# Smart Total Knee Replacement: Recognition of Activities of Daily Living Using Embedded IMU Sensors and a Novel AI Model in a Cadaveric Proof-of-Concept Study

## TL;DR

A smart knee replacement with embedded sensors can accurately detect daily activities, aiding personalized rehabilitation.

## Contribution

A novel AI model using embedded IMU sensors in knee prostheses for ADL recognition is introduced.

## Key findings

- The AI model achieved 95.68% overall accuracy in recognizing activities of daily living.
- Sitting, standing, stance, and knee bending were recognized with 100% accuracy.
- Walking and stair navigation had high F1 scores of 0.98 and 0.92, respectively.

## Abstract

What are the main findings?

Data from dual IMU devices embedded in a knee replacement prosthesis can accurately recognize activities of daily living.

What is the implication of the main finding?

Enhances the possibility of remote minimally intrusive personalised rehabilitation.

Total knee replacement (TKR) is a reliable treatment for end-stage degenerative conditions of the knee. Patient-reported outcome measures (PROMs) are central to assessing TKR outcomes, but they have limitations. Activities of daily living (ADLs) in the early post-operative period complement PROMs for holistic patient assessment. This study presents a method for capturing ADL parameters from data generated by inertial measurement unit (IMU) devices embedded in TKR prosthesis. A conventional posterior stabilized TKR was modified to create chambers in the femoral and tibial components. The prosthesis was implanted into a cadaver knee and movement was simulated using a hydraulic actuated knee simulator (AMTI, VIVO, MA, USA). A powered IMU device was placed in each of the chambers. The simulator was activated for various ADLs and the generated data was collected wirelessly. The pre-processed data was fed into a novel multimodal deep learning artificial intelligence model created to recognize specific ADL (trained on 70% of the data, with 30% reserved for validation and testing). The model achieved 95.68% overall accuracy, with 100% for sitting, standing, stance, and knee bending. Walking, stair navigation, and jogging showed F1 scores of 0.98, 0.92, 0.91, and 0.89, respectively. This technology enables seamless knee activity recognition and reporting with positive implications for patient-specific rehabilitation protocols.