# Measuring Assistive Technology Outcomes via AI-Based Kinematic Modeling of Individualized Routine Learning in Elite Boccia Athletes with Severe Cerebral Palsy: A Longitudinal Case Series

## TL;DR

An AI system tracks movement patterns of elite Boccia athletes with severe CP, offering new insights into motor learning and performance.

## Contribution

Introduces an AI-based kinematic model for measuring assistive technology outcomes in athletes with severe disabilities.

## Key findings

- The AI system achieved high alignment with expert ratings (MSE 1.14, MAE 1.13).
- For the most impaired athlete, a 4.69% reduction in kinematic variability was detected despite stable performance scores.
- Training loss decreased from 7.45 to 1.19, showing system stability and convergence.

## Abstract

Objectives: This longitudinal single-case series evaluated an AI-based routine-learning system as assistive technology (AT) for elite Boccia athletes with severe Cerebral Palsy (CP). The study aimed to provide an innovative outcome measurement approach for individualized monitoring by integrating performance scores and longitudinal kinematic variability indicators. Methods: Three national-level players performed 694 throws over eight weeks. To ensure technical credibility, trials were rated through a consensus-based assessment by a panel of two experts, serving as ground truth for AI modeling. The system utilized a Bidirectional Long Short-Term Memory (Bi-LSTM) architecture to extract 29 kinematic features and perform regression-based scoring, providing real-time augmented feedback. Results: High-baseline tasks maintained stable scores (7–9), while intermediate tasks showed significant score increases, reflecting motor learning transitions. The model achieved a Mean Squared Error of 1.14 and a Mean Absolute Error of 1.13, demonstrating high alignment with expert standards. Training demonstrated stable convergence, with loss reducing from 7.45 to 1.19. Notably, for the most severely impaired athlete, the AI system detected a 4.69% reduction in kinematic variability despite stagnant performance scores. This provides empirical evidence of movement stabilization within the cognitive stage that traditional observation might overlook. Conclusions: The Bi-LSTM system enabled accurate tracking of performance and motor variability, revealing distinct learning curves based on task difficulty. These findings demonstrate the feasibility of AI-enabled motion analysis as an AT for outcome measurement, supporting data-driven coaching where conventional evaluation is constrained by the rarity and severity of disabilities.