BioMed Engineer North Allegheny High School Sewickley, Pennsylvania, United States
Introduction: Falls in elderly nursing homes are a massive issue currently, and I saw this firsthand with my grandma. Many times patients are falling, and then it can take hours for them to get medical attention. The first part of my project is detecting falls, to do this I built a gadget that is installed on the patients, detects when the patient falls, then alerts a caretaker or medical worker on an app(downloaded on mobile phone) when there is a fall, the gadget also has a buzzer which sounds when there is a fall. Then to understand major reasons for falls, I used an LLM(ChatGPT because the most data has been passed through it), where the 2 coders will write a free-text description about a video of falls in nursing homes. After this, the LLM categorizes the description into 4 major categories.
Materials and
Methods: Materials: Adafruit gadget, computer, usb cables, pouch to hold gadget, scissors, mobile phone to hear ringtone
Methods: For the detection part of the project, the subject(myself) fell onto a mattress at various angles and velocities 15 times in sets of 3. For the understanding reasons for falls, I ran each description into the LLM three times to analyze consistency and more reliable results.
Results, Conclusions, and Discussions: Overall the results for the detection part of the project was that on the app the ringtone sounds 87% of the time from the mobile phone. The buzzer on the gadget rang 100% of the time. Analyzing this further on the ROC curve, an optimal TPR(True Positive Rate) is 1 and FPR(False Positive Rate) is 0, for my procedure the TPR is 0.93 and FPR is 0.07. This shows high accuracy with the device. For understanding reasons for falls, in the categorization analysis I clearly saw that the category "Shifting Of Weight" was the most major reason for falls. Then with the probability analysis(A unique feature of ChatGPT to be able to associate a probability that applies to each description), I found "Shifting Of Weight" to be the major reason as well. In the LLM procedure, a big limitation I faced was that with each categorization of the description there was variability with the results which could have slightly diluted my conclusion. In the future, I would like to be able to run my app in the background, so the person does not have to be on the app to hear the ringtone when there is a fall, and eventually put the device in a nursing home with real patients. Finally, on the understanding reasons part, I would like to use an NLP to automatically create the 4 major reasons for falls, instead of a human which could introduce bias.
Acknowledgements (Optional): Mr.Harshal Mulerkhar, who helped guide me through issues that occurred when I was making my app Mr. Kurt Beschorner(Professor from Pitt who specialized in analyzing falls), helped me with analyzing reasons for falls and identifying patterns Mr. Robert Zacharias(Professor from Carnegie Mellon University who specialized in Arduino gadgets), introduced me to the adafruit gadget I used in my procedure.
