(P61) Artificial Intelligence to Prevent Falls

Objectives: We developed a proof-of-concept artificial intelligence (AI) prototype algorithm which identifies patients with unsteady gaits from emergency department (ED) CCTV footage. The algorithm then alerts healthcare workers to these patients so that falls may be prevented. In this initial phase, we developed the algorithm and performed preliminary tests.

Background: Falls within hospitals are a major safety issue . 30-50% of falls result in injuries, contributing to an average increased hospitalization length of stay of 6.3 days and $14,000 additional medical costs per admission . In the elderly, falls often result in decreased mobility and loss of independence , thus burdening healthcare systems .
Crowds, blind corners, and long corridors in the ED make it difficult to have direct line of sight to all patients. Using existing CCTV to keep an additional ‘eye’ on patients who may be unsteady and at risk of falling can improve safety.

Methods: We collaborated with engineers from Nanyang Polytechnic (NYP) in Singapore to develop this algorithm.

Step 1 – Data collection and de-identification

Skeletal data from 6 weeks of CCTV footage from Singapore General Hospital ED was extracted using OpenPose . OpenPose is a real-time AI system which detects key skeletal points on the body. The skeletal point data subsequently used for analysis.

Step 2 – Dynamic Time Warping and Backpropagation Classifier

Dynamic Time Warping (DTW) , which is a method of calculating the degree of synchronization between two temporal sequences, was used to compare a patient’s walking pattern to a standard steady walking pattern. In our study, we found that left and right ankle points were most useful for gait analysis and hence concentrated on these points in this initial study phase.

The DTW scores of left ankle data points for unsteady gaits were much higher than the steady gaits [figure1]. These scores were then fed into a 3-layer backpropagation neural network model to determine the patient’s steadiness. The network was trained using 80 labelled data points.

Figure 2 [figure2] shows a summary of the algorithm development process.

Results: Our prototype was able to achieve an accuracy of 93.2% on the training datasets. The algorithm was also tested using 38 datasets (19 steady gaits, 19 unsteady gaits) captured in a controlled environment within NYP. Out of these 38 datasets, 35 (92.1%) of them were correctly classified.

Conclusions: Our proof-of-concept AI algorithm shows promising gait analysis potential. While it performs well in an controlled environment, for it to be realistically implemented in the ED, we will need to:
1. Improve accuracy by further training the algorithm with more numerous, diverse patient videos
2. Improve robustness by including the tracking the coordinates of other skeletal key points
3. Integrate with the existing CCTV system

Our team has applied for further funding to expand this project and we aim to further refine our prototype for real-life deployment and to make waiting rooms a safer place for all patients.

References (Optional): 1. Serious Reportable Events in Healthcare–2006 Update [Internet]. National Quality Forum. National Quality Forum; 2007 [cited 2022 Dec 14]. Available from: https://www.qualityforum.org/Publications/2007/03/Serious_Reportable_Events_in_Healthcare%E2%80%932006_Update.aspx

2. Joint Commission. Preventing falls and fall-related injuries in health care facilities. Sentinel Event Alert. 2015Sep28;(55):1–5.

3. Ambrose AF, Paul G, Hausdorff JM. Risk factors for falls among older adults: A review of the literature. Maturitas. 2013;75(1):51–61.

4. Choi NG, Choi BY, DiNitto DM, Marti CN, Kunik ME. Fall-related emergency department visits and hospitalizations among community-dwelling older adults: Examination of health problems and injury characteristics. BMC Geriatrics. 2019;19(1).

5. Qiao S, Wang Y, Li J. Real-time human gesture grading based on openpose. 2017 10th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI). 2017;

6. Berndt DJ, Clifford J. Using dynamic time warping to find patterns in time series. AAAI Technical Report WS. 1994Apr26;94(03):359–70.

Funding (Optional): A donation from Tesa Tape Asia Pacific Pte Ltd helped to fund this research.