Student Jesuit High School, Portland, OR Portland, Oregon, United States
Introduction:: With the ever-growing global aging population, falling is amongst the most serious public health problem faced by people in this risk group. Every nineteen minutes an older adult dies from a fall, making falls the leading cause of injury death to that group1. The most applicable and well known fall prevention initiative is STEADI (Stopping Elderly Accidents, Deaths & Injuries). STEADI uses a combination of walking, getting up from a chair as well as a series of survey questions. However, STEADI demonstrates high false negative rates among those categorized as minimal risk2. The tests are ineffective in picking up regular and continuous changes in baseline risk for the aged population. In most cases, by the time the next STEADI test is taken, the patient may have already fallen.
Gait metrics derived from IMU (inertial measurement unit) based sensors like accelerometers are highly effective in predicting fall risk thereby overcoming the shortcomings of the STEADI test3, however, this is done in an elaborate laboratory setting with many IMU based sensors attached to the body can be impractical for quick measurements especially for the elderly.
This work presents a novel sensor based solution, which uses one IMU wearable and video camera input (along with pose estimation) to assess and predict the risk of falls quickly and accurately. This dual input technique eliminates the need for multiple IMU sensors to derive gait metrics, overcomes the shortcomings of the current methods thereby providing an effective way to monitor and predict fall risk regularly.
Materials and Methods:: The design process consisted of the firstly designing and building a wearable prototype that is used for collecting IMU accelerometer and gyroscope data. After testing many wearable prototypes, a lightweight compact accelerometer and gyroscope sensor ‘Witmotion’ was selected to take the assessments. The assessment comprises of a participant walking for ten yards with the wearable worn at the waist. The IMU directional metrics (mediolateral, anterior-posterior and vertical acceleration) are recorded. The built-in Kalman filter provides additional accuracy with the measurements. . A camera mounted on a tripod is used to record the video of the participant while walking. Pose estimation libraries4 post process the video and convert the video frames into full body skeletal poses that can be used to estimate gait metrics like stride speed and stride length. Many deep learning models were chosen for training with the goal of achieving the highest accuracy (and least loss). Training was with more than 350,000 datapoints with up to 151,746 total parameters. Iterative epoch scaling as well as batch size scaling is performed to tune the model for highest accuracy. Participant data is used to generate classification on the trained model as a faller or non-faller. Accuracy of the prediction was also monitored over time for participants to gauge the increase or decline in the predicted classification such that preventative measures could be triggered to prevent future falls.
Results, Conclusions, and Discussions:: Results and discussions
We trained the data with a variety of machine learning and deep learning models. Based on our results, the LSTM (Long Short-Term Memory) network proved to be the best algorithm to predict falls. Within the LSTM family of networks, in addition to the vanilla LSTM network, we also evaluated the bi-directional LSTM and stacked LSTM networks. The stacked and bi-directional LSTM models showed the highest accuracy based on the accuracy and loss scores from the training results. The accuracy of prediction measured up to 94.26, with a loss of up to 16.89. The result of training showed high precision and recall for the training dataset, with an F1-score of 0.85 reported on the validation set. To the best of our knowledge, these are the best reported values for F1-score so far.
Our results show IMU sensor data and gait metrics are highly effective parameters in predicting fall risk. In addition, this solution can effectively be used to quickly assess, predict, and regularly track elderly participants so that any variation in baseline risk can be monitored and shared with the health care providers. Based on the assessment of the health experts, preventative measures can be taken immediately to prevent falls.
Conclusion:
We developed a novel sensor based solution that uses IMU sensor data and pose estimation methods to assess and predict the probability of falls in the elderly.
The results show strong applicability of machine learning (especially LSTM networks) to assess IMU and camera based data. This is possible because machine learning models can process large volumes of sensor data and the complexity of many parameters on which the results are dependent. Additionally, the ability of the models to integrate the smallest of changes from the sensors can quickly help monitor and track changes to the baseline risk for future falls.
In the future, this study can be expanded to monitor post-operative rehabilitation and motor function challenged patients. In addition, enhanced prediction by training for activities like bending, sitting, and ascending and descending stairs would enhance the solution to add additional coverage for fall scenarios.
Acknowledgements (Optional): : I would like to thank my mentor Dr. Omesh Tickoo for guiding me on this project. I would also like to thank all my volunteer elderly subjects who helped me.
References (Optional): : 1. World Health Organization. Falls. In World Health Organization. Retrieved December 02, 2023, from https://www.who.int/news-room/fact-sheets/detail/falls 2. Robert W. Nithman, Jennifer L. Vincenzo August 2019, How steady is the STEADI? Inferential analysis of the CDC fall risk toolkit. Archives of Gerontology and Geriatrics. Volume 83 3. Richard A. W. Felius, Marieke Geerars, Sjoerd M. Bruijn, Jaap H. van Dieën, Natasja C. Wouda and Michiel Punt (2022), Reliability of IMU-Based Gait Assessment in Clinical Stroke Rehabilitation. MDPI Sensors, 22(3), 908. 4. Open pose. In CMU-Perceptual-Computing-Lab/openpose Retrieved January 1, 2023, from https://github.com/CMU-Perceptual-Computing-Lab/openpose 5. Long short-term memory. In Wikipedia. Retrieved January 1, 2023, from https://en.wikipedia.org/wiki/Long_short-term_memory
