Artificial Intelligence for the Analysis of Video Data of Facial Movement, with a Focus on Myasthenia Gravis

The Face of Neuromuscular Dysfunction: Artificial Intelligence for the Analysis of Video Data of Facial Movement, with a Focus on Myasthenia Gravis

Rationale: Myasthenia Gravis (MG) is an autoimmune disorder (AID) with antibodies against the NMJ, resulting in various degrees of muscle fatigability and weakness. All striated muscles can be involved, although the extra-ocular muscles are most commonly affected, giving rise to a fluctuating ptosis and diplopia. Facial muscles are also commonly affected, resulting in eye closure weakness, difficulty chewing and swallowing or speech impairments. Antibodies against the acetylcholine receptor (AChR) are present in over 80% of generalized MG patients. In the pure ocular form, AChR antibodies are detectable in nearly 50% of all patients. In approximately 4%, antibodies against the postsynaptic muscle-specific receptor tyrosine kinase (MuSK) are found and in 15% of the patients with generalized disease, no serum antibodies are detected1-3. Approximately 15% of AChR MG patients has a thymoma, in which case the disease can be classified as a paraneoplastic syndrome2. With a prevalence of 1 to 2 per 10.000, MG is considered a rare disease2.

The rarity of MG can make it difficult to diagnose, specifically for general Neurologists who are likely to encounter a patient with MG only a handful of times throughout their career. In addition, the fluctuating nature of the disease makes it difficult to make appropriate treatment decisions, especially as patients throughout the country are usually treated at one specialized center (in the Netherlands, the LUMC). Currently, patients who are in doubt whether they are experiencing an exacerbation have to make an appointment and travel for several hours to undergo assessment by their specialized Neurologist. An objective, reliable biomarker for disease severity that can be used at home would therefore greatly improve quality of life for many MG patients. Emerging possibilities in modern technologies can support doctors with all kinds of medical challenges, like offering diagnostic support, treatment decisions or patient follow-up. A technology of special interest for this study is advanced facial recognition. We aim to study the ability of existing software (FaceReader, Noldus) versus a deep learning model specifically developed for this purpose by the group of Jan van Gemert at the TU Delft to differentiate between healthy controls and patients with MG and between MG patients with different levels of disease severity.

Primary objectives:

To determine and compare the diagnostic yield of two different methods (FaceReader technology and a deep learning model specifically developed for video data) to analyse facial weakness from video recordings (04:00m) with different standardized facial expressions to:

1. Differentiate between MG patients and healthy controls.
2. Differentiate between mild and moderate to severe disease severity.

Study Overview

• Status: Completed
• Conditions: Myasthenia Gravis

• Study Type: Observational
• Enrollment (Actual): 90

Contacts and Locations

Study Locations

• Netherlands
  • Zuid-Holland
    • Leiden, Zuid-Holland, Netherlands, 2333ZA
      • Leiden University Medical Center

Participation Criteria

Eligibility Criteria

• Ages Eligible for Study: Adult; Older Adult
• Accepts Healthy Volunteers: Yes

• Sampling Method: Non-Probability Sample

Study Population

To answer the main objectives we will obtain a 04:00 min standardized facial video recording of MG patients in the outpatient clinic and Neurology ward, recorded at the LUMC. For the webcam recordings we will recruit participants from the Dutch-Belgian MG registry (P15.287). Informed consent forms of the MG registry with permission to be approached for future research are available.

We will ask spouses and family to enrol as healthy controls, as well as employees of our department. The healthy control group will be age- and gender matched. Healthy controls are defined as subjects without medical conditions or medication which may influence the facial muscles, e.g. Graves thyroid disease, previous facial palsy due to a stroke or prednisone use. For the webcam recordings we will ask MG patients to introduce a healthy control of the same age.

Description

Inclusion Criteria:

• Male or female participants aged ≥ 18 years
• Subjects must understand the requirements of the study and provide written informed consent.

MG

• Clinical signs or symptoms suggestive of MG and at least one of the following:
• A serologic test for AChR antibodies or MuSK antibodies or
• A diagnostic electrophysiological investigation supportive of the diagnosis MG or
• A positive neostigmine test Healthy control group
• Volunteers from spouses, friends and family accompanying patient or employees from our department
• No medical conditions affecting the facial muscles, e.g. Graves' disease, previous stroke with a facial palsy
• No use of medication affecting the facial features, e.g. prednisone

Exclusion Criteria:

• Inability to give written informed consent
• Inability to read Dutch/ English video-instructions
• Participants with active Graves' disease or unilateral facial paralysis

Study Plan

How is the study designed?

Number of groups / cohorts

2

Cohorts and Interventions

Group / Cohort
healthy controls
myasthenia gravis

What is the study measuring?

Primary Outcome Measures

Outcome Measure	Measure Description	Time Frame
FaceReader	The primary parameter for the differentiation between healthy controls and MG and between different grades of MG disease severity with FaceReader is the diagnostic yield of individual muscles and combinations of muscles. The diagnostic yield is expressed as sensitivity, specificity and area under the curve of a receiver-operator curve (ROC) of the FaceReader algorithm. For this the quantitative data of facial weakness expressed in Action Units (AU), ranging between 0 (no activation) and +1 (maximal activation) will be used. Raw data from FaceReader provides the results of 20 AU's corresponding with 20 different facial movements of 20 facial muscles based on the Facial Action Coding System (FACS).	2020-2023
Narrow deep learning model	The primary parameter for the differentiation between healthy controls and MG and between different grades of MG disease severity with a working narrow deep learning model is the diagnostic yield. The diagnostic yield is expressed as sensitivity, specificity and area under the curve of a receiver-operator curve (ROC).	2020-2023
Disease severity	For comparison between different levels of disease severity the QMG score will serve as the gold standard. Groups based on disease severity are composed as following: mild QMG 0-9, moderate QMG 10-16 and severe QMG >16. For the home recording we will use the MG-Activities of Daily Living (MG-ADL) to measure disease severity since the QMG requires the physical presence of the patient. This is a commonly used tool in clinical trials. Groups based on disease severity are composed as following: mild MG-ADL 0-4, moderate-severe MG-ADL ≥5.	2020-2023

Secondary Outcome Measures

Outcome Measure	Measure Description	Time Frame
Longitudinal changes	Detection of medication effects by obtaining multiple videos (longitudinal) in a subset of patients. The QMG score or MG-ADL is the parameter for change in disease severity. A previous study found a minimal clinically important difference (MCID) in QMG score of ≥2 for a baseline QMG score between 0 and 16. For a baseline QMG score >16 the MCID is ≥3 points change in QMG score4. For the MG-ADL the MCID is a change of ≥2 points5. For detection of medication effects, our aim is to detect an intra-participant change in QMG ≥2 or ≥3, depending of baseline QMG score. For the home recorded group our aim is to detect a change in MG-ADL score of ≥2 points.	2020-2023
FaceReader vs deep learning model	A comparison of the diagnostic yield of FaceReader parameters and classification by the deep learning model.	2020-2023
site versus home recording	A comparison of the diagnostic yield of FaceReader and the deep learning model of videos recorded in the standardized LUMC setting (green screen, lights, 4K camera) and home recorded videos using a webcam.	2020-2023

Collaborators and Investigators

• Sponsor: Leiden University Medical Center
• Collaborators: Delft University of Technology
• Investigators: Principal Investigator: Martijn Tannemaat, MD, PhD, Leiden University Medical Center

Study record dates

Study Major Dates

• Study Start (Actual): December 1, 2020
• Primary Completion (Actual): December 31, 2022
• Study Completion (Actual): December 31, 2022

Study Registration Dates

• First Submitted: August 3, 2022
• First Submitted That Met QC Criteria: March 4, 2025
• First Posted (Actual): March 25, 2025

Study Record Updates

• Last Update Posted (Actual): March 25, 2025
• Last Update Submitted That Met QC Criteria: March 4, 2025
• Last Verified: March 1, 2025

More Information

Terms related to this study

• Additional Relevant MeSH Terms: Neurologic Manifestations; Musculoskeletal Diseases; Nervous System Diseases; Muscular Diseases; Neuromuscular Manifestations; Pathologic Processes; Neoplasms by Site; Neoplasms; Neuromuscular Diseases; Autoimmune Diseases; Immune System Diseases; Autoimmune Diseases of the Nervous System; Neurodegenerative Diseases; Paraneoplastic Syndromes, Nervous System; Nervous System Neoplasms; Paraneoplastic Syndromes; Neuromuscular Junction Diseases; Muscle Weakness; Myasthenia Gravis
• Other Study ID Numbers: P20.083

Plan for Individual participant data (IPD)

• Plan to Share Individual Participant Data (IPD)?: YES
• IPD Plan Description: Raw data, including de-identified individual participants' data, that support the findings of this study are available from the corresponding author, upon reasonable request. Privacy laws preclude the sharing of facial videos of individual study participants.
• IPD Sharing Supporting Information Type: STUDY_PROTOCOL; SAP; ICF; CSR

Drug and device information, study documents

• Studies a U.S. FDA-regulated drug product: No
• Studies a U.S. FDA-regulated device product: No
• product manufactured in and exported from the U.S.: No