Live AI Research System · OPERA-CT Architecture

AI-Powered
Respiratory
Triage System

Multimodal agentic AI combining cough acoustics, voice biomarkers, and symptom analysis for early respiratory disease detection using OPERA encoder and LangGraph orchestration.

Start Screening See How It Works
0%
Accuracy
0
Macro F1
0
AUROC
2–0s
Inference
🧬
📊
🔬
Live Analysis · OPERA-CT ● RUNNING
Cough Audio · 22 050 Hz
COPD Confidence
72%
Pneumonia Risk
18%
Voice Biomarker
61%
Symptom Index
55%
Triage: MODERATE · Refer within 48h · Confidence 84%
Scroll
88.4%
Classification Accuracy
0.91
AUROC Score
3
Disease Classes
2-tier
Agentic Pipeline
Core Capabilities

Three Signals.
One Intelligent Decision.

OPERA-CT fuses acoustic biomarkers, voice analysis, and clinical symptoms through a LangGraph agentic pipeline to deliver hospital-grade triage recommendations.

🎙️
Cough Acoustic Analysis
OPERA pre-trained encoder extracts 768-dim embeddings from cough audio. Detects COPD, pneumonia, and asthma patterns from a single 5-second recording.
OPERA Encoder
🔊
Voice Biomarker Profiling
MFCC, jitter, shimmer, and spectral features from sustained vowel phonation (/a/) track respiratory muscle fatigue and longitudinal decline.
Voice Biomarkers
📋
Symptom Scoring Engine
CAT-score–inspired multivariate symptom analysis covering dyspnea, wheezing, chest tightness, sleep quality, and energy levels with clinical thresholds.
CAT Score
🤖
LangGraph Agent Orchestration
Multi-agent pipeline with Tier 1 patient self-screen and Tier 2 doctor assessment nodes. Conditional routing with memory-aware state management.
LangGraph
📈
Longitudinal Risk Monitoring
3-signal fusion score tracks patient health trajectory over time. Drift detection alerts clinicians to deterioration before critical events.
Drift Detection
🏥
Dual-Portal Architecture
Separate patient self-screening and doctor review portals. Patients submit audio; doctors receive AI-assisted differential diagnoses with confidence scores.
2-Tier Triage
Pipeline Architecture

From Audio
to Clinical Decision

OPERA-CT's agentic pipeline processes raw audio and symptom inputs through a sequence of specialized AI agents, each contributing a confidence-weighted modality score.

OPERA Encoder LangGraph librosa SQLite Flask API GTX 1650 float16
01
Audio Ingestion & Preprocessing
Cough and vowel audio resampled to 22 050 Hz. Silence trimming, normalization, and format unification via soundfile/librosa.
02
OPERA Encoder Embedding
Pre-trained 768-dim acoustic embeddings fed into lightweight classification head (MLP). Runs on GTX 1650 with float16 quantization in ~1.2s.
03
Voice Biomarker Extraction
40 MFCCs, spectral centroid, jitter, shimmer, and HNR computed from vowel audio. Z-score normalized against patient baseline.
04
LangGraph Triage Agent
Rule engine fuses acoustic, voice, and symptom scores. Conditional graph routing determines severity tier and generates clinical reasoning.
05
Longitudinal Score & Referral
Session stored to SQLite. Drift detection compares against baseline. Final triage decision with referral urgency delivered to both portals.
Access Portal

Start Your Screening

Sign in to your existing account or create a new patient account to begin multimodal respiratory assessment.

🫁
RespiTriage AI
Multimodal Respiratory Screening System
Sign In
Create Account
For research purposes only · Not a certified medical device