We present AdoDAS, a privacy-preserving multimodal Grand Challenge
dataset for adolescent Depression, Anxiety, and Stress (D/A/S)
assessment. The benchmark combines a standardized reading passage
with open-ended interview prompts and provides labels from DASS-21,
including both subscale scores and item-level responses.
Adolescent mental health screening is critical for early identification
and timely intervention, yet scalable assessment remains challenging in
real-world practice. Depression, anxiety, and stress (D/A/S) are prevalent
during adolescence and frequently co-occur, while routine screening still
relies heavily on questionnaires and limited clinical resources.
Recent progress in multimodal affective computing suggests that audio-visual
behavioural cues during communication can provide signals complementary to
self-reports, motivating standardised, reproducible benchmarks for multimodal
D/A/S screening and assessment.
Despite growing interest, reproducible multimodal benchmarks for minors remain
scarce. A key barrier is privacy and ethics: releasing raw audio/video recordings
of children and adolescents can enable re-identification even after conventional
de-identification, making open data sharing difficult. In addition, existing
mental-health resources often differ in task definitions, data splits, and
evaluation protocols, which hinders fair comparison and may introduce subject
leakage.
We introduce AdoDAS, a feature-centric, privacy-preserving multimodal
Grand Challenge for adolescent D/A/S assessment with labels derived
from DASS-21. Data are collected using a tightly controlled school-based
protocol that combines a standardised reading passage and open-ended question
answering, balancing controlled elicitation and spontaneous expression while
improving comparability across participants and sites. To enable reproducible
research without exposing raw signals, AdoDAS withholds raw audio/video and
releases pre-computed acoustic descriptors, visual representations, and cross-modal
temporal metadatato support temporal modelling and multimodal fusion. In total,
AdoDAS includes 6,000 participants and 24,000 audio-video segments across four
sessions per participant.
The benchmark includes two tracks: (A1) multi-task binary classification for D/A/S
screening and (A2) DASS-21 item-level prediction for interpretable symptom modeling.
Both tracks use the same privacy-preserving feature inputs and official subject-disjoint
split protocol, with ranking generated by the official script for deterministic evaluation.
