be routine before any clinical claim is made. Reporting should follow TRIPOD+AI [109], including sample size, class balance, calibration and subgroup performance, and risk of bias should be assessed with PROBAST [24]. Anxiety and comorbidity need more attention, preferably with clinically labelled data. Finally, more work is needed in under-represented languages. Only two Arabic text studies were found [29,30], and none used clinical data from Arabic-speaking patients. Given the burden reported by the Saudi National Mental Health Survey [4], Arabic speech and text corpora with clinical labels would be a useful contribution, and they would support the health aims of Saudi Vision 2030 and Sustainable Development Goal 3 on good health and well-being.

For clinicians, the current evidence supports the use of these models as research tools and, at most, as aids to triage that are always followed by clinical assessment. They should not replace validated scales or clinical interviews.

Limitations of the evidence

The included studies used different datasets, labels, and metrics, so the results could not be pooled. Many studies used the same few public datasets, especially DAIC-WOZ, so their results are not independent. Sample sizes were small in most EEG, video, and clinical text studies. Labels in social media studies were often weak proxies for clinical diagnosis. Many abstracts did not report the validation design or any number.

Limitations of the review process

This review has important limitations. Only two databases were searched. Open Alex covers much of the engineering literature, but Scopus, Web of Science, and IEEE Xplore were not searched directly, so some studies may have been missed. Eligibility assessment and data extraction were based on the full abstract and bibliographic record, not on the full text of each paper, because of the size of the evidence base. Some details, such as sample size and the exact split, could therefore not be checked. Forty-three records could not be assessed because no abstract or full text was available. Screening was done by one reviewer with the help of a large language model, and a second independent reviewer was not available. Only English-language publications were included, and preprints were excluded. Finally, the field is moving fast, and many relevant studies will have appeared after the search date.

Deep transfer learning is now the main approach to automatic detection of depression and, to a smaller extent, anxiety. Pre-trained language and speech models give the best results, and fine-tuning generally works better than prompting a general LLM. However, most of the 318 included studies used random splits of small or public datasets, and only 14.2% tested models on new participants or new data sources. The high accuracies reported in many abstracts should therefore be read with caution. The next step for the field is fewer new architectures and more careful evaluation: participant-independent splits, external and multi-site validation, clinically labelled data, joint modelling of depression and anxiety, fairness testing, and data from under-represented languages such as Arabic.

Funding

This research received no external funding.

Conflicts of interest

The author declares no conflict of interest.

Data availability

The search strings, the screening decision log and the data extraction sheet are provided as supplementary files.

Author contributions

 T.S.M. designed the review, carried out the searches, screened the records, extracted and synthesized the data, and wrote and revised the manuscript.

Use of generative AI

A large language model (Claude, Anthropic) was used to assist with screening, data extraction, and language editing. The author checked all decisions and the final text and takes full responsibility for the content.

Acknowledgements

The author thanks the two anonymous reviewers, whose comments led to this revised version.

Supplementary materials

Supplementary File S1: data extraction sheet for the 318 included studies, with full bibliographic details. Supplementary File S2: record-level screening and eligibility decision log for all 586 screened records and the six reports from citation searching.