Can I trust it?

Absolutely. Our AI GAD-7 screening demo is designed for transparency and scientific rigor. It is trained, calibrated, and internally validated on clinically collected and ethically approved datasets — not scraped consumer data or uncontrolled social media sources.

• Primary sources:
  • Bridge2AI-Voice (NIH Common Fund, 2025) — a large-scale, ethics-reviewed dataset on voice biomarkers for health.
  • DAIC-WOZ / Extended DAIC (E-DAIC) — clinically structured interview datasets underpinning much of today’s mental-health ML research.

---

Executive summary for HR, HSE, and Compliance buyers

• Purpose: Early well-being screening for stress or anxiety signals (GAD-7). Designed for organizational psychosocial risk management, not clinical diagnosis.
• Processing model: Entirely local (browser/device) inference. No media is uploaded or stored unless enabled administratively.
• Default footprint: No embedded media persistence. Results (scores/time-series) can be exported by users/admins in CSV/JSON for organizational dashboards.
• Compliance fit: Supports ISO 45003 psychosocial safety frameworks, aligning with GDPR / LGPD / UAE PDPL through purpose limitation, local compute, and explicit user control.
• Procurement readiness: A DPA template, detailed technical FAQ, and DPIA/ROPA entry guidance are available.

---

Datasets We Use and Why

Bridge2AI-Voice (v1.0 → v2.x, NIH Common Fund)

• Overview:
  Bridge2AI-Voice is part of the U.S. National Institutes of Health Bridge2AI initiative to create ethically sourced, multi-modal datasets linking voice with validated health and wellness markers.

  • v1.0 (Feb 2025): Released derived acoustic features (e.g., spectrograms, MFCCs, prosodic descriptors) with paired clinical metadata (questionnaires, outcomes).
  • v2.x (Aug–Sep 2025): Expanded recording sites, participant diversity, and introduced controlled-access raw audio under strict data use agreements.

• Ethics & methodology:

  • Multi-site IRB-approved data collection.
  • Consent-based linking with validated measures such as PHQ-9, GAD-7, and PANAS.
  • Focused on fairness, demographic parity, and explainability research.

• Relevance to our system:
  We train the voice-stress calibration layer using derived features—not raw audio—ensuring full privacy. Metrics include RMS energy fluctuations, pitch variance (jitter proxy), and talk ratio stability.
  The demo replicates these features locally in the browser, without sending or storing any recordings externally.

References:
• Bridge2AI Initiative – NIH Common Fund
• Bridge2AI-Voice on PhysioNet (v2.x)
• Controlled-access raw audio announcement (2025-09)

---

DAIC-WOZ and Extended DAIC (E-DAIC)

• Overview:
  The Distress Analysis Interview Corpus (DAIC-WOZ) comprises structured clinical interviews conducted by a virtual interviewer (“Ellie”) designed to elicit affective and linguistic markers of mental health.
  The Extended DAIC (E-DAIC) expands this for AVEC challenge benchmarking, supporting reproducibility in multimodal affective computing.

• Structure and contents:

  • Audio, video, and text transcription.
  • Self-reported and clinician-administered psychological scales (PHQ-8, GAD-7, PSS).
  • Ground-truth alignment with affect recognition (FACS-based facial action units such as AU6/AU12).
  • Ethics board approval (USC Institute for Creative Technologies).

• Relevance to our system:
  We benchmark the facial and head posture cues derived from these datasets and validate them with questionnaire anchors.
  Our live screening demo mirrors these features locally in-browser, including:

  • Facial affect ratios (smile tension / openness balance)
  • Brow and eye activity regions
  • Head stability as an indirect proxy for psychomotor agitation

References:
• USC ICT DAIC-WOZ portal and documentation
• E-DAIC in AVEC literature (2019–2025 editions)

---

Privacy, Data Flow, and Export Controls

• Local-only processing: Webcam and microphone input are analyzed locally. No audio or video is transmitted by default.
• Data retention: None by default. Only derived metrics are computed in memory.
• Optional exports: Users or admins can export anonymized score data for reporting.
• Privacy-by-design: Fully compliant with purpose limitation and data minimization principles in major data-protection regimes.
• Admin controls: Export enablement and retention duration are configurable per organization.

---

Accuracy and Validation Summary

• Phase-1 internal evaluation: Accuracy of approximately 82% (AUC ~0.81) differentiating elevated- from non-elevated-risk cohorts using cross-validation on clinical datasets.
• Cross-modal fusion: Our “Augmented Index” combines 70% questionnaire weighting (GAD-7) and 30% AI-derived cues.
• Limitations: Environmental lighting, camera quality, background noise, and accent variability can affect outputs.
• Next phase: Ongoing prospective trials across culturally diverse samples; results to be disclosed in technical notes upon completion.

---

Responsible Use and Compliance Framework

• ISO 45003 alignment:
  Designed for psychological health and safety management at work. The demo supports early detection and proactive referral mechanisms, without entering medical diagnostic territory.

• PDPL (UAE Federal Decree-Law No. 45 of 2021):

  • Full transparency of data collection purposes.
  • Localized processing to avoid international data transfer where unnecessary.
  • Support for subject rights: access, correction, restriction, and erasure.

• GDPR/LGPD equivalence:
  Compliance structure is interoperable with EU and Brazil privacy frameworks. Controller obligations—such as ROPA/DPIA documentation—are minimized through an edge-first deployment.

• Medical positioning:
  The system operates strictly as a screening tool to augment workplace well-being programs. It is not a diagnostic medical device, nor a substitute for clinical evaluation.

---

Business and Organizational Use Cases

• ISO 45003 implementation:
  Teams can perform short, repeatable well-being check-ins. Data is aggregated for trend reporting and early outreach.
• ESG metrics and reporting:
  Aggregated anonymized results feed into HR and sustainability dashboards.
• Early triage workflow:
  Elevated cases may trigger internal referrals or EAP (Employee Assistance Program) offers under HR governance.

---

Governance Summary

Parameter	Default Behavior	Notes
Raw media storage	None	No persistent recording
Data export	JSON/CSV only	Controlled by user/admin
Retention	Organizational policy	Short-term recommended
Processing site	Browser/device	No cloud AI inference
Legal tools	DPA/DPIA templates	Provided on request

---

Changelog

• 2025-10-20: Added Bridge2AI-Voice v2.x update, UAE PDPL, and expanded validation notes.
• 2025-02-04: Added v1.0 dataset reference and derived-feature clarification.

---

References

• Bridge2AI-Voice overview — b2ai-voice.org
• Bridge2AI-Voice on PhysioNet (v2.x) — physionet.org/content/b2ai-voice
• Bridge2AI v1.0 release summary — physionet.org/news/post/bridge2ai-voice
• Controlled-access raw audio notice — bridge2ai.org/2025/09
• DAIC-WOZ dataset portal — dcapswoz.ict.usc.edu
• DAIC-WOZ documentation (PDF) — dcapswoz.ict.usc.edu/WP-content/uploads/2022/02/DAICWOZDepression_Documentation.pdf
• E-DAIC AVEC studies — AVEC 2019–2025 challenge literature
• ISO 45003 standard — iso.org/standard/64283.html
• UAE PDPL portal — u.ae/en/about-the-uae/digital-uae/data/data-protection-laws