Optimized Edge-AI Streaming for Smart Healthcare and IoT Using Kafka, Large Language Model Summarization, and On-Device Analytics

This exponential growth poses serious challenges for traditional cloud-based architectures due to their high latency, bandwidth consumption, and privacy vulnerabilities. This study presents a novel Edge-AI streaming framework that integrates Apache Kafka for scalable real-time data streaming, DistilBERT for summarizing unstructured clinical notes on-device, and XGBoost for edge-based predictive analytics. The framework leverages Kafka’s partitioned topic structure for high-throughput streaming, deploying a quantized DistilBERT model via ONNX to generate compact clinical summaries, and executing real-time classification tasks using XGBoost on edge devices (Raspberry Pi 4 and Jetson Nano). Performance was evaluated on a simulated dataset of 10,000 patient records using latency, throughput, ROUGE/BLEU scores, and F1-score as key metrics. The system achieved a 40% latency reduction (185 ms vs. 680 ms cloud), 35% throughput improvement (850 vs. 630 records/s), and high summarization performance (ROUGE-1: 0.87, ROUGE-2: 0.82, BLEU: 0.76). Predictive analytics delivered an F1-score of 0.94 for anomaly detection. HIPAA compliance was ensured through local data processing, and resource usage remained minimal (≤1.5 GB RAM, ≤70% CPU). By tightly coupling Kafka streaming, LLM summarization, and edge analytics, the proposed framework provides a scalable and secure infrastructure for real-time healthcare IoT. Future work includes incorporating federated learning, multimodal data fusion, and dynamic compression to enhance adaptability and clinical utility.