Blue Modern Elegant Presentation

5/21/2026. 1. A Universal Plug-and-Play Edge-AI Platform for Privacy-Preserving Healthcare Applications.

Introduction. AI in Healthcare – Growing Need AI in Healthcare — The Growing Need • Healthcare AI market: $187.9 billion by 2030 (CAGR ~37%) • Applications: medical imaging, emotion recognition, TB screening, skin disease diagnosis • Critical gap: AI benefits rarely reach rural and low-resource settings Why Cloud AI Fails in Healthcare • 🔒 Privacy risk — raw patient data transmitted to third-party servers • ⏱ Latency — network delays of 150–500ms affect real-time diagnosis • 📡 Connectivity — 3.5 billion people lack reliable internet access.

5/21/2026. 3. PROBLEM STATEMENT. Fragmented Systems: Hospitals currently need separate AI systems for different tasks (ECG, Imaging, etc.), leading to high costs. Cloud Dependency: Most medical AI relies on the internet, causing high latency and operational expenses. Privacy Risks: Centralized data processing in the cloud poses significant risks to sensitive patient information. Inflexibility: Existing solutions are task-specific and difficult to update or adapt to new clinical models..

THE EDGE-AI SOLUTION. UNIVERSAL AI GATEWAY. We engineered a model-agnostic platform that decentralized diagnostics by bringing the brain to the clinic..

5/21/2026. 5. PROJECT GOALS. Universal Platform: A single, model-agnostic Edge-AI platform that runs multiple healthcare models on a single device. Plug-and-Play: A registry-based architecture that allows for dynamic loading of new models without altering core code. Local Processing: All inference is done on-device (Raspberry Pi), ensuring 100% offline capability and data privacy. User-Centric: Includes a lightweight web-based dashboard for clinicians with no technical expertise..

5/21/2026. 6. RESEARCH APPROACH. Literature Review: Analyzed existing Edge-AI frameworks and identified the lack of "Universal" platforms in the biomedical sector. Hardware Selection: Benchmarked Raspberry Pi 4 (8GB) to ensure it could handle real-time inference without thermal throttling. Multi-Modal Validation: We didn't just test one model; we researched and validated the platform across: Vision: Skin Disease & Facial Emotion. Audio: Tuberculosis Predicting using Cough Data. Data: Chronic Disease Risk Forecasting. Optimization Research: Studied weight pruning and efficient threading to ensure the web dashboard remains responsive during heavy AI computation..

5/21/2026. 7. METHODOLOGY. Modular Design Approach: The platform is built using a layered system architecture: Input Layer: Handles diverse data sources (USB cameras, microphones, CSV uploads). Preprocessing Layer: Automatically resizes and normalizes data based on the selected model's requirements. Model Management Layer: A JSON-based registry that maps model files to their clinical category. Inference Engine: Uses the TensorFlow Lite Interpreter for optimized execution on the ARM processor. Quantization: We utilized 8-bit integer quantization to reduce model size by 4x while maintaining ~98% of the original accuracy..

Hardware Details. 5/21/2026. 8. Component Specification Device Raspberry Pi 4 Model B RAM 8 GB LPDDR4 Processor Quad-core Cortex-A72 (ARM v8) 64-bit SoC @ 1.5 GHz Storage 128 GB MicroSD Card Operating System Raspberry Pi OS (64-bit Linux-based) Connectivity Wi-Fi, Ethernet, Bluetooth Power Consumption Approx. 5–7 W Display Interface HDMI Output Peripheral Support USB Camera, Microphone, Sensors.

Software Details. 5/21/2026. 9. Model Name Task Model Type Input Type Deployment Format Facial Emotion Recognition Emotion Detection CNN Facial Image TensorFlow Lite Skin Disease Classification Dermatological Diagnosis CNN Skin Image TensorFlow Lite Thermal Emotion Recognition Physiological Emotion Detection CNN Thermal Image TensorFlow Lite TB Chest X-ray Detection Tuberculosis Screening Deep CNN Chest X-ray Image TensorFlow Lite Multimodal TB Prediction TB Risk Analysis Gradient Boosting Audio + Clinical Data Optimized ML Model.

Model Deployment Pipeline. 5/21/2026. 10. [image] Preprocessing Module Processed Data (Consolidated) • Image Processing Model Query Model Info Response Loaded A1 Model Processed Data Model Registry (Multiple A1 Models Stored) Raw Prediction Output User Interface (Input Layer) Input Data Processed Data & Application Model Logic Layer fTFLite Inference Result Data Validated Data Data Validation Module User Dashboard / Result Display • Audio Processing . Clinical Data Processing Engine Prediction Module (TensorFlow Lite Edge Inference).

5/21/2026. 11. [image] uuoneld . Ivabp3. Component Specification Edge Device Raspberry Pi 4 (8GB RAM) Peripherals USB Camera, Microphone, 3.5-inch Integrated Display Runtime Python 3.10 with TensorFlow Lite Interpreter Libraries OpenCV (Imaging), Flask (Web Dashboard), NumPy Optimization Quantized TFLite models for memory efficiency.

Performance Evaluation. 5/21/2026. 12. Inference latency CPU usage System stability Multi-model execution Real-time prediction accuracy.

5/21/2026. 13. Developed a Functional Universal Edge AI Platform A fully operational edge-based healthcare AI system was successfully implemented on Raspberry Pi 4, capable of running multiple AI models locally without cloud dependency..

5/21/2026. 14. Developed a Functional Universal Edge AI Platform A fully operational edge-based healthcare AI system was successfully implemented on Raspberry Pi 4, capable of running multiple AI models locally without cloud dependency..

5/21/2026. 15. VITAL EDGE A1 - LIVE MODEL DIAGNOSTIC (STRESS TEST) INFO: Created TensorFlow Lite WNPACK delegate for CPU. @ Thermal emotion model.tflite I Input: @ fer model.tflite @ skin disease.tflite @ multimodal tb.tflite I Input: [ I Input: [ I Input: [ [ 1 128 128 1 48 II I 1 64 €4 31 1 1 soo soo 31 1 Latency: 53.10ms Latency: 26.16ms Latency: I. 36ms I Latency: 307.16ms.

5/21/2026. 16. RESULT. User Interface & Experience Features Model selection dropdown Dynamic input interface File upload & camera activation Real-time prediction display Confidence score visualization Dual Deployment: Accessible via a local web browser or a native desktop application. Non-Technical Interaction: Clinicians simply select a model, upload data (or use live camera), and view confidence scores. Robustness: Integrated error handling for corrupted files or improper dimensions, preventing system crashes..

5/21/2026. 17. CONCLUSION. Project Impact Successful Integration: Developed and validated a Universal Edge-AI Gateway that breaks the "one-device-one-task" barrier in medical diagnostics. Privacy & Autonomy: Proved that complex healthcare AI can run offline, ensuring total patient data sovereignty and HIPAA-compliant workflows. Optimized Performance: Achieved real-time inference (under 500ms) on low-cost Raspberry Pi hardware, making high-end AI accessible to resource-limited settings. Plug-and-Play Flexibility: Demonstrated a modular registry system that allows for rapid scaling—new medical models can be deployed without software re-engineering..

5/21/2026. 18. AI Usage Disclosure. AI Tools Used ChatGPT Google Gemini How AI Was Used AI tools were used as supporting assistants during development, including: Concept exploration and brainstorming of Edge-AI healthcare architecture Coding assistance and debugging during development of the inference pipeline and system modules Guidance for TensorFlow Lite model conversion and optimization workflows Assistance in troubleshooting environment and dependency issues Documentation and report structuring for technical clarity Presentation preparation and slide organization Mock evaluation and possible judge-question preparation Validation and Team Ownership All AI-generated suggestions were reviewed, verified, and modified by the team. Model training, integration, system design, and testing were implemented by the team members. All technical components of the system are fully understood and explainable by the team during judging. Accountability Statement The final architecture, code implementation, system testing, and results are entirely the responsibility of the project team..

5/21/2026. 19. Research Publication (Conference Submission).

5/21/2026. 20. THANK YOU.