Al-Assisted Early Detection System for Osteoarthritis (OA) Risk Markers in North Eastern Region (NER)

1. Executive Snapshot

Osteoarthritis (OA) is a highly prevalent, debilitating musculoskeletal disorder that disproportionately impacts rural, elderly, and agrarian populations in India's North Eastern Region (NER) due to terrain demands, heavy physical labor, and severe shortages of orthopaedic specialists [PS] [KNOWN]. Existing diagnostic pathways rely on costly hospital-based X-rays, MRI scans, and specialist evaluations that are largely inaccessible at the primary healthcare level in hilly, remote hamlets [PS] [RESEARCHED]. This project proposes ArthriSense-NER, a low-cost, portable, hardware-assisted, and offline-first AI screening solution [INFERENCE] [ASSUMPTION]. ArthriSense-NER combines a low-cost Bluetooth Low Energy (BLE) Inertial Measurement Unit (IMU) knee strap with mobile computer vision (MediaPipe Pose) and a digital clinical questionnaire (WOMAC/KOOS) to capture joint kinematics, gait symmetry, and mobility indicators in field conditions [INFERENCE]. Operating completely offline on standard Android smartphones used by Accredited Social Health Activists (ASHAs) and Auxiliary Nurse Midwives (ANMs), the system computes an explainable OA Risk Score and severity indication, generates ABDM-compliant FHIR R4 screening records, and syncs automatically with district health dashboards upon re-establishing connectivity [INFERENCE].

Metadata Table

Fundamental System Objective: To empower frontline healthcare workers in remote, low-resource regions of NER with a portable, offline-capable hardware-software screening tool that objectively identifies early Osteoarthritis risk markers through multimodal gait, joint kinematic, and symptom analysis before structural joint damage becomes irreversible [PS] [INFERENCE].

2. Problem Deconstruction

The problem statement provided by MDoNER requires an AI-assisted, portable screening system specifically designed for the geographical and infrastructural realities of the North Eastern Region [PS]. Below is the atomic requirement deconstruction:

3. Stakeholders & User Needs

4. Requirements Traceability

5. Constraints & Environment Realities

[NER Field Operational Environment]
 ├── Geography: Rugged, mountainous terrain; isolated hamlets (Arunachal, Meghalaya, Nagaland, Mizoram)
 ├── Connectivity: 0 to 2G intermittent signal in hills; broadband restricted to urban centers
 ├── Power: Frequent power grid outages (6-12 hours/day); require battery/power bank operation
 ├── Language: High linguistic diversity (Assamese, Bengali, Bodo, Khasi, Garo, Mizo, Meitei, Nagamese)
 ├── Digital Literacy: Low among elderly patients; moderate (basic smartphone usage) among ASHA/ANM workers
 └── Field Conditions: High humidity, rain, dust; screening performed in open village community halls/camps

1. Geographical & Terrain Isolation: The North Eastern Region features dense forest coverage and mountainous topography [KNOWN]. Transportation between rural hamlets and District Hospitals takes 4–10 hours [RESEARCHED]. Screening tools must fit in a backpack and weigh less than 1.5 kg [INFERENCE] [ASSUMPTION].
2. Connectivity Deficit: Cellular connectivity in rural NER is highly inconsistent, with frequent total signal blackouts in hill valleys [RESEARCHED]. Cloud-dependent AI architectures (e.g., streaming video to AWS/GCP) will fail completely [INFERENCE]. 100% of diagnostic inference must occur locally on the smartphone [INFERENCE].
3. Power Reliability: Rural Primary Health Sub-Centres frequently experience prolonged power outages [RESEARCHED]. Hardware devices must operate on rechargeable Li-ion batteries providing at least 10 hours of continuous field screening, chargeable via standard USB-C/power banks [INFERENCE] [ASSUMPTION].
4. Linguistic Diversity: Over 220 languages/dialects are spoken across the 8 NER states [RESEARCHED]. Assamese, Bengali, Bodo, Khasi, Garo, Mizo, and Meitei represent major regional blocks [RESEARCHED]. Text-only interfaces are insufficient; audio instructions in native dialects are mandatory [INFERENCE].
5. Low Digital Literacy: Rural elderly citizens cannot navigate complex touchscreen apps [KNOWN]. ASHA workers must act as proxy operators, guiding the patient through physical movements while the app provides step-by-step visual and verbal cues [INFERENCE].

6. Research Questions

1. Kinematic Accuracy: How closely do smartphone-based MediaPipe Pose estimation and single-strap BLE IMU sensors correlate with gold-standard optical motion capture (Vicon/OptiTrack) for measuring knee flexion/extension and gait asymmetry in field settings?
2. Clinical Correlation: Can an ensemble ML model trained on spatiotemporal gait parameters and WOMAC scores reliably differentiate early-stage OA (Kellgren-Lawrence Grade 1–2) from healthy control joints without immediate radiologic imaging?
3. Hardware BOM Optimization: What is the minimal viable sensor configuration (e.g., dual 6-axis MPU6050 vs single 9-axis ICM-20948) that maintains acceptable knee joint angle measurement accuracy (< 3° error) while keeping total kit cost under ₹2,500?
4. Offline Sync & Conflict Resolution: What data synchronization protocol guarantees zero data loss or duplication when 500+ ASHA workers upload cached patient reports simultaneously after weeks of offline field operation?
5. ABDM Compliance: How can FHIR R4 JSON bundles containing raw IMU kinematic time-series and CV pose vectors be structured to comply with National Health Authority (NHA) digital health record standards?

7. Data Requirements & Sources

Data Classification Categories:

• A. Data Explicitly Mentioned in PS: Joint movement angles, gait patterns, posture assessments, pain scores, mobility screening inputs, medical imaging/sensor data, digital patient symptom reports [PS].
• B. Data Required by Inference: Patient demographics (age, gender, BMI), occupational physical intensity (farming, hill carrying), historical joint trauma, regional language audio assets [INFERENCE].
• C. Data That Must Be Collected: Field validation IMU/CV paired datasets gathered from volunteers/patients during pilot trials [INFERENCE].
• D. Data That Can Be Synthetically Generated: Simulated high-volume offline patient screening records for stress-testing database synchronization and backend queue performance [INFERENCE] [ASSUMPTION].
• E. Data That Still Needs Validation: Normative gait and joint ROM baseline values specific to tribal/hill populations in NER performing high-grade gradient walking [NEEDS VALIDATION].

8. Existing Solutions & Gap Analysis

[Existing Solution Landscape]
 ├── Marker-based MoCap (Vicon): Gold standard ($50k+), laboratory restricted ──> GAP: Inaccessible in rural NER
 ├── OpenCap / Video MoCap: Smartphone camera CV ──> GAP: Fails in poor lighting, loose clothing, or non-standard angles
 ├── Wearable IMU Systems (APDM): High accuracy ($5k+) ──> GAP: Extremely expensive, complex setup for field workers
 └── Paper Questionnaires (WOMAC): Subjective, recall bias ──> GAP: No objective kinetic data, manual record management

The Unmet Gap:

No current solution provides an integrated, low-cost (< ₹2,500) hardware-software ecosystem that combines BLE IMU sensor kinematics with smartphone computer vision, runs 100% offline edge ML, supports multilingual NER audio workflows, and outputs ABDM FHIR R4 digital health records directly from rural health camps [PS] [INFERENCE].

9. Candidate Solution Approaches

[Evaluated Architectural Approaches]
 ├── Approach 1: Pure Mobile Computer Vision App (Camera-only pose estimation + WOMAC)
 ├── Approach 2: Hybrid Hardware-Software Ecosystem (BLE IMU Knee Strap + Vision CV + Edge ML)  <-- [SELECTED]
 └── Approach 3: Cloud-Centric X-Ray Diagnostic AI Platform (Deep Learning Radiograph Analysis)

10. Recommended Approach & Why

Selected Solution: Hybrid BLE IMU Motion Strap + Smartphone Computer Vision + Edge ML Risk Classifier (Approach 2)

Why Approach 2 Wins:

1. Fulfills the Hardware Category Mandate: Problem Statement SIH26004 is explicitly listed under the Hardware category and requires a "portable assessment interface or sensor-assisted screening mechanism" [PS]. Approach 1 (software-only) violates the category mandate and fails under low-light/traditional clothing conditions in field camps [INFERENCE].
2. Dual-Sensor Complementarity: IMU sensors measure precise micro-rotations, angular velocity, and flexion angles directly on the joint without occlusion issues [1.1, 1.5]. Computer Vision captures macro-body posture, walking speed, cadence, and step symmetry [2.2, 2.10]. Combining both inputs yields a clinical accuracy profile exceeding 88% sensitivity for early OA indicators [1.1, 2.8].
3. 100% Offline Capability: All data processing—from BLE telemetry parsing and MediaPipe pose estimation to LightGBM ML inference and report compilation—occurs locally on an off-the-shelf $120 Android phone without needing cloud connectivity [INFERENCE].
4. Extreme Affordability: The total hardware Bill of Materials (BOM) cost is under ₹2,000 ($24), allowing state health departments to equip every ASHA worker across NER at negligible cost [INFERENCE] [ASSUMPTION].

Rejection of Alternatives:

• Approach 1 (Pure Computer Vision App) Rejected: High error rates in knee flexion angle estimation when patients wear traditional loose attire (e.g., Mekhela Chador in Assam, Phanek in Manipur) or when screening in poorly lit village huts [INFERENCE]. Software-only approach fails the SIH Hardware category criteria [PS].
• Approach 3 (Cloud X-Ray AI Platform) Rejected: Over 85% of rural primary sub-centres in NER lack functioning diagnostic X-ray infrastructure [PS] [RESEARCHED]. Furthermore, transmitting heavy radiological files over 2G network connections in mountain districts is infeasible [INFERENCE]. X-rays only detect structural changes in moderate-to-late OA (KL Grade 2+), missing early functional/kinematic risk markers [KNOWN].

11. Technical Architecture

==================================================================================================
                                ARTHRISENSE-NER SYSTEM ARCHITECTURE
==================================================================================================

[1. HARDWARE & INGESTION LAYER - FIELD DEPLOYMENT]
 ┌─────────────────────────────────────────┐       ┌──────────────────────────────────────────┐
 │ Dual-IMU Knee Motion Strap (Hardware)  │       │ Smartphone RGB Camera (30 FPS)           │
 │ - ESP32-C3 Microcontroller (BLE 5.0)   │       │ - Monocular walking video capture        │
 │ - 2x MPU6050 6-Axis Accelerometer/Gyro │       │ - Guided 5-meter gait assessment track   │
 │ - 3.7V 800mAh LiPo Battery (12hr run)  │       │ - Real-time lighting quality check       │
 └────────────────────┬────────────────────┘       └────────────────────┬─────────────────────┘
                      │ BLE 5.0 Stream (50 Hz CSV Packets)              │ Frame Buffer
                      └────────────────────┐  ┌─────────────────────────┘
                                           ▼  ▼
[2. MOBILE EDGE PROCESSING LAYER - REACT NATIVE / ANDROID (OFFLINE)]
 ┌────────────────────────────────────────────────────────────────────────────────────────────┐
 │  ON-DEVICE DATA FUSION & FEATURE EXTRACTION ENGINE                                          │
 │  ├── BLE Packet Parser & Quaternion Goniometer (Calculates exact knee flex/ext angle)     │
 │  ├── MediaPipe Pose Engine (TFLite/ONNX - Tracks 33 2D/3D skeletal keypoints)             │
 │  ├── Kinematic Feature Extractor:                                                          │
 │  │   └─ ROM, Angular Velocity, Stance/Swing Ratio, Cadence, Gait Speed, Asymmetry Index    │
 │  └── Multilingual Clinical Questionnaire (WOMAC 10-item icon scale in 8 NER languages)      │
 └─────────────────────────────────────────┬──────────────────────────────────────────────────┘
                                           │ Formatted Feature Vector
                                           ▼
[3. EDGE AI INFERENCE & LOCAL STORAGE LAYER]
 ┌────────────────────────────────────────────────────────────────────────────────────────────┐
 │  ON-DEVICE LIGHTGBM RISK & SEVERITY CLASSIFIER (ONNX Runtime)                              │
 │  ├── Inputs: [IMU Kinematics + CV Gait Features + WOMAC Score + Age + BMI + Occupation]   │
 │  ├── Outputs: OA Risk Score (0-100%), Severity Tier (Low/Mod/High), SHAP Feature Drivers   │
 │  └── Local Encrypted Storage (SQLite / WatermelonDB with SQLCipher - AES-256)             │
 └─────────────────────────────────────────┬──────────────────────────────────────────────────┘
                                           │ Encrypted JSON Packet / PDF Report
                                           ▼
[4. DATA SYNCHRONIZATION & ABDM INTEGRATION LAYER (UPON CONNECTIVITY)]
 ┌────────────────────────────────────────────────────────────────────────────────────────────┐
 │  OFFLINE DELTA SYNC ENGINE (Vector Clocks & CRDT Conflict Resolution)                     │
 │  ├── ABDM Integration Gateway: FHIR R4 Bundle Generator (DiagnosticReport & Observation)   │
 │  └── Secure TLS 1.3 Upload to Cloud Backend (FastAPI / PostgreSQL)                        │
 └─────────────────────────────────────────┬──────────────────────────────────────────────────┘
                                           │ HTTPS / REST APIs
                                           ▼
[5. CENTRAL CLOUD & ANALYTICS DASHBOARD - MDONER & HEALTH OFFICERS]
 ┌────────────────────────────────────────────────────────────────────────────────────────────┐
 │  MDONER DISTRICT HEALTH DASHBOARD (React.js Web App)                                        │
 │  ├── GIS Heatmaps: District & Block-level OA Risk Prevalence in NER                        │
 │  ├── PHC Triage Queue: High-risk patient referral management for Orthopaedists             │
 │  └── Supply Chain & Camp Analytics: ASHA performance, battery stats, equipment tracking   │
 └────────────────────────────────────────────────────────────────────────────────────────────┘
==================================================================================================

Step-by-Step Data Flow Narrative:

1. Patient Registration: ASHA worker launches the offline mobile app in a village health camp, selects the local language (e.g., Khasi), and enters basic demographic data (age, gender, BMI, occupation) or scans the patient's ABDM QR code to pull their 14-digit ABHA ID [INFERENCE].
2. Hardware Pairing & Calibration: The worker straps the lightweight ESP32 BLE IMU unit onto the patient's knee. The mobile app automatically pairs via BLE 5.0 and performs a 3-second static standing calibration to establish the baseline zero-angle position [INFERENCE].
3. Multimodal Screening Protocol:
   • Active ROM Test: Patient performs 3 active knee extension/flexion movements. Dual MPU6050 sensors stream raw acceleration and angular velocity data at 50 Hz to the phone. The BLE parser calculates precise joint angle changes ($\Delta \theta$) [INFERENCE].
   • 5-Meter Gait Test: Patient walks 5 meters along a straight path. Smartphone camera records video. MediaPipe Pose extracts skeletal keypoints (hip, knee, ankle) in real time to calculate gait speed, cadence, stance-phase symmetry, and dynamic knee valgus [2.10, 4.5].
   • Digital WOMAC Questionnaire: ASHA worker administers a 10-question visual/voice questionnaire capturing pain, stiffness, and daily function difficulty [1.1].
4. On-Device Edge AI Scoring: The feature extraction engine computes 18 kinematic and clinical metrics. The local ONNX LightGBM model processes the feature vector in < 150 milliseconds, generating an OA Risk Probability (0–100%), Severity Tier (Low, Moderate, High), and SHAP attribution breakdown (identifying primary risk drivers) [INFERENCE].
5. Report Generation & Local Storage: The app generates a structured PDF report and localized voice explanation for the patient. The complete record is encrypted (AES-256) and saved to local SQLite storage [INFERENCE].
6. Cloud Sync & ABDM Publishing: When the ASHA worker returns to a PHC with mobile connectivity, the Delta Sync Engine pushes encrypted records to the FastAPI backend and publishes ABDM-compliant FHIR R4 bundles to the National Health Authority (NHA) gateway [INFERENCE] [RESEARCHED].

12. Technology Stack

13. AI/ML Justification

Model 1: MediaPipe Pose Kinematic Feature Extractor

1. Exact Problem Solved: Tracks lower-limb skeletal joints (Hip, Knee, Ankle keypoints) from smartphone walking video to extract dynamic gait parameters [2.10, 4.5].
2. Input: 1080p / 720p 30 FPS monocular video frame buffer [2.10].
3. Output: 3D coordinates $(x, y, z)$ for keypoints 23 (Left Hip), 24 (Right Hip), 25 (Left Knee), 26 (Right Knee), 27 (Left Ankle), 28 (Right Ankle) with confidence scores [4.1, 4.10].
4. Why Simpler System Fails: Manual frame-by-frame angle measurement (e.g., Kinovea) takes 15 minutes per patient and requires expert biomechanists [2.5, 4.7].
5. Data Required: 2D/3D video frames across diverse body shapes and rural clothing [4.7].
6. Evaluation Metric: Mean Absolute Error (MAE) in knee angle degrees compared to optical motion capture (target MAE < 4.5°) [2.5, 2.10].
7. When Wrong: False keypoint detection under heavy skirts/pants. Mitigated by confidence thresholding ($> 0.7$) and primary reliance on BLE IMU readings [INFERENCE].
8. Training Required: Pre-trained COCO/BlazePose weights fine-tuned on lower-limb gait trajectories [INFERENCE].
9. Fine-Tuning: Not required for MVP; pre-trained MediaPipe Pose model achieves high relative reliability (ICC > 0.89) [4.7].
10. RAG: No [INFERENCE].
11. Computer Vision: Yes (Object Detection & Pose Estimation) [2.10, 4.5].
12. Speech Processing: No [INFERENCE].
13. Forecasting: No [INFERENCE].
14. Human Oversight: ASHA worker reviews skeleton overlay on screen before submitting video analysis [INFERENCE].

Model 2: LightGBM Early OA Risk & Severity Classifier

1. Exact Problem Solved: Predicts OA risk probability and severity tier by fusing joint ROM, gait speed, asymmetry index, age, BMI, and WOMAC pain score [1.1, 2.8].
2. Input: 18-element normalized feature vector [IMU ROM, Peak Flexion, Stance Time Ratio, Cadence, Gait Speed, Asymmetry Index, WOMAC Pain, WOMAC Stiffness, WOMAC Function, Age, Gender, BMI, Occupational Class] [INFERENCE].
3. Output: OA Risk Index (0–100%), Categorical Risk Tier (Low / Moderate / High), SHAP Feature Importance Values [INFERENCE].
4. Why Simpler System Fails: Fixed clinical threshold rules (e.g., "if WOMAC > 15 then high risk") miss non-linear combinations (e.g., moderate pain + severe gait asymmetry in a 45-year-old farmer indicates early structural degradation) [1.1, 2.8].
5. Data Required: OAI NIH tabular dataset (4,796 patients) merged with sensor-based knee gait datasets [1.2, 1.3].
6. Evaluation Metric: Area Under ROC Curve (AUC-ROC > 0.88), Sensitivity (> 85%), Specificity (> 82%) [1.1, 2.8].
7. When Wrong: False positive over-refers to PHC; false negative misses early intervention. System explicitly displays confidence intervals and mandates doctor review [INFERENCE].
8. Training Required: Yes (Supervised classification on historical cohort data) [1.1, 2.8].
9. Fine-Tuning: Periodic retraining when field validation data is uploaded to central servers [INFERENCE].
10. RAG: No [INFERENCE].
11. Computer Vision: No (Tabular Gradient Boosting) [INFERENCE].
12. Speech Processing: No [INFERENCE].
13. Forecasting: No [INFERENCE].
14. Human Oversight: Essential. Results are flagged as "Preliminary Risk Screening" requiring PHC Medical Officer verification [PS] [INFERENCE].

Where AI Should NOT Be Used:

• No AI for Final Clinical Diagnosis: AI must NEVER issue an absolute medical diagnosis of Osteoarthritis. Final diagnosis requires clinical examination and radiological correlation by a licensed physician [PS] [INFERENCE].
• No AI for Patient Consent: Consent for screening and ABDM health record linking must be obtained via explicit user interaction, never automated AI decision-making [3.4, 4.6].
• No Generative AI for Medical Advice: Prevention advice must be retrieved from deterministic, pre-approved medical templates validated by orthopaedic specialists, avoiding LLM hallucination risks [INFERENCE].

14. Data Architecture & Offline/Low-Network Design

==================================================================================================
                                LOCAL SQLITE SCHEMA & OFFLINE SYNC
==================================================================================================

┌───────────────────────────┐       ┌───────────────────────────┐       ┌───────────────────────────┐
│     PATIENTS TABLE        │       │    SCREENINGS TABLE       │       │    IMU_GAIT_DATA TABLE    │
├───────────────────────────┤       ├───────────────────────────┤       ├───────────────────────────┤
│ patient_id (UUID v4) PK   │1    N │ screening_id (UUID v4) PK │1    1 │ gait_id (UUID v4) PK      │
│ abha_number (VARCHAR 14)  ├───────┤ patient_id (UUID) FK      ├───────┤ screening_id (UUID) FK    │
│ name (TEXT ENCRYPTED)     │       │ timestamp (DATETIME)      │       │ imu_rom_flexion (FLOAT)   │
│ age (INTEGER)             │       │ womac_pain_score (INT)    │       │ imu_rom_extension (FLOAT) │
│ gender (VARCHAR 10)       │       │ womac_stiffness_score(INT)│       │ cv_gait_speed (FLOAT)     │
│ bmi (FLOAT)               │       │ oa_risk_score (FLOAT)     │       │ cv_cadence (FLOAT)        │
│ occupation_type (TEXT)    │       │ risk_tier (VARCHAR 10)    │       │ asymmetry_index (FLOAT)   │
│ created_at (DATETIME)     │       │ sync_status (INTEGER)     │       │ raw_imu_csv_blob (BLOB)   │
└───────────────────────────┘       └───────────────────────────┘       └───────────────────────────┘
                                                  │
                                                  │1
                                                  │
                                                  │N
                                    ┌───────────────────────────┐
                                    │    SYNC_QUEUE TABLE       │
                                    ├───────────────────────────┤
                                    │ queue_id (INTEGER) PK     │
                                    │ payload_json (TEXT)       │
                                    │ retry_count (INTEGER)     │
                                    │ last_attempt (DATETIME)   │
                                    └───────────────────────────┘
==================================================================================================

Local Offline Data Persistence:

• Local Storage Engine: WatermelonDB backed by SQLite with SQLCipher AES-256 database encryption [INFERENCE].
• Sync Status Flags: Every local record contains a sync_status flag: 0 (Unsynced), 1 (Syncing), 2 (Synced), 3 (Conflict) [INFERENCE].

Synchronization & Conflict Resolution Protocol:

1. Vector Clock Timestamping: Every record mutation is tagged with a client UUID and monotonically increasing sequence number [INFERENCE].
2. Delta Compression: To minimize bandwidth over 2G connections, the app transmits compressed JSON delta payloads containing only unsynced screening records (sync_status = 0) [INFERENCE].
3. Conflict Handling Rule: If a patient record is modified on both client and server, the server applies Last-Write-Wins (LWW) based on NTP-validated timestamps, storing superseded versions in an audit table [INFERENCE].
4. Exponential Backoff Retry Strategy: Failed upload requests retry at intervals of $2^n$ seconds (5s, 10s, 20s, 40s, max 15 mins) to conserve phone battery during network outages [INFERENCE].
5. Bandwidth Reduction: Raw 50 Hz IMU CSV telemetry and MediaPipe video frames are processed locally into summary metrics (~2 KB JSON packet per screening). Raw video files are discarded immediately after feature extraction to respect phone storage limits [INFERENCE].
6. Offline UX Indicators: A persistent sync status banner informs the ASHA worker: "12 Screenings Saved Offline - Will Auto-Upload When Network Available" [INFERENCE].

15. Security, Privacy & Governance

1. Personal Data & PII Protection: Patient names, phone numbers, and addresses are encrypted at rest using AES-256-GCM encryption keys derived from Android Keystore System [INFERENCE]. The app strictly complies with India's Digital Personal Data Protection (DPDP) Act 2023 [4.6].
2. Location Data Governance: GPS coordinates recorded during screening are blurred to town/village center level to protect patient privacy while providing sufficient spatial accuracy for MDoNER GIS heatmaps [INFERENCE].
3. Authentication & Access Control: Role-Based Access Control (RBAC) enforces strict privileges:
   • ASHA / ANM: Can create patient records, run screening protocols, and view local patient reports [INFERENCE].
   • PHC Doctor: Can review full diagnostic reports, confirm/override risk tiers, and issue official referrals [INFERENCE].
   • MDoNER Administrator: Accesses aggregated, anonymized district-level epidemiological analytics [INFERENCE].
4. Encryption Standards: All data in transit between mobile devices and central servers uses TLS 1.3 with Certificate Pinning to prevent Man-in-the-Middle (MitM) attacks [INFERENCE].
5. ABDM Consent Framework: Integrates NHA ABDM Consent Manager APIs [4.1, 4.6]. Patient explicit opt-in consent (captured via Aadhaar OTP or digital signature) is required before linking screening reports to their 14-digit ABHA ID [4.1, 4.6].
6. Regulatory Validation Disclaimer: The dossier explicitly labels the hackathon system as a "Prototype Decision Support Tool for Screening". It is not a certified Class B/C medical device under Indian Medical Devices Rules (MDR) 2017 until formal clinical trials and CDSCO approval are completed [INFERENCE].

16. MVP Scope (Hackathon Build)

==================================================================================================
                                    HACKATHON MVP SCOPE
==================================================================================================

  ┌──────────────────────────────────────────────────────────────────────────────────────────┐
  │ MUST BUILD (Core Value Proposition)                                                     │
  │  ├── ESP32 BLE Hardware Knee Strap streaming real-time joint angles                      │
  │  ├── React Native Mobile App with MediaPipe Pose Gait Analysis Engine                    │
  │  ├── Multilingual WOMAC Questionnaire (English, Assamese, Bengali)                       │
  │  ├── On-Device LightGBM OA Risk & Severity Classification Engine                         │
  │  ├── Offline-First SQLite Data Storage with Auto-Sync Engine                             │
  │  └── ABDM FHIR R4 DiagnosticReport JSON Generator                                        │
  └──────────────────────────────────────────────────────────────────────────────────────────┘
  ┌──────────────────────────────────────────────────────────────────────────────────────────┐
  │ SHOULD BUILD (High Impact Differentiators)                                              │
  │  ├── Voice-guided audio prompts for illiterate patient instruction                       │
  │  ├── MDoNER Web Dashboard with interactive district GIS heatmaps                         │
  │  └── Printable/Exportable PDF Screening Summary with SHAP feature explanations            │
  └──────────────────────────────────────────────────────────────────────────────────────────┘
  ┌──────────────────────────────────────────────────────────────────────────────────────────┐
  │ NICE TO HAVE (Optional Polish)                                                           │
  │  ├── Expanded language support (Khasi, Garo, Bodo, Mizo, Manipuri)                       │
  │  └── Automated exercise guidance video player in app                                     │
  └──────────────────────────────────────────────────────────────────────────────────────────┘
  ┌──────────────────────────────────────────────────────────────────────────────────────────┐
  │ DO NOT BUILD DURING HACKATHON (Described Architecture Only / Out of Scope)               │
  │  ├── Custom silicon PCB manufacturing (Use off-the-shelf ESP32 breakout boards)          │
  │  ├── Full CDSCO medical device regulatory certification                                  │
  │  └── Direct cloud DICOM X-ray PACS server integration                                    │
  └──────────────────────────────────────────────────────────────────────────────────────────┘
==================================================================================================

17. Live Demo Simulation Plan

To deliver a flawless, realistic live presentation for SIH judges without relying on live government server availability or physical patient presence:

[Live Demo Simulation Architecture]
 ┌────────────────────────────────┐      ┌────────────────────────────────┐
 │ Mock BLE Sensor Hardware       │      │ Pre-Recorded Patient Cohorts   │
 │ - ESP32 broadcasting real-time │      │ - Normal Gait (Healthy Control)│
 │   joint flex/ext telemetry     │      │ - Antalgic Gait (Moderate OA)  │
 └───────────────┬────────────────┘      └───────────────┬────────────────┘
                 │ BLE Stream                            │ Video Stream
                 └────────────────┐  ┌───────────────────┘
                                  ▼  ▼
 ┌────────────────────────────────────────────────────────────────────────┐
 │                    REACT NATIVE SCREENING MOBILE APP                   │
 ├────────────────────────────────────────────────────────────────────────┤
 │ 1. Live BLE IMU Knee Flexion Waveform Display                          │
 │ 2. Live Camera MediaPipe Skeleton Overlay & Angle Extraction           │
 │ 3. Real-Time LightGBM Risk Scoring & SHAP Attribution Rendering       │
 │ 4. Offline Mode Switch Toggle -> PDF Report Generation                 │
 │ 5. ABDM Sandbox API Sync -> Push FHIR Bundle to ABDM Health Locker     │
 └────────────────────────────────┬───────────────────────────────────────┘
                                  │ HTTPS
                                  ▼
 ┌────────────────────────────────────────────────────────────────────────┐
 │                    MDONER WEB ANALYTICS DASHBOARD                      │
 └────────────────────────────────────────────────────────────────────────┘

Step-by-Step 3-Minute Live Demo Script:

1. Phase 1: Hardware & Patient Setup (0:00 - 0:45): Demonstrate the physical ESP32 IMU knee strap. Strap it onto a team member's leg. Show real-time Bluetooth live angle telemetry ($0^\circ \rightarrow 110^\circ$ flexion) rendering dynamically on the React Native mobile app screen [INFERENCE].
2. Phase 2: Multimodal Gait & Symptom Capture (0:45 - 1:30): Point the mobile camera at the team member walking. Show MediaPipe Pose tracking hip-knee-ankle keypoints in real time, displaying walking speed ($1.1 \text{ m/s}$) and stance symmetry. Fill out 3 quick WOMAC pain sliders in Assamese UI [INFERENCE].
3. Phase 3: Edge AI Inference & Offline PDF Generation (1:30 - 2:15): Press "Analyze Risk". Show instant (< 200 ms) output: OA Risk Index: 74% (High Risk - Moderate OA Indicator). Display SHAP explanation: "Primary drivers: Reduced Knee Extension (-12°), High WOMAC Pain Sub-score". Toggle Phone Airplane Mode to ON, proving full offline functionality and PDF report generation [INFERENCE].
4. Phase 4: ABDM Sync & MDoNER Dashboard (2:15 - 3:00): Turn Airplane Mode OFF. App automatically triggers background sync. Show ABDM Sandbox API response accepting the FHIR R4 DiagnosticReport bundle. Switch to the MDoNER Web Dashboard showing the newly recorded screening appearing on the district map of East Khasi Hills, Meghalaya [INFERENCE].

18. Implementation Roadmap

Target Deadline: 30 September 2026 [PS] (Timeline sized for student development velocity).

==================================================================================================
                                IMPLEMENTATION ROADMAP (SIH 2026)
==================================================================================================

 [PHASE 1: RESEARCH & HARDWARE PROTO-TYPING] (Weeks 1-3 | Target: Oct 15)
 ├── Finalize ESP32-C3 firmware & MPU6050 Complementary Filter algorithm
 ├── Collect baseline IMU knee flexion datasets and benchmark against manual goniometer
 └── Set up React Native project template with WatermelonDB/SQLite

 [PHASE 2: VISION CV & EDGE ML DEVELOPMENT] (Weeks 4-6 | Target: Nov 15)
 ├── Integrate MediaPipe Pose TFLite engine into React Native Android build
 ├── Train LightGBM model on OAI NIH dataset + open gait kinematic datasets
 └── Export LightGBM model to ONNX format and integrate ONNX Runtime Mobile

 [PHASE 3: APP WORKFLOW, I18N & ABDM INTEGRATION] (Weeks 7-9 | Target: Dec 15)
 ├── Build multilingual UI engine (English, Assamese, Bengali) with audio guidance
 ├── Implement ABDM FHIR R4 JSON schema generator and Sandbox API client
 └── Build offline Delta Sync Engine with Vector Clock conflict resolution

 [PHASE 4: WEB DASHBOARD & SYSTEM INTEGRATION] (Weeks 10-11 | Target: Jan 15)
 ├── Develop FastAPI cloud backend and PostgreSQL/PostGIS database
 ├── Build React.js MDoNER Web Analytics Dashboard with Leaflet GIS maps
 └── Perform end-to-end integration testing (Hardware -> Mobile -> Cloud -> Web)

 [PHASE 5: FIELD TESTING & DEMO POLISH] (Weeks 12-14 | Target: Sep 30, 2026 [PS])
 ├── Conduct stress testing (1,000 simulated offline records sync)
 ├── Perform hardware battery endurance and field durability testing
 └── Finalize live demo pitch deck, simulation scripts, and video documentation
==================================================================================================

19. Testing & Validation Plan

20. Risk Register

21. Cost & Scalability

Hackathon Prototype Cost (BOM for 1 Hardware-Software Unit):

• ESP32-C3 Microcontroller Board: ₹350 ($4.20) [INFERENCE]
• Dual MPU6050 IMU Sensors: ₹240 ($2.90) [INFERENCE]
• 3.7V 800mAh LiPo Battery & TP4056 Charger: ₹220 ($2.65) [INFERENCE]
• 3D Printed Enclosure & Elastic Knee Strap: ₹250 ($3.00) [INFERENCE]
• Wiring, Switches, USB-C Port, Fasteners: ₹140 ($1.70) [INFERENCE]
• Total Hardware Kit Cost: ₹1,200 ($14.45) [INFERENCE] [ASSUMPTION]
• Cloud Hosting (Hackathon): ₹0 (Free tier Render/Supabase/Vercel) [INFERENCE].

Production Scale Deployment (1,000 PHC Units Across NER):

• Mass Hardware BOM (Bulk Sourcing): ₹850 ($10.20) per kit x 1,000 = ₹8,50,000 ($10,200) [INFERENCE] [ASSUMPTION].
• Cloud Infrastructure (AWS/NIC Cloud - 100,000 screenings/yr): Managed PostgreSQL + FastAPI ECS Containers = ~₹25,000 ($300) / month [INFERENCE] [ASSUMPTION].
• Maintenance & Replacement Reserve: 10% annual hardware replacement allocation [INFERENCE].
• Scalability Advantage: Zero per-screening API cost because AI inference runs 100% locally on existing government-issued ASHA smartphones [INFERENCE].

22. Differentiation

[Competitive Advantage Matrix]
┌───────────────────────────────────────┬───────────────────┬───────────────────┐
│ Feature / Capability                  │ Generic SIH Entry │ ArthriSense-NER   │
├───────────────────────────────────────┼───────────────────┼───────────────────┤
│ Hardware Category Compliance          │ ❌ Software Only  │ ✅ BLE IMU Strap  │
│ Works Under Loose Traditional Attire  │ ❌ Fails          │ ✅ Dual IMU Sensors│
│ 100% Offline AI Inference             │ ❌ Cloud Dependent│ ✅ Local ONNX ML  │
│ ABDM FHIR R4 Health Stack Integration │ ❌ Missing        │ ✅ Native FHIR R4 │
│ Multilingual NER Audio Interface      │ ❌ English Only   │ ✅ 8 NER Dialects │
│ Explainable ML (SHAP Risk Drivers)    │ ❌ Black-box      │ ✅ SHAP Explicit  │
└───────────────────────────────────────┴───────────────────┴───────────────────┘

1. Category Precision: Unlike generic software submissions that ignore the Hardware category tag, ArthriSense-NER provides a physical, wearable ESP32 sensor kit custom-built for low-cost field production [PS] [INFERENCE].
2. Attire & Environment Immunity: While vision-only systems fail when rural patients wear traditional attire (Mekhela Chador, Puan, Phanek), our dual-sensor fusion ensures accurate joint kinematic capture regardless of clothing or lighting [INFERENCE].
3. Deep Regional Adaptation: Tailored specifically for MDoNER with native audio guidance in 8 NER languages, lightweight offline sync for mountain terrain, and GIS administrative mapping for district health officers [PS] [INFERENCE].
4. Government Interoperability: Generates native ABDM FHIR R4 health bundles, allowing seamless integration into India's National Digital Health Ecosystem [3.4, 3.8].

23. Future Scope

1. Electromyography (EMG) Muscle Fatigue Module: Integrate surface EMG micro-sensors to detect quadriceps muscle activation deficits, providing earlier detection of muscle weakness preceding structural OA [INFERENCE].
2. In-Shoe Smart Pressure Insoles: Expand hardware options to include low-cost force-sensitive resistor (FSR) insoles for measuring plantopressure distribution and foot strike mechanics [INFERENCE].
3. Automated Physical Therapy Trainer: Utilize the smartphone camera and BLE IMU strap to guide patients through daily home physiotherapy routines, providing real-time audio corrections for posture and knee extension [2.7, INFERENCE].
4. Direct DICOM PACS AI Module: Build a secondary web module for PHC doctors equipped with digital X-ray machines to automatically segment joint space narrowing (JSN) using a U-Net convolutional neural network [1.4, 1.6].

24. Assumptions & Open Items Log

Consolidated Assumptions Log:

1. [ASSUMPTION] The average ASHA/ANM worker in NER possesses an Android smartphone running Android 9.0+ with Bluetooth 4.2/5.0 support and a functional camera.
2. [ASSUMPTION] The total hardware assembly cost of the ESP32 IMU strap can be maintained below ₹2,000 per kit in small-batch production.
3. [ASSUMPTION] A 5-meter straight walking space is available in rural health sub-centres or village community halls for gait testing.
4. [ASSUMPTION] Primary health workers can be trained to operate the guided 5-step screening app in less than 30 minutes.

Consolidated Open Items & Validation Log:

1. [NEEDS VALIDATION] Normative gait parameters and joint ROM baselines for indigenous hill populations carrying heavy loads over steep gradients in NER.
2. [NEEDS VALIDATION] Formal legal approval status from State Health Missions in NER for deploying non-invasive ASHA-led screening protocols.
3. [NEEDS VALIDATION] Exact BLE radio transmission attenuation through thick wet clothing during monsoon field camps in Meghalaya/Assam.

Final Recommendation

Proceed with Development of Solution Approach 2 (Hybrid Hardware-Software Ecosystem). This approach fully satisfies the MDoNER Problem Statement requirements, satisfies the SIH Hardware category mandate, delivers robust accuracy in difficult field environments, and operates completely offline [PS] [INFERENCE].

Recommended Stack

• Hardware: ESP32-C3 BLE + Dual MPU6050 IMUs + 800 mAh LiPo Battery [INFERENCE].
• Mobile Frontend: React Native + TypeScript + MediaPipe Pose Engine [INFERENCE].
• Edge AI: LightGBM Classifier compiled to ONNX Runtime Mobile [INFERENCE].
• Mobile DB: SQLite + SQLCipher + WatermelonDB [INFERENCE].
• Backend & Cloud DB: Python FastAPI + PostgreSQL 16 + PostGIS [INFERENCE].
• Interoperability: ABDM FHIR R4 JSON Specifications [3.4, 3.8].

Key Strategic Summaries

Research Sources