Source-linked AI summary
AdoDAS: A Privacy-Preserving Multimodal Challenge for Adolescent Depression, Anxiety, and Stress Assessment
Zhaojie Luo, Junkun Wang, Tianhua Qi, Yuxuan Wu, Xin Zhao, Tetsuya Takiguchi, Tomoko Matsui, Kun Qian, Fei Wang, Shuqiong Wu, Zhengjun Yue, Hiroshi Ishiguro, Xinyuan Qian, Haizhou Li
TL;DR
AdoDAS addresses adolescent D/A/S assessment with a privacy-preserving benchmark that withholds raw recordings and evaluates complementary screening and item-prediction tracks. Audio-visual baselines scored 0.4604 mean F1 and 0.2675 mean QWK; leading submissions reached 0.5921 mean F1 and 0.2776 mean QWK.
Problem
Adolescent assessment is constrained by reliance on self-report, limited specialist capacity, and difficulty observing symptoms consistently, while privacy concerns complicate benchmarks using minors’ raw recordings.
Source: §1 · p. 2Method
AdoDAS releases anonymized audio-visual representations and ASR text for participant-disjoint screening and ordinal DASS-21 item-prediction tracks.
Source: §1 · p. 2, §2.3 · p. 3Results
Audio-visual baselines scored 0.4604 mean F1 and 0.2675 mean QWK, while leading systems reached 0.5921 mean F1 and 0.2776 mean QWK.
Source: §6 · p. 4Takeaways & Limitations
AdoDAS provides a bounded, evidence-based starting point for privacy-preserving multimodal adolescent D/A/S research.
Source: §6 · p. 4The split tests unseen participants under matched collection conditions, not unseen schools, cultures, devices, or clinical populations, and withheld raw recordings do not provide a formal privacy guarantee.
Source: §6 · p. 4
Abstract
from arXiv · showhide
Adolescent depression, anxiety, and stress (D/A/S) call for scalable tools that complement, rather than replace, professional evaluation. Under a privacy-preserving policy, the AdoDAS Grand Challenge withholds minors' raw recordings and distributes anonymized audio-visual representations and ASR-derived text. Its 6,000 participants provide 24,000 segments across one scripted-reading and three open-response sessions. Two tracks assess multi-task binary D/A/S screening and ordinal prediction of 21 DASS-21 item responses. From 191 registrations, the final leaderboards included 95 eligible screening teams and 64 item-prediction teams. Audio-visual baselines achieved 0.4604 mean F1 and 0.2675 mean Quadratic Weighted Kappa; leading submissions reached 0.5921 and 0.2776. Representative systems emphasize cross-session modelling, temporal multimodal fusion, psychometric structure, and task-aware calibration.
1 Introduction
AdoDAS addresses constraints on adolescent D/A/S assessment while treating computational outputs as screening signals and research measurements, not diagnoses. It tests whether privacy-preserving representations can support complementary subscale screening and item-level prediction across varied sessions.
1 Introduction
- Existing benchmarks cover depression analysis and multimodal settings, but few combine multiple adolescent elicitation sessions, coarse and item-level DASS targets, and a release policy designed to reduce direct exposure of minors’ recordings.Raw voices and faces can reveal biometric identity even after obvious metadata are removed [30] [31] [32].Source: §1 · p. 2
- Subscale screening captures broad D/A/S risk, whereas ordinal item prediction retains symptom-specific variation and response order without implying diagnosis.The DASS-21 organizes 21 ordinal items into Depression, Anxiety, and Stress subscales.Source: §1 · p. 2
- The four sessions pair scripted reading, which constrains lexical content, with open responses that elicit personal narratives and spontaneous verbal and non-verbal behaviour.Participant-linked sessions support subject-level cross-session evaluation.Source: §1 · p. 2
- The representation-only release makes privacy–utility a benchmark question: which task-relevant signals remain useful when anonymized text and derived features replace minors’ raw recordings?The release reduces direct biometric exposure while testing what information survives the interface.Source: §1 · p. 2
- AdoDAS provides a subject-disjoint benchmark with standardized derived representations, anonymized ASR text, and hidden-test tracks for D/A/S screening and DASS-21 item prediction.It also reports participation, results, and representative high-performing methods.Source: §1 · p. 2
2 Challenge Design
AdoDAS pairs participant-disjoint adolescent data and complementary DASS targets with a representation-only release and hidden-test evaluation. Its design specifies what is released, what each track predicts, and how submissions are scored.
2.1 Dataset and Protocol
- The 6,000-participant cohort contributes 24,000 segments across one scripted reading and three open responses, split by participant 70/10/20 to prevent session-level leakage.Track A1 converts subscales to binary labels; A2 retains ordinal item responses, whose zero-response prevalence ranges from 54.40% for d09 to 84.47% for d21; Figure 1 summarizes split sizes and response distributions.Source: §2.1 · p. 2, Figure 1 · p. 3
2.2 Privacy-Preserving Release
- The release withholds raw audio, video, and identifiable frames, instead providing anonymized ASR text and audio-visual representations aligned to a common 25 Hz timeline.Derived text and representations may still retain identity, institutional, recording-condition, or demographic information, and results do not measure raw-signal performance ceilings or establish absence of re-identification risk.Source: §2.2 · pp. 2–3
2.3 Tracks and Evaluation
- A1 submissions provide three probabilities and are ranked by mean binary F1 at a fixed 0.5 threshold; A2 submissions provide integer predictions for 21 items and are ranked by mean item-level QWK.Mean AUROC breaks A1 ties, mean absolute error breaks A2 ties, and the scorer rejects invalid submissions before scoring.Source: §2.3 · p. 3
3 Baseline and Challenge Protocol
The organizer baseline fuses temporal audio and video features, then aggregates session representations for participant-level predictions. On the released features, audio-visual fusion outperformed either unimodal baseline on both primary metrics.
3 Baseline and Challenge Protocol
- The baseline projects heterogeneous feature groups to a common width, encodes each modality with residual dilated temporal convolutions, and fuses audio and video summaries per session.Masked mean pooling aggregates four sessions into a participant representation; attentive pooling uses validity masks with VAD and video-quality signals.Source: §3 · p. 3
- The A1 baseline uses three sigmoid outputs, while A2 uses 21 CORAL-style ordinal heads with validation-selected decoding parameters.The organizers fixed the split, input schema, submission validation, and scoring software across teams, and the public implementation provides an executable reference for the protocol.Source: §3 · pp. 3–4
- Audio-visual fusion improved mean F1 by 0.0511 and mean QWK by 0.0917 over video, which exceeded audio on both primary metrics.The baseline’s weakest selected-video target was Stress at F1 0.3067, versus 0.4390 for Depression and 0.4821 for Anxiety; these reference systems use frozen organizer-provided representations, not raw-signal fine-tuning.Source: §3 · p. 4, Table 1 · p. 4
4 Participation and Results
HITSZ-HLT led A1 screening with 0.5921 mean F1, while HEALTH FIRST led A2 item prediction with 0.2776 mean QWK; 105 teams entered the leaderboards.
Participation and Results
- 95 teams were eligible for A1 and 64 for A2, from 105 unique leaderboard teams; 54 entered both tracks.Among dual-track teams, the primary metrics had a Spearman correlation of 0.4429.Source: §4 · p. 4
- Eligible A1 scores had a median F1 of 0.4236, while A2 had a median QWK of 0.2126 and a compressed high-performing group.A1’s best F1 was 0.5921.Source: §4 · p. 4
- 0.5921 mean F1 for HITSZ-HLT led A1, while 0.2776 mean QWK for HEALTH FIRST led A2.A1 ranks one and two differed by 0.0067 F1, while the entire A2 top-five range was 0.0116 QWK.Source: §4 · p. 4
5 Top-Performing Methods and Lessons
Leading systems combine cross-session and multimodal approaches with ordinal-item modelling, psychometric structure, and calibration.
Top-Performing Methods and Lessons
- Strong systems paired cross-session and frame-level audio-visual interaction, ordinal-item query heads, or ASR-derived summaries with confidence-aware refinement.HITSZ-HLT used staged correction, HEALTH FIRST validation-selected experts, and Ritsumei multimodal evidence.Source: §5 · p. 4
- The strongest systems’ long-range linguistic structure and multimodal fusion are consistent with prior interview-based modelling.Source: §5 · p. 4
- Cross-session modelling, temporal multimodal fusion, psychometric links between item and subscale tasks, and calibration recur as lessons across strong systems.Missingness-aware aggregation matters for heterogeneous streams; calibration and post-processing are especially important for A2 under QWK and imbalanced ordinal labels.Source: §5 · p. 4
6 Discussion and Conclusion
AdoDAS reports stronger audio-visual baseline performance and higher best-submission scores across its two tracks. Its results are screening signals, and the benchmark’s privacy and generalization boundaries remain explicit.
Discussion and Conclusion
- 0.4604 mean F1 for the audio-visual baseline rose to 0.5921 for the best A1 system, while 0.2675 mean QWK rose to 0.2776 for the best A2 system.The A1 top two formed a distinct tier, while the A2 top five spanned only 0.0116 QWK.Source: §6 · p. 4
- The moderate cross-track correlation (ρ= 0.4429) and recurring modelling mechanisms suggest partially shared capability alongside task-specific bottlenecks, but do not establish causal attribution.Submissions were heterogeneous and lacked common ablations.Source: §6 · p. 4
- Results are screening research signals, not diagnoses; withholding raw recordings reduces direct exposure but is not a formal privacy guarantee.The participant split tests matched collection conditions, not transfer to unseen schools, cultures, devices, or clinical populations.Source: §6 · p. 4