Our abstract titled “Diagnosing Rare Diseases by Movement Primitive-Based Classification of Kinematic Gait Data” was accepted as poster presentation and will be presented by our collaboration partner Jing Xu from Marburg.
Xu, Jing; Greß, Hannah; Seefried, Sabine; van Drongelen, Stefan; Schween, Raphael; Sommer, Claudia; Endres, Dominik; Krüger, Björn; Stief, Felix
Diagnosing Rare Diseases by Movement Primitive-Based Classification of Kinematic Gait Data Proceedings
Bernstein Conference, 2023.
@proceedings{JingXu2023,
title = {Diagnosing Rare Diseases by Movement Primitive-Based Classification of Kinematic Gait Data},
author = {Jing Xu and Hannah Greß and Sabine Seefried and Stefan van Drongelen and Raphael Schween and Claudia Sommer and Dominik Endres and Björn Krüger and Felix Stief},
url = {https://abstracts.g-node.org/conference/BC23/abstracts#/uuid/31c21041-91a0-46bd-87dc-46271501fdc0},
doi = {10.12751/nncn.bc2023.313},
year = {2023},
date = {2023-01-10},
urldate = {2023-01-10},
booktitle = { Bernstein Conference 2023},
abstract = {Of over 6.000 known rare diseases, a considerable portion involves motor symptoms [1]. Whereas aiding diagnosis by artificial intelligence based on non-motor symptoms has shown promise [2], the potential of using movement data to this purpose has not yet been fully investigated. We therefore aim to implement a machine learning algorithm inspired by biological motor control to aid diagnosis of rare diseases by classifying data from standard kinematic clinical gait analysis.
Starting from 42-degrees-of-freedom time series of joint angles extracted from motion capture data with custom routines [3], we employ a Gaussian process-based temporal movement primitive algorithm [4] in order to reduce the data to sets of movement primitives and weight vectors that capture the essential characteristics of the gait movement. The primitives are participant (and disease) -independent and represent general human gait. The weights are participant-specific and thus contain disease-specific information. A weighted combination of the primitives can thus generate participant specific gait data. We then apply standard classification tools such as Support Vector Machines and Random Forests to the weights to distinguish the disease from the control gait. The primary goal is to reliably differentiate patients from age-matched controls in an existing data set on patients with Legg–Calvé–Perthes disease (LCPD). A secondary goal is to allow the classifier to expand the set of diseases using nonparametric methods such as the Dirichlet process.
Importantly, our movement primitive algorithm is inspired by current theories of biological motor control with a potential edge over standard algorithms in training on small case numbers. The temporal primitives are analogous to central pattern generators in the spinal cord [5], whereas the weights reflect activation of these central patterns by more central mechanisms in a hierarchical control scheme. In such a control scheme, disease-specific changes in weights may be caused directly by disease-specific influences on neural signaling, such as in the Stiff Person Syndrome [6], or indirectly through pain-avoidance in orthopedic conditions such as LCPD.
With further development, our approach holds potential for facilitating early detection and improving treatment strategies across a wide range of rare movement disorders and orthopedic conditions.},
howpublished = {Bernstein Conference},
keywords = {},
pubstate = {published},
tppubtype = {proceedings}
}
Of over 6.000 known rare diseases, a considerable portion involves motor symptoms [1]. Whereas aiding diagnosis by artificial intelligence based on non-motor symptoms has shown promise [2], the potential of using movement data to this purpose has not yet been fully investigated. We therefore aim to implement a machine learning algorithm inspired by biological motor control to aid diagnosis of rare diseases by classifying data from standard kinematic clinical gait analysis.
Starting from 42-degrees-of-freedom time series of joint angles extracted from motion capture data with custom routines [3], we employ a Gaussian process-based temporal movement primitive algorithm [4] in order to reduce the data to sets of movement primitives and weight vectors that capture the essential characteristics of the gait movement. The primitives are participant (and disease) -independent and represent general human gait. The weights are participant-specific and thus contain disease-specific information. A weighted combination of the primitives can thus generate participant specific gait data. We then apply standard classification tools such as Support Vector Machines and Random Forests to the weights to distinguish the disease from the control gait. The primary goal is to reliably differentiate patients from age-matched controls in an existing data set on patients with Legg–Calvé–Perthes disease (LCPD). A secondary goal is to allow the classifier to expand the set of diseases using nonparametric methods such as the Dirichlet process.
Importantly, our movement primitive algorithm is inspired by current theories of biological motor control with a potential edge over standard algorithms in training on small case numbers. The temporal primitives are analogous to central pattern generators in the spinal cord [5], whereas the weights reflect activation of these central patterns by more central mechanisms in a hierarchical control scheme. In such a control scheme, disease-specific changes in weights may be caused directly by disease-specific influences on neural signaling, such as in the Stiff Person Syndrome [6], or indirectly through pain-avoidance in orthopedic conditions such as LCPD.
With further development, our approach holds potential for facilitating early detection and improving treatment strategies across a wide range of rare movement disorders and orthopedic conditions.
Starting from 42-degrees-of-freedom time series of joint angles extracted from motion capture data with custom routines [3], we employ a Gaussian process-based temporal movement primitive algorithm [4] in order to reduce the data to sets of movement primitives and weight vectors that capture the essential characteristics of the gait movement. The primitives are participant (and disease) -independent and represent general human gait. The weights are participant-specific and thus contain disease-specific information. A weighted combination of the primitives can thus generate participant specific gait data. We then apply standard classification tools such as Support Vector Machines and Random Forests to the weights to distinguish the disease from the control gait. The primary goal is to reliably differentiate patients from age-matched controls in an existing data set on patients with Legg–Calvé–Perthes disease (LCPD). A secondary goal is to allow the classifier to expand the set of diseases using nonparametric methods such as the Dirichlet process.
Importantly, our movement primitive algorithm is inspired by current theories of biological motor control with a potential edge over standard algorithms in training on small case numbers. The temporal primitives are analogous to central pattern generators in the spinal cord [5], whereas the weights reflect activation of these central patterns by more central mechanisms in a hierarchical control scheme. In such a control scheme, disease-specific changes in weights may be caused directly by disease-specific influences on neural signaling, such as in the Stiff Person Syndrome [6], or indirectly through pain-avoidance in orthopedic conditions such as LCPD.
With further development, our approach holds potential for facilitating early detection and improving treatment strategies across a wide range of rare movement disorders and orthopedic conditions.