Article
Exploring the hidden information in gait using wearables and deep learning
Search Medline for
Authors
-
Ricardo Smits Serena
- Department of Orthopaedics, Institut of AI and Informatics in Medicine, Technical University of Munich, München, Germany
-
Simone Beischl
- Klinik und Poliklinik für Orthopädie und Sportorthopädie, Klinikum rechts der Isar, TU München, München, Germany
-
Igor Lazic
- Klinik für Orthopädie und Sportorthopädie, Klinikum rechts der Isar, Technische Universität München, München, Germany
-
Barbara Vogel
- Zentrale Physiotherapie, Klinikum rechts der Isar, Technische Universität München, München, Germany
-
André Bergmann
- Zentrale Physiotherapie, Klinikum rechts der Isar, Technische Universität München, München, Germany
-
Florian Pohlig
- Klinik für Orthopädie, Klinikum rechts der Isar, Technische Universität München, München, Germany
-
Fritz Seidl
- Klinik für Orthopädie, Klinikum rechts der Isar, Technische Universität München, München, Germany
-
Rainer Burgkart
- Klinikum rechts der Isar, Technische Universität München, München, Germany
-
Rüdiger von Eisenhart-Rothe
- Klinik für Orthopädie, Klinikum rechts der Isar, Technische Universität München, München, Germany
-
Florian Hinterwimmer
- Institute for AI and Informatics in Medicine, Technical University of Munich, München, Germany
| Published: | October 23, 2023 |
|---|
Outline
Text
Objectives: Precision medicine is a new approach to disease treatment and prevention that considers the individual variability of each person. This concept will make it possible to predict more accurately which treatment and prevention strategies for a particular disease will be successful in which groups of people. To establish precision medicine approaches in orthopaedics (e.g., musculoskeletal diseases), objective data as well as powerful data analysis tools are required.
In this study we describe an approach to utilize Deep Learning methodology to analyse gait data recorded with common wearables to predict patient specific traits.
Methods: The presented prospective, experimental study was conducted at our Biomechanical Lab using 4 IMUs (2 MetaMotionS, 1 iPhone, and 1 Apple Watch) placed on the right distal tibia, right lower thigh, right upper thigh, and left wrist respectively (Figure 1 [Fig. 1]). The measurements were done on 15 healthy participants (8 women and 7 men, 32 ±5.5years, 1.75 ±0.09m). For each participant we recorded the accelerometer and gyroscope data for two 1-minute forward walks recorded on 2 different days. A state-of-the-art time series classification Deep Learning model was used to try to predict the Age, Height, and Gender from each person.
Results and conclusion: For each prediction a separate model was trained and 4-fold cross-validation was performed. We achieved 81.2%, 71.8% and 84.3% accuracies for predicting Age, Height, and Gender, respectively.
The main finding of this study is that with very low amounts of data (only movement data from one leg and one arm) accurate predictions of person-specific characteristics were feasible. We hypothesize that by replicating the study with the correct patient subset, we would be able to predict of gait-related issues such as osteoarthritis of the knee or hip joint.
