DystoniaDBSNet: A novel deep learning biomarker of predictive treatment outcomes in dystonia

12 

Abstract Submission 

Dongren Yao¹, Harith Akram², Kailash Bhatia³, Thomas Foltynie², Patricia Limousin², Eoin Mulroy², Nutan Sharma⁴, Joshua Wong⁵, Jun Yu⁶, Ludvic Zrinzo², Kristina Simonyan⁷

¹Mass Eye and Ear, Harvard Medical School, Boston, MA, ²UCL Queen Square Institute of Neurology, London, UK, ³UCL Queen Square Institute of Neurology, Longon, UK, ⁴Harvard Medical School, Boston, MA, ⁵University of Florida, Gainesville, FL, ⁶Univeresity of Florida, Gainesville, FL, ⁷Mass Eye and Ear, Mass General Hospital, Harvard Medical School, Boston, MA

Dongren Yao  
Mass Eye and Ear, Harvard Medical School
Boston, MA

Harith Akram, MD  
UCL Queen Square Institute of Neurology
London, UK

Kailash Bhatia, MD  
UCL Queen Square Institute of Neurology
Longon, UK

Thomas Foltynie, MD  
UCL Queen Square Institute of Neurology
London, UK

Patricia Limousin, MD  
UCL Queen Square Institute of Neurology
London, UK

Eoin Mulroy, MD  
UCL Queen Square Institute of Neurology
London, UK

Nutan Sharma, MD, PhD  
Harvard Medical School
Boston, MA

Joshua Wong, MD  
University of Florida
Gainesville, FL

Jun Yu, MD  
Univeresity of Florida
Gainesville, FL

Ludvic Zrinzo, MD  
UCL Queen Square Institute of Neurology
London, UK

Kristina Simonyan  
Mass Eye and Ear, Mass General Hospital, Harvard Medical School
Boston, MA

Dystonia is a debilitating neurological disorder characterized by involuntary sustained or intermittent muscle contractions causing abnormal movements, postures, or both. Among the established therapeutic options is invasive circuit-based neuromodulation with deep brain stimulation (DBS) of bilateral globus pallidus (GPi), which yields an average of 30%-60% clinical improvement in different forms of dystonia. Yet, only about 5% of dystonia patients undergo DBS surgery and, among those treated, around 25% of patients have poor response. A limiting factor in patient selection for successful DBS treatment is the absence of a pathophysiologically relevant biomarker to inform treatment outcomes prior to neurosurgical intervention.

We developed and tested a deep learning algorithm, DystoniaDBSNet, which uses a structural brain MRI of patients who underwent DBS-GPi to automatically identify the neural biomarker of predictive treatment efficacy. Whole-brain T1-weighted MRIs from 130 patients with focal, multifocal, segmental, or generalized dystonia (64 M/66 F, age 45.78±18.62 years) treated at Massachusetts General Hospital, University College London, and the University of Florida were included in this study. Clinical improvement was defined as at least a 30% reduction of symptom severity based on the standardized Burke-Fahn-Marsden Dystonia Rating Scale. The DystoniaDBSNet model was trained and validated using phenotype-, sex-, age-, and surgical site-matched patient cohorts, allocating 80% of patients for training and 20% for testing.

The training model of DystoniaDBSNet achieved the area under the receiver operating characteristic curve (AUC) of 100% in discriminating DBS benefiting from non-benefiting patients. DystoniaDBSNet automatically identified a neural biomarker of DBS treatment outcome, which included clusters in the bilateral precentral and middle frontal gyri, left superior frontal gyrus, anterior cingulate cortex, thalamus, and right postcentral gyrus. In the testing set of patients with different clinical phenotypes of dystonia, DystoniaDBSNet achieved an overall accuracy of 96.0%, with 100% sensitivity, 85.7% specificity, and a 3.87% referral rate in predicting the DBS treatment outcome.

DystoniaDBSNet yielded a fully automated, objective, and highly accurate predictive outcome of DBS treatment in patients with different forms of dystonia from a single structural MRI that was collected prior to neurosurgical intervention. The components of the DystoniaDBSNet biomarker included brain regions known for their contribution to dystonia pathophysiology. The translational significance of DystoniaDBSNet is in its potential to enhance clinical decision-making in DBS candidate selection and ultimately to deliver improved clinical care to patients with dystonia.

Deep Brain Stimulation ¹

Other Methods ²

DISORDERS

Machine Learning

Movement Disorder

STRUCTURAL MRI

Other - dystonia