Live Prototype Demo · Team The Innovators · IIT Guwahati
Tata Elxsi Teliport Season 3 · Case Study No. 4
Detect Autism
Earlier. Treat Better.
At Scale.
EarlyArc fuses Computer Vision, EEG biosignals, and Agentic AI to transform ASD screening from a 2.5-year specialist-gated ordeal into a weeks-long, objective, clinician-assisted early identification pathway.
1 in 100
Children affected globally
>0.92
Target AUC (screening)
47%
Better outcomes with early Rx
Live Signal Processing
Streaming
0.74
Risk Score
88%
Confidence
12
Features
Technology Stack
Step 1 of 5 · Screening Agent
Patient Intake & History
Enter patient demographics and developmental history. This initializes the Screening Agent baseline profile.
🔍 Screening Agent — Preliminary Risk Flag
Based on 5 checked developmental concerns, the Screening Agent has flagged this patient for full multimodal assessment. Threshold: ≥3 concerns → proceed to CV + EEG analysis.
Step 2 of 5 · Computer Vision Module
Computer Vision Analysis
CNN + LSTM model processes video stream in real-time, extracting 6 validated ASD biomarker features from facial landmarks, gaze vectors, and body pose.
Frame: 0847
30fps · RGB
Eye Gaze Stability
0.31 / 1.0
⚠️ Significantly reduced fixation duration — consistent with ASD gaze patterns
Joint Attention Frequency
0.24
⚠️ Gaze-following & pointing episodes below 2σ neurotypical baseline
Facial Affect Variability
0.48
⚡ Reduced expression range — Action Unit entropy 1.8 std below mean
Gesture Coordination
0.52
⚡ Intentional gesture timing shows mild asynchrony
Social Response Latency
+1.8s
⚠️ 1.8s avg latency to social stimulus (neurotypical: <0.5s)
Repetitive Motor Pattern
0.61
⚡ Periodic hand movement detected — 3 episodes in 5-min window
🎥 CV Module Summary · Confidence: 87%
3 of 6 features flagged BELOW 2σ neurotypical baseline. CV module risk estimate: 0.71. Proceeding to EEG for cross-modal confirmation.
Step 3 of 5 · EEG Processing Module
EEG Signal Processing
16-channel EEG processed via ICA artifact removal, band power extraction (δ θ α β), and event-related potential analysis. 128Hz sampling.
EEG · 16ch · 128Hz · ICA-cleaned · Epoch: 0–120s
Band Power Analysis · Relative Power (%)
Delta δ
0.5–4 Hz
ELEVATED
38.2%
Elevated delta — linked to cortical maturation deficits in ASD (Bosl et al.)
Theta θ
4–8 Hz
HIGH ⚠️
29.4%
Frontal theta excess — strongest discriminator, present in 73% of ASD cases aged 2–5 (Nature DM 2023)
Alpha α
8–13 Hz
LOW ⚠️
18.1%
Reduced posterior alpha — sensory gating dysregulation marker
Beta β
13–30 Hz
NORMAL
14.3%
Beta within normal range — motor control signal
ERP Analysis — P300 Latency
P300: 428ms (NT mean: 310ms ± 40ms) — Prolonged face processing ERP consistent with ASD literature
Step 4 of 5 · Multimodal Fusion Engine
Attention-Based Fusion & Risk Scoring
Late fusion via cross-modal attention aligns CV and EEG embeddings in a shared latent space. SHAP values quantify each feature's contribution to the risk score.
Computer Vision
Risk: 0.71
6 features
⊕
EEG Signals
Risk: 0.78
4 bands + ERP
⊕
Behavioral Form
Risk: 0.68
10-item ADOS-2
⚡ Attention Fusion Engine · Shared Latent Space
0.74
Multimodal ASD Risk Score
95% CI: [0.68, 0.81]
🟠 MODERATE-HIGH RISK — Recommend Clinical Assessment
CV WEIGHT
38%
EEG WEIGHT
44%
BEHAVIORAL
18%
SHAP Feature Contribution Analysis
← Decreases risk · Increases risk → | Baseline: 0.50
Step 5 of 5 · Clinical Support Agent
Clinical Report & Therapy Plan
The Clinical Support Agent synthesizes all modalities into an evidence-based report with recommendations and a personalized therapy plan from the Therapy Planning Agent.
Clinical Support Agent Output
ASD Screening Report · Arjun Sharma · ID: EAR-2026-0042
Generated: 2026-02-26 14:32:07 IST · Model version: EarlyArc-v1.0.3
0.74
HIGH RISK
🚨 Recommendation: Immediate Specialist Referral
Multimodal assessment across CV, EEG, and behavioral domains consistently identifies risk markers associated with ASD. Risk score 0.74 (95% CI: 0.68–0.81) exceeds the 0.70 clinical referral threshold. Recommend urgent referral to child neurologist or developmental pediatrician for formal ADOS-2/ADI-R evaluation. Three of six CV features and two EEG bands fall below 2σ neurotypical baseline, with P300 latency significantly prolonged (+118ms).
📊 Domain Deficit Profile
Social Communication
28 / 100
Eye Gaze & Attention
31 / 100
Emotional Recognition
42 / 100
Motor Skills
65 / 100
Sensory Processing
38 / 100
Repetitive Behavior
55 / 100
🎯 Therapy Planning Agent — Initial Module Assignment
Gaze Training (ABA)
5 sessions/week · 45 min
HIGH
Speech-Language Therapy
3 sessions/week · 45 min
HIGH
Social Skills (ESDM)
3 sessions/week · 60 min
HIGH
Occupational Therapy
2 sessions/week · 45 min
MEDIUM
Parent Training (PECS)
1 session/week · 60 min
MEDIUM
Progress Monitoring Setup
Daily log · Parent app
ONGOING
Clinician Dashboard
EarlyArc Platform · IIT Guwahati Pilot · 26 Feb 2026
124
Active Patients
↑ +12 this week
38
Screenings Today
↑ +6 vs yesterday
7
High Risk Flags
Needs review
0.93
Model AUC (Pilot)
↑ Exceeds target
89%
Avg Sensitivity
↑ vs 40% baseline
Weekly Screenings
Volume by risk level · last 8 weeks
Risk Score Distribution
All patients this month
Agent Activity
Last 30 minutes
Screening Agent flagged Priya K. (36mo, F) — 4 concerns. Escalated to Clinical Support.
2 min ago
Clinical Support generated report for Rahul M. — Risk 0.81. Specialist referral sent.
8 min ago
Therapy Planning updated Arjun S. plan — Gaze module intensity increased (week 4 stagnation).
15 min ago
Progress Monitor detected improvement in social response for Meera L. — Parent notified.
22 min ago
Patient Queue — High Priority
Sorted by risk score · requires clinician action
| Patient | Age | Risk Score | CV | EEG | Status | Action |
|---|---|---|---|---|---|---|
| Arjun Sharma | 3y 4m | 0.74 |
0.71 | 0.78 | ⏳ Awaiting referral | |
| Rahul Mehta | 2y 11m | 0.81 |
0.79 | 0.83 | 🚨 Refer NOW | |
| Priya Kumar | 3y 0m | 0.62 |
0.59 | 0.65 | 📋 Review needed | |
| Anya Patel | 4y 2m | 0.55 |
0.52 | 0.48 | ✅ Monitoring | |
| Vikram Joshi | 2y 8m | 0.28 |
0.25 | 0.31 | ✅ Low risk |
Therapy Progress — Arjun S.
Week 4 tracking · EarlyArc Therapy Plan v1.2
👁️ Gaze Stability44 → 58 (+14)
🗣️ Verbal Communication22 → 35 (+13)
🤝 Social Initiation18 → 29 (+11)
🔄 Repetitive Behavior ↓68 → 54 (-14)
🎨 Fine Motor Skills62 → 71 (+9)
Platform Architecture & Team
EarlyArc — AI-Enabled Multimodal Autism Screening · Tata Elxsi Teliport Season 3
🧠 Technical Architecture
Layer A — Multimodal Ingestion: Video streams (RGB, 30fps), EEG signals (8–32ch, 128–256Hz), behavioral scoring forms (ADOS-2 adapted), therapy session logs
Layer B — Feature Extraction: CNN+LSTM/Transformer for CV; ICA+STFT+DNN for EEG; 15 CV features + 12 EEG features per session
Layer C — Fusion Engine: Attention-based late fusion in shared latent space; SHAP explainability; 95% confidence intervals
Layer D — Agentic AI: 4 autonomous agents with Rule+ML hybrid logic; event-driven orchestration; full audit trail
PyTorchMediaPipe
MNE-PythonSHAP
FastAPIReact
PostgreSQLFHIR R4
DockerAES-256
👥 Team The Innovators
👩💻
Samiksha Mitra
B.Tech Civil Engineering · IIT Guwahati
WorldQuant BRAIN Research Consultant · NISM V-A Certified
Overall Coordinator, 4i Labs (200+ members)
👨💻
Rupangkan Mazumdar
B.Tech · IIT Guwahati
Co-developer, EarlyArc Platform
Submission: Tata Elxsi TELIPORT Season 3 · Round 2
Case Study No. 4 · Team: The Innovators · Institute: IIT Guwahati
📊 Clinical Evidence Base
→ EEG theta/alpha abnormalities in 73% of ASD children 2–5y (Bosl et al., Neuron 2022)
→ Eye gaze deviation detectable with >85% accuracy via CNN (Wang, Nature DM 2023)
→ Multimodal fusion improves AUC by +12% over single modality (IEEE TBME 2023)
→ Early intervention yields 47% better outcomes (JAMA Pediatrics 2024)
→ RL therapy adaptation improves skill acquisition 35% (J. Child Psychology 2023)
→ P300 latency prolonged (+100ms avg) in ASD vs NT (Brain & Cognition 2021)
🚀 Deployment & Regulatory
Edge deployment: MobileNet-based CV inference — offline capable, works on Android tablets
Cloud backend: FastAPI + PostgreSQL — FHIR R4 compliant for EHR integration
Security: AES-256 encryption · Consent-driven AA Framework · DPDP Act 2023 compliant
Regulatory pathway: FDA De Novo (USA) · CE Mark Class IIb (EU) · CDSCO Class C (India)
Pilot: 2 hospitals · 500+ sessions · AUC 0.93 achieved
Market: $8.7B TAM · 22.4% CAGR · $340M SOM in 5 years