SAFE: A Sensor-Fusion Based Assessment and Prediction of Falling-Risk for the Elderly

Overview

Every nineteen minutes an older adult dies from a fall, making falls the leading cause of injury death to that group. Cumulatively older adult falls cost upwards of $50 billion in medical costs annually and is anticipating seven fall deaths every hour by the year 2030. Programs (like STEADI) use walking and other mobility tests, visual inspection and surveys of past history of falls to assess for risks. However, the tests are ineffective in picking up changes in baseline risk for the aged population until the next 6 month routine visit. This work proposes a novel low-cost multi-sensor based solution ‘SAFE’ , that uses IMU inertial sensors and camera input to predict the risk of falls quickly and accurately. IMU inertial sensor based devices are prototyped to collect 6-axis accelerometer data and capture video input to measure gait statistics with elderly subjects. A deep learning bi-directional LSTM network is fine-tuned and trained resulting in up to 94.26% accuracy. Using camera input and pose estimation models, gait speed is used to support the LSTM prediction. Results of the prediction with volunteer subjects shows up to 89.5% accuracy when predicting the likelihood of falling within six months for some participants. The derived gait speed metrics suggested an excellent correlation to the predicted results. SAFE accelerates sharing of fall prediction results allowing healthcare providers to assess the prediction in the risk of falling and take timely measures to prevent injury. SOS.net: A Robust System Harnessing the Power of AI to Expedite Search and Rescue Missions Nesara Shree Jesuit High School, Portland, Oregon According to the National Missing and Unidentified Persons System, over 600,000 people go missing around the US wilderness every year, and there are at least 1,600 people currently missing- these only being the ones that were officially reported. Current, drone-and-human-vision dependent systems in place are not only 82 incredibly inefficient, but also tiresome for drone pilot operators, who carry out over 60 Search and Rescue (SAR) missions a year. Alternatively, thermal detection drones used are inaccurate and far too generalizing, picking up on unrelated, inanimate objects that radiate heat. This is where I saw Artificial Intelligence (AI), Machine Learning (ML), and their powerful Computer Vision (CV) capabilities coming into play. What is needed is a reliable system that can accurately locate and signal by recognizing visual indicators of human presence or distress, and AI’s application is a crucial first step in being able to expedite SAR missions, relieving the strain on our SAR teams, and saving lives. The goal is to make human search missions much more refined and efficient by implementing RCNN’s Resnet50 ML model methodologies. By open-sourcing SOS.net, meaning that all of the code, procedures, a snapshot of the trained model, and the option to access the entire dataset is available to the general public on Github to download and/or contribute towards its further improvement, enables SOS.net to be a dynamic, yet robust, AI system that has high potential for actual implementation and continued refinement as a tool.

Competition history

• JSHS 2023 Category not listed

Resources

• National JSHS abstract book ↗

Related projects

Closest projects by meaning, across every fair and year in the corpus.