# Anthropometry and diagnostic aware deep learning for exercise assessment

## TL;DR

This paper introduces ADA, a deep learning framework that uses sensor data and personal features to assess exercise technique and predict injury risk.

## Contribution

ADA combines IMU kinematics with anthropometric and diagnostic data using attention-based fusion for improved movement assessment.

## Key findings

- ADA achieved 94.8% sequence-level accuracy in movement quality classification.
- Binary risk prediction reached 97.8% accuracy with diagnostic context.
- Personalized fine-tuning improved accuracy by 3%–5% depending on window length.

## Abstract

Correct technique during strength exercises such as squats and Romanian deadlifts (RDLs) is fundamental for performance and injury prevention.

We introduce ADA (Anthropometry and Diagnostic Aware), a multimodal deep-learning framework that integrates IMU kinematics with anthropometric and diagnostic features to classify movement quality and predict movement related risk.

Seventeen-sensor IMU data were collected from 15 healthy subjects performing correct and incorrect squat and RDL trials. A CNN-LSTM branch processed kinematic sequences and a fully connected branch processed static anthropometric/diagnostic inputs; feature fusion used attention weighting.

Incorporating anthropometry and diagnostic context increased sequence-level accuracy from 86.5% (kinematics only) to 94.8% (ADA) and enabled binary risk prediction at 97.8%. Personalized (transfer learning) fine tuning further improved accuracies (mean gains 3%–5% depending on window length).

ADA demonstrates that subject-specific static features improve movement quality classification and risk stratification, supporting wearable-based personalized feedback in training and rehabilitation.