Contextualized Early Detection of Depression

• Contextualized early detection of depression is defined as assessing integrated cues from language, audio, visuals, and physiology within their surrounding context.
• It leverages multimodal fusion techniques such as BiLSTM, transformers, and graph attention to combine diverse data sources for reliable risk estimation.
• Evaluation metrics span accuracy, AUROC, ERDE, and calibration, ensuring timely interventions while balancing predictive performance.

Contextualized early detection of depression denotes the identification of depressive risk from signals interpreted within their surrounding context rather than from isolated lexical, acoustic, visual, or physiological cues. In current work, that context may be the semantic and prosodic evolution of a clinical interview, the thematic structure of question–answer dialogue, the temporal pattern of social-media activity, the history of emotionally intense posts, the object and scene composition of vlogs, or the dynamics of physiological recordings. The “early” component is likewise task-dependent: some systems issue threshold-based warnings during an ongoing interview, some flag risk from a single post, some operate on streaming user histories, and some distinguish pre-diagnosis “high-risk” content from post-diagnosis depression content (Singh et al., 6 Feb 2025, Chen et al., 18 Mar 2025, Zhang et al., 2022, Cha et al., 2024).

1. Conceptual boundaries of contextualized early detection

In clinical-interview work, early detection is often formulated as a continuous risk-estimation problem. One multimodal DAIC‑WOZ framework defines a mental disorder score as a sigmoid output $\hat{y} \in [0,1]$ and then applies a threshold: if the score exceeds a predefined threshold, the system triggers “Early Intervention”; otherwise it schedules “Regular Monitoring” (Singh et al., 6 Feb 2025). In that formulation, “contextualized” refers to jointly modeling what is said and how it is said across the dialogue.

In social-media research, the same phrase can denote substantially different operational tasks. LLM-MTD treats early detection as post-level classification on Reddit, where a single post may be sufficient to flag depression risk and to generate a medically informed explanation (Chen et al., 18 Mar 2025). Other work frames early detection as user-level or stream-level inference, where posts arrive over time and the system must decide whether to classify now or wait for more evidence, as in eRisk-style early risk detection (Zhang et al., 2022, Bucur et al., 2020). A recent eRisk benchmark makes the contextual requirement even more explicit by defining Task 2 as “Contextualized Early Detection of Depression” from full conversation trees, with the target subject appearing either in the submission or in comments and evaluation performed after 1, 100, 500, and 1000 writings (Marmol-Romero et al., 24 Sep 2025).

Video-based work extends the notion of context beyond text and audio. MOGAM constructs “high-risk depression” vlogs from videos uploaded before a user’s first diagnosis vlog and “depression” vlogs from videos uploaded after diagnosis, thereby turning pre-diagnosis versus post-diagnosis status into an early-detection proxy (Cha et al., 2024). This suggests that “contextualized early detection” is not a single standardized task but a family of related formulations defined by when evidence becomes available, which modality carries that evidence, and how intervention timing is operationalized.

2. Context sources and data regimes

Clinical-interview studies typically use semistructured or structured dialogues in which linguistic, prosodic, and interactional context are tightly coupled. The DAIC‑WOZ interview corpus provides synchronized audio at 16 kHz and text transcripts for 20–60 minute interviews with a virtual agent. In one multimodal system, each text segment is represented as a 768-dimensional BERT embedding and each audio segment as a 193-dimensional vector including MFCCs, Delta MFCCs, Delta$^2$ MFCCs, Chroma, Mel spectrogram, Spectral contrast, Tonnetz, and Pitch (Singh et al., 6 Feb 2025). PDIMC treats these interviews as explicitly multi-theme dialogues and extracts theme-specific text for “family,” “work,” “mental,” “medical,” and “overall,” thereby redefining context in terms of intra-theme and inter-theme relations rather than raw turn sequence alone (Zhao et al., 16 Feb 2025). A later interview-based study combines raw 16 kHz audio, full transcripts, and a tabular feature set built from 74 COVAREP variables, 5 formants, and 7 SpaCy-based lexical metrics, yielding a 550-feature structured representation aligned with routine patient interviews (Weber et al., 26 Aug 2025).

Social-media work uses a broader variety of contextual signals. LLM-MTD operates on the Reddit Self-Reported Depression Dataset and emphasizes intra-post linguistic context, where a transformer encodes the full post and explanations are conditioned on prompts referencing medical knowledge and diagnostic criteria (Chen et al., 18 Mar 2025). A different Reddit study derives temporal behavior from created_utc, including month, day, hour, weekday, is_working_hour, and is_weekend, then models text and time jointly for classification across mental-health-related subreddits, including depression (Saeed et al., 6 Mar 2025). Mood2Content constructs DepCOV, anchors user timelines at self-reported COVID-19 infection and later depression signals, aggregates tweets into daily documents over up to four weeks, and treats daily mood swings as psychiatric signals (Wu et al., 2023). DORIS restricts user histories to the six months before the last post, retrieves high emotional intensity posts, and summarizes a “mood course” with an LLM (Lan et al., 2024). DKDD further distills user histories into time-ordered entities representing symptoms, major life events, and treatments, thereby replacing raw post streams with clinically interpretable digital traces (Zhang et al., 2023).

User-generated video introduces yet another kind of context. MMFformer uses D‑Vlog and LMVD to model facial action units, landmarks, gaze, head pose, and audio features such as eGeMAPS or VGGish embeddings, explicitly targeting spatio-temporal high-level patterns in vlogs (Haque et al., 8 Aug 2025). MOGAM converts each vlog into a graph of object co-occurrences detected by YOLOv5, augments that graph with ResNet visual features, KoBERT title and description embeddings, and vlog duration, and then treats this object-centered multimodal representation as contextual evidence (Cha et al., 2024).

Physiological work broadens context beyond discourse altogether. One study uses GSR, skin temperature, and pupillary dilation recorded while participants watch videos, with 85 total input features and four depression-level classes (“None,” “Mild,” “Moderate,” “Severe”) (Yang, 2022). Another uses resting-state EEG from only Fp1, Fpz, and Fp2, extracting power-spectrum statistics and nonlinear features such as correlation dimension, Rényi entropy, and C0 complexity for depressed versus control classification (Shi et al., 2020). This suggests that the adjective “contextualized” can refer not only to conversational or temporal context but also to experimentally controlled physiological context.

3. Modeling paradigms

A large part of the literature uses sequence models or transformer encoders to fuse complementary sources of context. The DAIC‑WOZ multimodal system in (Singh et al., 6 Feb 2025) adopts a dual-branch architecture: BERT embeddings are passed through a BiLSTM for text, audio features through an LSTM for speech, and the two 64-dimensional branch outputs are fused by element-wise averaging,

$$z = \frac{1}{2}\left(u^{(\text{text})} + u^{(\text{audio})}\right),$$

followed by a sigmoid output for binary depression or PTSD classification. A Reddit model similarly combines a BiLSTM text module and an LSTM temporal-feature module with cross-modal attention to form a context-aware classifier for depression and related classes (Saeed et al., 6 Mar 2025). MMFformer extends this pattern to video by pairing a ViT-style spatial transformer for visual features with a transformer encoder over audio spectrogram patches and then testing late, intermediate, and attention-based fusion strategies (Haque et al., 8 Aug 2025). MOGAM replaces recurrent sequence modeling with graph attention over object co-occurrence and then applies cross-attention,

$F = ATTENTION(F_g, F_a, F_a),$

where $F_g$ is the object-graph representation and $F_a$ is the concatenation of visual and metadata features (Cha et al., 2024). In interview diagnosis, a separate multimodal design trains Wav2Vec2 on 30-second audio chunks, BERT on full transcripts, and XGBoost on engineered tabular features, then performs weighted late fusion followed by logistic-regression calibration (Weber et al., 26 Aug 2025).

A second paradigm is explicitly knowledge-aware and explanation-oriented. LLM-MTD uses a single pre-trained LLM with a classification head and a generation head, optimized by the multi-task objective

$$\mathcal{L}_{total} = \lambda \mathcal{L}_{cls} + (1-\lambda)\mathcal{L}_{gen},$$

so that classification and medically grounded explanation generation share the same contextual representation (Chen et al., 18 Mar 2025). DORIS separates representation learning from classification: it constructs a diagnostic-criteria feature from DSM-5-like symptom annotations, a mood-course feature from high emotional intensity posts plus an LLM-generated temporal summary, and a post-history embedding, then forms

$$F_u = \mathrm{Concat}\big(F_u^{MC} + F_u^{PH},\ F_u^{DC}\big),$$

which is fed to Gradient Boosting Trees (Lan et al., 2024). DKDD encodes symptom, life-event, and treatment entities with an LSTM, augments them with temporal attention and ontology attention derived from a depression ontology, and fuses the two forms of knowledge with user-specific weights before final classification (Zhang et al., 2023). HAN-BERT (Psych) begins even earlier in the pipeline by screening posts against template sentences derived from clinical depression scales, retaining only the top-$K$ risky posts for hierarchical encoding and online early risk detection (Zhang et al., 2022). PDIMC adds an explicitly interactive variant by using in-context learning to extract themes, attention to model intra-theme and inter-theme relations, and an Interactive Theme Adjustment Strategy to reweight themes according to simulated clinician feedback (Zhao et al., 16 Feb 2025).

Low-resource conversational settings have produced a third, transfer-oriented line of work. In (Lorenc et al., 2022), chunked multi-turn conversations are encoded with sentence embeddings and biLSTMs, and transfer from eRisk Reddit data to the conversational DAIC‑WOZ domain yields a UAR of 0.803 with Chunk-biLSTM + $USE_5$. This suggests that contextualized early detection can also be cast as a domain adaptation problem in which rich social-media representations are used to compensate for the scarcity of labeled conversational data.

4. Early-decision mechanisms and evaluation regimes

Early detection systems differ not only in architecture but in the mechanism by which they turn continuous evidence into a timed decision. In interview-based multimodal scoring, the sigmoid output itself is the “mental disorder score,” and intervention is a direct thresholding problem: if $^2$0 exceeds a predefined threshold, the system triggers “Early Intervention”; otherwise it schedules “Regular Monitoring” (Singh et al., 6 Feb 2025). In streaming eRisk-style settings, the decision process is more elaborate. HAN-BERT (Psych) maintains an evolving queue of up to $^2$1 risky posts and runs the full model only when a new post changes that queue; if the resulting score crosses threshold $^2$2, the system issues an alert and stops. In the reported experiments, this design required inference for only 10.41% of posts while improving ERDE and F1 over baselines (Zhang et al., 2022). A related system learns an explicit confidence scalar $^2$3 alongside topic-based depression classification and then decides at week $^2$4 according to a decaying threshold,

$^2$5

so that high confidence is demanded early and lower confidence is tolerated later (Bucur et al., 2020).

Evaluation protocols reflect these different decision regimes. Classical discrimination metrics remain common: the multimodal DAIC‑WOZ interview model reports 0.92 accuracy for depression (Singh et al., 6 Feb 2025); LLM-MTD reports 0.882 accuracy, 0.820 F1, and 0.880 AUPRC on RSDD (Chen et al., 18 Mar 2025); Mood2Content reports AUROC 0.9317 and AUPRC 0.8116 for depression risk in COVID-19 patients (Wu et al., 2023); DORIS reports AUPRC 0.8134, AUROC 0.9715, and F1 0.7596 on SWDD (Lan et al., 2024); PDIMC reports accuracy 0.94, precision 0.89, recall 0.92, and F1-score 0.90 on DAIC‑WOZ (Zhao et al., 16 Feb 2025); and the interview fusion system in (Weber et al., 26 Aug 2025) reports AUROC 0.88 and F1-score 0.75, together with calibration and decision-analytic results. This suggests that direct numerical comparison across papers is usually inappropriate, because datasets, labels, unit of prediction, and the meaning of “early” differ substantially.

Early-detection benchmarks therefore supplement classical metrics with explicit timeliness measures. eRisk work uses ERDE$^2$6 and ERDE$^2$7, where lower scores indicate better trade-offs between correctness and delay; HAN-BERT (Psych) reports ERDE$^2$8 10.72, ERDE$^2$9 8.12, and F1 60.3 on eRisk2017 (Zhang et al., 2022). The eRisk 2025 contextualized conversation task adds latencyT, speed, Flatency, and ranking metrics such as P@10 and NDCG@10 after 1, 100, 500, and 1000 writings. SINAI-UJA’s best run attains F1 0.39 with latencyT 3.00 and speed 0.99, and its runs are described as among the fastest in issuing early predictions even though the system ranks 8th out of 12 by F1 (Marmol-Romero et al., 24 Sep 2025). The interview model in (Weber et al., 26 Aug 2025) extends evaluation in a different direction by reporting calibration error 0.04 and higher net clinical benefit than “treat all” and “treat none” strategies, thereby linking prediction to action thresholds rather than only to discrimination.

5. Explainability, clinical integration, and trust

Explainability is central because depression detection is repeatedly treated as a sensitive, high-stakes task. LLM-MTD integrates explainability into the model objective itself: when a post is classified as positive, the LLM generates a free-form explanation grounded in diagnostic symptoms such as hopelessness, anhedonia, fatigue, and worry. In a human evaluation of 100 correctly classified positive posts, explanations from LLM-MTD receive average scores of 4.5 for relevance, 4.2 for completeness, and 4.6 for medical accuracy, compared with 3.2, 2.8, and 3.0 for a rule-based baseline (Chen et al., 18 Mar 2025). DORIS makes explanation multi-level by exposing per-post symptom annotations, a mood-course summary over time, and a final natural-language explanation conditioned on both (Lan et al., 2024). DKDD visualizes attention on diagnosis-related entities such as “self harm,” “suicidal ideation,” or major life events, so the model’s decision can be traced to time-stamped digital traces rather than to opaque embeddings alone (Zhang et al., 2023). HAN-BERT (Psych) links each risky post to a diagnostic basis derived from clinical-scale templates and supplements that with post-level attention weights (Zhang et al., 2022). PDIMC exposes theme-level structure and theme-weight adjustment, which provides an interpretable intermediate space closer to clinical reasoning than an undifferentiated transcript embedding (Zhao et al., 16 Feb 2025).

Clinical integration follows directly from these explanation mechanisms. The DAIC‑WOZ multimodal scoring framework in (Singh et al., 6 Feb 2025) is explicitly tied to “Early Intervention” versus “Regular Monitoring,” and its Flask-based interface returns JSON outputs such as disorder status and probability. The interview model in (Weber et al., 26 Aug 2025) is described as an adjunct to psychiatric evaluation and is positioned for routine interviews, primary care, telepsychiatry, and longitudinal monitoring; calibration and decision-curve analysis are used to estimate clinical utility rather than only classification strength. Social-media systems are similarly framed as triage or moderation aids rather than autonomous diagnosticians: LLM-MTD is proposed for clinical review, platform support, and research surveillance (Chen et al., 18 Mar 2025), while DORIS is presented as a way to support screening by combining mood trajectories with clinically guided features (Lan et al., 2024). Across these papers, trust is rarely reduced to a single attribution map; it is more often built through clinically grounded data, explicit symptom categories, transparent fusion roles, calibrated probabilities, and human-readable rationales.

6. Limitations, controversies, and future directions

A recurrent limitation is the instability of labels and task definitions. Some studies use clinically grounded interview corpora such as DAIC‑WOZ with PHQ-derived or clinician-validated labels (Singh et al., 6 Feb 2025, Weber et al., 26 Aug 2025), whereas others rely on subreddit names, self-reports, or timing relative to a self-disclosed diagnosis (Saeed et al., 6 Mar 2025, Cha et al., 2024). Weak supervision via community membership can conflate disorder-specific language with platform-specific norms, and pre-diagnosis labeling in vlogs is a proxy rather than a clinical assessment. This suggests that contextualized early detection remains partly a label-engineering problem: what counts as “early,” “depressed,” or “high-risk” depends strongly on data construction.

Generalization is also constrained by domain, language, and scale. Several papers explicitly note that DAIC‑WOZ is relatively small, semistructured, and not necessarily representative of telemedicine or informal conversation (Singh et al., 6 Feb 2025). Social-media corpora are often English-specific, culturally narrow, or imbalanced, and Reddit users are not representative of broader populations (Chen et al., 18 Mar 2025, Weber et al., 26 Aug 2025). Transfer learning partly mitigates this scarcity in conversational settings, but the need to adapt from social-media data to dialogue data in (Lorenc et al., 2022) underscores the mismatch between source and target domains. Physiological studies face a different version of the same issue: the three-channel EEG study reports 72.25% accuracy on a very small sample, and the physical-characteristics study reports 83% accuracy with a 3-layer network and Kaiming initialization on 192 records, but both remain far from broad, demographically validated deployment (Shi et al., 2020, Yang, 2022).

Future work in the literature is notably convergent. Multiple papers call for richer multimodality, including gestures, facial expressions, physiology, images, social interaction data, or wearable signals (Singh et al., 6 Feb 2025, Haque et al., 8 Aug 2025). Others call for explicit longitudinal modeling of user posting history or depression trajectories rather than isolated post classification (Chen et al., 18 Mar 2025, Lan et al., 2024). Cross-lingual adaptation, multilingual transformers, fairness auditing, and stronger privacy protections are repeated concerns (Singh et al., 6 Feb 2025, Chen et al., 18 Mar 2025, Zhang et al., 2023, Weber et al., 26 Aug 2025). Interactive systems are also becoming more prominent: PDIMC uses interactive theme adjustment in interviews (Zhao et al., 16 Feb 2025), and the eRisk 2025 pilot task shows that structured conversational strategies with LLM-powered personas can top all evaluation metrics in a depression-assessment setting (Marmol-Romero et al., 24 Sep 2025). A plausible implication is that the field is moving from static detection toward adaptive, multimodal, and human-in-the-loop assessment systems in which contextualization is not merely additional input information but the organizing principle of inference, explanation, and intervention.