Leveraging Machine Learning to Effortlessly Track Patient Movement in the Clinic.
- Conditions
- Movement Disorders
- Registration Number
- NCT04074772
- Lead Sponsor
- University of Colorado, Denver
- Brief Summary
The objective of this study is the development of a system that will allow for the precise measurement of movement kinematics in a clinical exam setting using natural video from three cameras and machine learning to track points of interest. The investigators aim to implement such system in an unobtrusive and simply-incorporated way into the physical exam to provide exact, objective measures to detect patient movement abnormalities in ways not feasible with current tracking technologies.
- Detailed Description
Aim 1: Develop 3D tracking capable of capturing behavior of healthy controls during physical exams. In aim 1, the investigators will recruit healthy volunteers to perform a simplified physical exam in a replica exam room while being recorded with three synchronized FLIR cameras. The simplified exam will consist of four tasks: assessment of tremor, finger chase, finger-to-nose movements, and finger tapping. Study staff will then use DeepLabCut (DLC) software -technology that trains artificial neural networks to identify user defined features in an image - to recognize body parts of interest in physical exam videos. Once the network is fully trained the investigators will test its ability to generalize on different patients and different contexts. Additional analysis of volunteers' movement during the physical exam will be performed to assess for characteristics such as tremor, speed, and tortuosity of movement.
Aim 2: Apply 3D tracking to the clinic to track physical exam behaviors in motor disorder patients. In aim 2, the investigators will apply the trained network to the clinic to examine the physical exam characteristics of movement disorder patients. Aim 2a will test the DLC network's ability to capture movement disorder abnormalities during the physical exam in patients and healthy age-matched controls. DLC scores of each test variable will be compared to the physician's score of movement according to a standardized scale. The investigators expect to find that the DLC tracking method is able to objectively score movement disorders in ways that mirror and surpass the ability of the physician. In Aim 2b, the investigators will explore the population of recruited patients to see whether it is possible to pull out characteristic movements that correspond to certain disease states. In this exploratory aim, the investigators expect to be able to separate different disease groups (e.g.: Parkinsonian and ataxic patients) from each other based simply on the tracked movement characteristics.
Research Methods:
In Aim 1, a movement arena will be built on the University of Colorado Denver Graduate School campus using three FLIR cameras with a custom built synchronization and initiation system. The investigators will recruit up to 30 healthy 18-70-year-old controls from the University of Colorado Denver Graduate School to perform the simplified physical exam (assessment of tremor, finger chase, finger-to-nose movements, and finger tapping) while video is captured from three angles at 100 Hz. The investigators expect this testing to take no more than 5 minutes per subject. This video will be used to train the DLC artificial neural network to recognize limb features. The investigators will measure the ability of our trained DLC network to characterize twelve points of interest on each limb during a physical exam: the tips of the four fingers and the thumb, all four metacarpophalangeal joints, the center of the hand, the elbow, and the shoulder. A successful outcome will be a network that maintains the ability to recognize features of interest at high confidence between different individuals and different room contexts.
In Aim 2a, a tracking arena will be set up in a University of Colorado Movement Disorder Clinic exam room. The investigators will recruit up to 100 patients between 18-70 years old that are visiting for a movement disorder related appointment as well as spouses and relatives of the patients at the appointment for healthy age-matched controls. Patients in the clinic will be asked after their visit if they would like to participate in the study. If they consent, the physician will obtain written consent and fill out a patient form that includes the patient's age, race, sex, and diagnosis (or putative diagnosis). Video recording will be started and the physician will perform the simplified physical exam mentioned above. The physicians will judge the finger chase and finger-to-nose task as is described in the Scale for the Assessment and Rating of Ataxia (SARA, items 5 \& 6) from 0-4. The postural tremor and finger tapping will be judged according to the Unified Parkinson Disease Rating Scale (UPDRS, items 21 \& 23) from 0-4. If the patient is visiting with a person that consents to be an age-matched control (within 10 years of the patient's age) the physical exam will be repeated as above. The investigators expect this testing to take no more than 5 minutes per subject, beginning to end. The investigators will then use the DLC algorithm to score the physical exam in a way analogous to the physician scoring to assess the accuracy of the system.
In Aim 2b, the investigators will explore the patient data from Aim 2a for movement features specific to individual diseases. Data clustering methods (PCA and t-SNE) will be used to separate data into groups using high-dimensional DLC tracking data from each physical exam task. Success will be measured as the ability to separate diseases from one another based solely on the analysis of movement data.
Recruitment & Eligibility
- Status
- COMPLETED
- Sex
- All
- Target Recruitment
- 25
- Healthy controls: within age range
- Age-Matched controls: within age range
- Movement Disorder Patients: have diagnosed or putative movement disorder
- Healthy controls: have diagnosed or putative movement disorder; outside of age range
- Age-Matched controls: have diagnosed or putative movement disorder; outside of age range
- Movement Disorder Patients: outside of age range
Study & Design
- Study Type
- OBSERVATIONAL
- Study Design
- Not specified
- Primary Outcome Measures
Name Time Method Successful Tracking Achieved in Clinic On the day of physical exam If the neural network can generalize to different patients and contexts with accurate tracking, such that it can track all 12 points of interest with \>99% accuracy in \>95% of frames of novel video data, the investigators will consider this outcome a success.
- Secondary Outcome Measures
Name Time Method Identification of Diseases by Movement Tracking On the day of physical exam If the investigators can separate different movement disorders from one another based on tracking data alone this outcome will be considered a success. Specifically, in LOTO cross validation, individual patient data must be assigned the correct disease state with 95% confidence.
Trial Locations
- Locations (1)
University of Colorado Hospital
🇺🇸Aurora, Colorado, United States