Background

In addition to clinical 3D gait analysis (3DGA), advanced musculoskeletal (MSK) models offer deeper insights in movement dynamics. However, their requirement for specialized expertise, high computational resources, and time consumption limits their clinical utility. Machine learning (ML) methods can help overcome this limitation by significantly accelerating MSK modelling workflows. However, ML come with their own challenges. For instance, many healthcare professionals lack a sufficient understanding of ML, rendering it difficult for them to interpret the results and, as a consequence, leading to reduced trust in the technology. In the project DeepForce XAI, I address this issue and apply methods of explainable AI (XAI) to make ML model results easier to understand.

Project Content

DeepForce XAI builds on the DeepForce project. It employs explainable AI (XAI) to enhance the transparency and understandability of the ML models used in DeepForce. Most XAI methods were designed for image and video data. The main challenge is to find an XAI method that prepares gait data in a way that makes the most relevant information stand out and easy to interpret.

Goals

The primary goals pursued in DeepForce XAI are:

• Identifying the most suitable XAI methods for gait analysis.
• Tailoring these methods to process gait analysis data.
• Evaluating and refining the methods based on user feedback.

Methods

DeepForce XAI uses a gait analysis dataset of 3,000 patients collected during the DeepForce project. We conducted a systematic review to summarise the state of the art in machine-learning methods for predicting lower-extremity joint loads. In parallel, we are running a qualitative study to identify the XAI techniques best suited to musculoskeletal (MSK) simulations and clinical decision-making. Self-explaining and post-hoc methods like BLA, DiCE, and GradCAM are among the top candidates due to their high comprehensibility.

In the next step, these methods are modified to meet the requirements of gait analysis and the needs of clinicians. For quantitative evaluation, techniques such as White Box checks, preservation checks, and data randomization checks are used. For qualitative evaluation, we ask clinical experts to provide feedback on the clarity, relevance, and usefulness of the XAI models. In addition, a usability study with healthcare professionals is carried out to assess whether our methods perform as intended and meet expectations of their future users.

Funding

The content does not necessarily represent the view of the state of Lower Austria or the funding agency. Neither the state of Lower Austria nor the funding agency can therefore be held responsible for the content.