Challenge

Data management and scalability limitations slowed AI-driven care innovation

SafelyYou’s AI solution provides patient safety and continuous care management by leveraging computer vision models that process real-time data from millions of hours of video footage. Before implementing FiftyOne as part of their ML data management and model development pipeline, SafelyYou’s AI team faced challenges in scaling their AI-driven solutions for fall detection and healthcare monitoring.

• Custom data analysis: The AI team spent at least half their time maintaining custom-built scripts for reviewing AI predictions, false positives, and negatives for their Respond solution, considerably slowing their model development workflow.
• Data management dependencies: The team had limited control over their data and relied on their internal Case & Notification team (which was responsible for managing the clinical transactional database) to implement queries and visualizations for AI model evaluation. The workflow wasn’t ideal and hindered development and quick iterations.
• Scalability: Their transactional data pipeline was difficult to scale, which created bottlenecks in processing the growing volume of video data required for training AI models.

As the team moved to an event-based architecture to reduce bottlenecks, the AI team needed a data solution that was scalable and flexible, and which allowed them to:

• Explore and filter large volumes of video data quickly and share interesting samples without downloading them
• Instantly visualize object annotations and identify mistakes
• Protect the personally identifiable information (PII) associated with healthcare

Solution

Scalable, privacy-compliant model development workflow for accurate fall detection and care management using FiftyOne

SafelyYou found FiftyOne as the perfect fit to address their data management and scalability challenges for their Respond and Clarity solutions. With FiftyOne as the primary resource for ML work, the team was able to:

• Streamline the ML workflow by visualizing and exploring videos and metadata such as start/end timestamps, sensor IDs, and AI scores in production. They were able to cross-reference videos with their data warehouse to construct datasets for evaluation and explore their data visually and dynamically, without downloading the data to a remote server or incurring dependencies on other teams.
• Easily find annotation mistakes with FiftyOne field visualizations and tagging to distinguish incorrect labels and predictions
• Fine-tune and validate AI models by examining model performance to track false positives/negatives and create quality datasets for training and evaluation.

With FiftyOne, SafelyYou was able to meet their privacy compliance needs by ensuring personally identifiable information (PII) in healthcare data was protected during the annotation and model training process.

Key Results

Streamlined ML workflow reclaimed 60-80 hours of development time per month; 77% reduction in images that require manual verification

FiftyOne has helped SafelyYou process over 60 million minutes of video footage, totalling 3TB of data. Since implementing the tool in their ML pipeline, the team saw the impact on their workflow almost immediately.

These results have delivered significant business benefits to SafelyYou. Their technology has reduced falls by 40% and improved patient safety. It has also helped extend residents’ stays by an average of 4+ months, enabling care facilities to retain long-term residents and provide consistent care.