DxToniaNet: A Neural Network for Estimating the Severity of Leg Dystonia in Children with Cerebral Palsy

Importance: Leg dystonia affects 50-80% of children with cerebral palsy (CP), but severity assessment remains subjective and time-consuming. Objective tools are urgently needed to quantify leg dystonia severity in clinical care and clinical trials. Objective: To develop and validate a deep neural network (DxToniaNet) for leg dystonia severity estimation using 2D clinical video recordings. Design, Setting, and Participants: We retrospectively analyzed video recordings of 298 children with cerebral palsy undergoing a seated hand open-close task as part of routine clinical care (Center 1: 193 videos, 720p/30fps; Center 2: 105 videos, 4K/60fps). Videos were rated for leg dystonia severity by a group of 7 experts using the Global Dystonia Rating Scale (GDRS) leg subscore (range 0-40) which has a minimum clinically important difference (MCID) of 4 points. Videos displaying leg dystonia (median expert GDRS score ≥1) were used to develop and test a machine learning model (DxToniaNet) for estimating leg dystonia severity. Open-source software (OpenPose) estimated bilateral knee, ankle, and toe coordinates, normalized for height and camera positioning, and 69 variance and amplitude features were calculated based on established expert-cited features of leg dystonia. A neural network, DxToniaNet, was developed, trained and validated using 5-fold cross-validation on Center 1 videos before final testing on independent Center 2 videos. Main Outcome(s) and Measure(s): Mean absolute error (MAE), root mean square error (RMSE), and accuracy in differentiating between videos with clinically meaningful differences in GDRS scores were calculated relative to experts' median GDRS scores. Results: When estimating experts' leg GDRS scores in testing data, the proposed machine learning (ML) model yielded a MAE of 1.71 ± 1.68 and RMSE of 2.21 ± 0.34 (significantly less than the GDRS MCID of 4), and 90-94% accuracy in differentiating between videos with greater than a 4-point difference in the GDRS. Conclusions and Relevance: We have demonstrated that DxToniaNet can accurately estimate leg dystonia severity compared to experts' assessments with a small (clinically insignificant) error margin. Our methodology supports generalizability since DxToniaNet was developed and tested using independent videos datasets from different centers, recorded under different conditions. In the future, DxToniaNet may help facilitate standardized leg dystonia severity estimation in clinical care and clinical trials.

Authors

Institutions

Publication Details

Journal
medRxiv
Published
2026-11-13
DOI
https://doi.org/10.64898/2026.11.11.26367859
Primary Topic
Cerebral Palsy and Movement Disorders
Type
preprint

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

DxToniaNet: A Neural Network for Estimating the Severity of Leg Dystonia in Children with Cerebral Palsy

Bhooma R. Aravamuthan, Keerthana Chintalapati, Sushma Gandham, Jennifer A. O’Malley et al.
2 links
medRxiv
Cerebral Palsy and Movement Disorders
preprint

DxToniaNet: A Neural Network for Estimating the Severity of Leg Dystonia in Children with Cerebral Palsy

Bhooma R. Aravamuthan, Keerthana Chintalapati, Sushma Gandham, Jennifer A. O’Malley, Dararat Mingbunjerdsuk, Rishabh Bajpai, Susie Kim, Rose Gelineau‐Morel, Timothy J. Feyma, Jeff L. Waugh, Joel S. Perlmutter, Michael C. Kruer, Joanna Blackburn, Laura S. Tochen, Emma Lott, Alyssa Rust, Steve Wu
preprint en

Abstract

Importance: Leg dystonia affects 50-80% of children with cerebral palsy (CP), but severity assessment remains subjective and time-consuming. Objective tools are urgently needed to quantify leg dystonia severity in clinical care and clinical trials. Objective: To develop and validate a deep neural network (DxToniaNet) for leg dystonia severity estimation using 2D clinical video recordings. Design, Setting, and Participants: We retrospectively analyzed video recordings of 298 children with cerebral palsy undergoing a seated hand open-close task as part of routine clinical care (Center 1: 193 videos, 720p/30fps; Center 2: 105 videos, 4K/60fps). Videos were rated for leg dystonia severity by a group of 7 experts using the Global Dystonia Rating Scale (GDRS) leg subscore (range 0-40) which has a minimum clinically important difference (MCID) of 4 points. Videos displaying leg dystonia (median expert GDRS score ≥1) were used to develop and test a machine learning model (DxToniaNet) for estimating leg dystonia severity. Open-source software (OpenPose) estimated bilateral knee, ankle, and toe coordinates, normalized for height and camera positioning, and 69 variance and amplitude features were calculated based on established expert-cited features of leg dystonia. A neural network, DxToniaNet, was developed, trained and validated using 5-fold cross-validation on Center 1 videos before final testing on independent Center 2 videos. Main Outcome(s) and Measure(s): Mean absolute error (MAE), root mean square error (RMSE), and accuracy in differentiating between videos with clinically meaningful differences in GDRS scores were calculated relative to experts' median GDRS scores. Results: When estimating experts' leg GDRS scores in testing data, the proposed machine learning (ML) model yielded a MAE of 1.71 ± 1.68 and RMSE of 2.21 ± 0.34 (significantly less than the GDRS MCID of 4), and 90-94% accuracy in differentiating between videos with greater than a 4-point difference in the GDRS. Conclusions and Relevance: We have demonstrated that DxToniaNet can accurately estimate leg dystonia severity compared to experts' assessments with a small (clinically insignificant) error margin. Our methodology supports generalizability since DxToniaNet was developed and tested using independent videos datasets from different centers, recorded under different conditions. In the future, DxToniaNet may help facilitate standardized leg dystonia severity estimation in clinical care and clinical trials.

medRxiv
Northwestern University (US), Seattle Children's Hospital (US), George Washington University (US), University of Washington (US), Washington University in St. Louis (US), Gillette Children's Specialty Healthcare (US), University of Arizona College of Medicine- Phoenix (US), University of Cincinnati (US), University of Missouri–Kansas City (US), The University of Texas Southwestern Medical Center (US), Stanford University (US)
Pediatric Epilepsy Research Foundation, National Institute of Neurological Disorders and Stroke
Good health and well-being
Cerebral Palsy and Movement Disorders
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon