AI Driven Seizure Detection for Smarter Healthcare Monitoring

Overview

A leading provider of EEG solutions in the United States partnered with D3V Technology Solutions for a cutting-edge healthcare initiative. The core objective of this project was to develop a sophisticated, domain specific machine learning (ML) model capable of accurately detecting seizures from EEG video recordings. This high impact AI initiative focused on computer vision and was executed by the expert team at D3V Technology Solutions.

The Challenge

The primary challenge was to engineer a machine learning model with exceptionally high sensitivity to ensure no seizure events were missed, while maintaining high specificity to minimize false alarms. The healthcare provider required a system that could deliver timely and reliable seizure notifications to remote monitoring technologists, helping them enhance patient monitoring capabilities and response times.

Our Solution

D3V Technology Solutions designed and implemented a cloud native proof of concept (POC) system on Google Cloud Platform (GCP). The architecture followed an event driven microservices approach to ensure scalability and efficiency.

Key components of the solution included:

• A custom tuned EfficientNetV2-M machine learning model hosted on Vertex AI Endpoint for real time inference
• A FastAPI based backend deployed on Cloud Run to provide API services
• A video processing service running on Cloud Run to manage video workflows and trigger inference processes
• Google Cloud Storage for secure and scalable data storage

The project was executed using an agile methodology over a 12 to 16 week period, covering:

• Data ingestion and preprocessing
• Model development and training
• Continuous optimization and evaluation

The model development progressed through three stages:

V1: Initial model selection and baseline development
V2: Performance optimization and preprocessing improvements
V3: Large scale validation and refinement

Impact

The final model version significantly exceeded the original performance goals by achieving more than 95% sensitivity and specificity in seizure detection. This represented a substantial improvement over the initial targets of 85% sensitivity and 75% specificity.

The AI powered solution created a scalable foundation for future enhancements, including model retraining with larger datasets and continuous performance improvements. The highly accurate and automated detection system empowers healthcare monitoring teams with faster, more relevant, and reliable insights.

Key Accomplishments

• Successfully designed and delivered a high performance custom machine learning model for seizure detection that exceeded expected accuracy targets
• Built a scalable and resilient cloud native solution on Google Cloud using modern serverless technologies
• Achieved results strong enough to qualify the project for partner award submissions in the Artificial Intelligence category
• Delivered detailed technical documentation, API documentation, and handover materials to support seamless integration and long term maintenance