Voice Biomarker Research Protocol

Acoustic detection of aspiration risk and dysphagia screening — HKEX IFS 2026 funded programme

RESEARCH-STAGED — Not for clinical use

Formal Research Protocol (White Paper)

The full CareEZ Voice Biomarker Research Protocol v1.0 white paper is available for download. It covers hypotheses, study design (N=500, 5 RCHEs), biomarker panel, statistical analysis plan, ethics, timeline, and budget.

↓ Download Protocol PDF (v1.0, 2026-05-25)   Markdown source

Status: Pre-funding stub. Y1 formal study commences Q1 post-HKEX IFS award. Author: CareEZ Research (Y0 by Raymond Chau + AI; Y1 lead by Clinical Advisor, TBC).

Overview

CareEZ is developing a non-invasive, audio-based dysphagia screening module that analyses voice and swallowing sounds to detect biomarkers associated with aspiration risk. This page documents the research design, planned biomarker panel, and current integration status under the HKEX IFS 2026 funded programme.

The voice biomarker module complements the existing Snap-to-IDDSI vision classifier and EAT-10 symptom screener to form a multimodal dysphagia risk pipeline.

Research Rationale

Silent aspiration — food or liquid entering the airway without triggering cough — affects an estimated 40–70% of post-stroke patients and up to 50% of nursing home residents in Hong Kong (Cichero et al., 2017; HKCSS Care-Food White Paper, 2023). Current gold-standard assessment (VFSS / FEES) requires specialist equipment and hospital access, creating a diagnostic gap in community and residential care settings.

Acoustic biomarkers derived from voice and swallowing sounds offer a low-cost, non-invasive first-line screening pathway that can be administered via smartphone.

Planned Biomarker Panel

1. Voice Quality Index (VQI)

Mel-spectrogram analysis of sustained phonation (/a/, /i/) to detect jitter, shimmer, and harmonics-to-noise ratio (HNR) changes associated with laryngeal pathology and reduced glottic closure — a predictor of penetration/aspiration events.

2. Swallow Count & Timing

Acoustic event detection of cervical auscultation sounds during swallowing trials (standardised water swallow test protocol). Counts swallow events and measures swallow latency as proxy for pharyngeal phase delay.

3. Cough Acoustics

Classification of voluntary vs. reflexive cough signatures. Absent or attenuated cough reflex is a known aspiration risk factor. Cough peak expiratory flow proxies derived from microphone signal amplitude.

4. Wet Voice Score (WVS)

Spectral analysis of post-swallow vocalisation for formant turbulence and low-frequency resonance patterns associated with pharyngeal pooling and laryngeal penetration ("wet/gurgly" voice quality). Validated against FEES findings in published literature.

Programme Timeline (HKEX IFS 2026)

Phase Period Milestone Status
Pre-study 2025 Q4 – 2026 Q1 IRB application; recording protocol finalisation; pilot microphone calibration study (n=20) Planning
Y1 — Data Collection 2026 Q2 – 2026 Q4 Recruit 150 participants (50 healthy controls, 50 dysphagia mild-moderate, 50 severe); FEES-correlated audio corpus; annotation pipeline Planned
Y1 — Model Training 2026 Q3 – 2027 Q1 Train convolutional + attention-based classifier on annotated corpus; internal validation; benchmark against EAT-10 AUC Planned
Y2 — Clinical Pilot 2027 Q1 – 2027 Q3 Prospective pilot (n=100) in 2 HK residential care homes; SLP-concurrent assessment; sensitivity / specificity vs. VFSS ground truth Planned
Y2 — API Integration 2027 Q2 – 2027 Q4 Replace stub endpoint with live inference; regulatory review (MDCO HK); commercial deployment planning Planned

Current Integration Status

To validate the clinical workflow integration (upload path, CORS handling, content-type negotiation, file-size limits) before the acoustic model is trained, CareEZ has shipped a live stub endpoint at:

POST https://www.seniordeli.com/api/iddsi-classify-audio

The stub accepts multipart/form-data (field: audio) or application/json (field: audio_base64), validates content type (mp3, m4a, wav, ogg, webm), enforces a 10 MB limit, and returns a structured research_staged: true response with the planned biomarker panel keys set to "research-pending".

This pattern — ship the contract first, fill in the model later — follows API-first design and allows frontend/clinical-IT integration to begin before the acoustic model exits the training pipeline.

Sample response

{ "research_staged": true, "audio_mode": "metadata-stub", "audio_received_bytes": 10240, "duration_seconds_estimate": null, "biomarker_panel": { "voice_quality": "research-pending", "swallow_count": "research-pending", "cough_count": "research-pending", "wet_voice_score": "research-pending" }, "stub_disclosure": "...", "recommendation": "For aspiration risk screening today, use /api/voice-aspiration-screen", "research_protocol_url": "https://careez.org/research/voice-biomarker-protocol" }

Screening Available Today

While acoustic biomarker analysis is research-staged, clinicians and care workers can use:

Ethics & Safety Commitments

Research disclaimer: This programme is at the data-collection and model-design stage. The stub API endpoint does not perform any acoustic analysis. No output from this system constitutes medical advice or clinical diagnosis. All clinical decisions must be made by a qualified speech-language pathologist following a formal swallowing assessment. Programme milestones are subject to HKEX IFS 2026 funding approval and IRB clearance.

Contact

For collaboration inquiries, research data sharing agreements, or clinical pilot participation: [email protected]

Carewells Limited · Hong Kong · Social Enterprise Registration pending