Can Artificial Intelligence Analyze Force Plate Data for a Squat?



Advisor / Mentor: Chelette, Amber (Amber.Chelette@sfasu.edu)


Artificial intelligence is constantly evolving as it is prompted with new information. The use of artificial intelligence (AI) is becoming ever more prominent as this technology can analyze data at a rate non comparable to that of humans. Advances in AI have allowed for machine learning algorithms to be capable of analyzing the quality and trajectory of human movement via 3D kinematic data. The implementation of AI tools in clinical settings can help physicians eliminate bias and make diagnoses sooner. However, there should be limits placed on the data acquired to protect the patient. PURPOSE: The purpose of this study was to see if generative AI can analyze force plate data collected from a series of squat movements. We hypothesized that AI would provide partially correct information. However, the responses would be inconsistent due to AI constantly absorbing new information. METHODS: To determine the capabilities of AI, squat displacement data collected from a force plate was utilized. Three distinct AI models were used: Chat GPT, Google Gemini, Microsoft Copilot. Each AI underwent a series of questions designed to evaluate its aptitude in data analysis. The questioning process followed a hierarchical structure, starting with general requests and progressively refining to more specific inquests. RESULTS: When prompted, "Can you describe the movement taking place with the data presented" followed by the squat displacement data, Chat GPT, Google Gemini, and Microsoft Copilot were unable to generate a response. When prompted, "This measures force displacement between the right and left leg during a series of movements, can you tell me the movements taking place" followed by the squat displacement data, Google Gemini generated the following response: "The data you provided likely measures the displacement of force ... during a series of squatting movements" followed by how the AI bot came to that conclusion. When prompted "… tell me how many squats the subject performed" Google Gemini generated the following response:  "… it is impossible to determine the exact number of squats performed based on the provided data … to determine the number of squats we would need additional information like squat depth, time stamps, and thresholds." The AI's tested were unable to generate a response when prompted to identify whether the subject distributes force evenly and to time the pace of the squats. The response of Copilot only described the first and last numbers of the data set. It also described the pattern of the numbers as a downward trend. The only following responses of the AI were incorrect. CONCLUSION: When provided with more specific context, Chat GPT, Google Gemini, and Microsoft Copilot were able to identify patterns within the force displacement data to determine the movement taking place was a squat. The way in which prompts are worded matters, such that changing a single word can generate a completely different response. Our results indicate that the AI's tested can identify movements when provided displacement data. However, limitations prevail when inquiring with more specific prompts. This could be studied further provided access to more advanced AI models.

This document is currently not available here.