TL;DR
The Neurobooth platform is an open-source, high-throughput system designed to capture time-synchronized, multimodal behavioral data (video, audio, motion, and eye-tracking) within a standard neurology clinic. Over its first two years, the project successfully collected rich "digital phenotypes" from 470 participants, proving that comprehensive digital assessment can be seamlessly integrated into clinical workflows to improve the diagnosis and monitoring of neurodegenerative diseases.
What Problem Does This Paper Address?
Traditional neurological examinations, while powerful, are often subjective and limited by human perception. A clinician might miss subtle changes in eye movement or minor irregularities in gait that signal the early stages of diseases like Parkinson’s or Ataxia.
Furthermore, existing digital phenotyping research is fragmented. Most studies are small, focus on a single domain (e.g., just gait or just speech), and use proprietary "black box" systems that don't talk to each other. This lack of standardization makes it nearly impossible to aggregate the large, high-dimensional datasets needed to train modern machine learning models for disease prediction and tracking.
Methods
The authors developed Neurobooth, a physical and digital infrastructure for "multimodal digital phenotyping."
- Hardware: A sound-insulated, wheelchair-accessible booth equipped with a sophisticated sensor array:
- Vision: Three Intel RealSense depth cameras (3D body tracking), a high-speed FLIR camera (facial expressions), and an iPhone 12.
- Oculomotor: An EyeLink Portable Duo for high-precision eye tracking.
- Motion: Five Mbient wearable inertial sensors (IMUs) on the wrists, ankles, and waist.
- Audio: A Blue Yeti microphone for high-fidelity speech capture.
- Software (NeuroboothOS): A custom Python-based library that coordinates stimulus delivery (via PsychoPy) and data collection. Crucially, it uses the Lab Streaming Layer (LSL) protocol to ensure all data streams—from the 1000Hz eye tracker to the 60Hz depth cameras—are perfectly time-synchronized.
- Tasks: Participants perform a 30–40 minute battery of tasks covering cognition (Multiple Object Tracking), oculomotor function (saccades, smooth pursuit), motor skills (finger-to-nose, sit-to-stand), and speech (passage reading).
Datasets
The paper details the first two years of the Neurobooth Dataset collected at Massachusetts General Hospital (MGH):
- Participants: 470 unique individuals, including 82 healthy controls and 388 patients with a wide range of neurological diagnoses (Parkinson’s, ALS, various Ataxias, etc.).
- Sessions: 782 total sessions, with 213 participants providing longitudinal data spanning up to 20 months.
- Clinical Context: The digital data is paired with "gold standard" clinical outcomes, including clinician-performed rating scales (like SARA or MDS-UPDRS) and 19 different Patient-Reported Outcome Measures (PROMs) ingested via REDCap.
Key Insights and Ideas
- The Power of Synchronization: By syncing eye-tracking with mouse movement or speech with facial video, researchers can study "cross-domain" coordination. For example, they can quantify how eye-voice coordination breaks down in Friedreich’s ataxia—a metric a human observer simply cannot measure.
- Clinical Feasibility: The study proves that a high-tech "booth" can exist inside a busy outpatient clinic. Most patients participated immediately before or after their scheduled appointments, and 81.1% reported they would be willing to do the full assessment again.
- Open Science as a Catalyst: By making the hardware specs and NeuroboothOS open-source, the authors aim to turn the platform into a "library" of behavioral tasks that other institutions can adopt, facilitating massive-scale data aggregation.
- Digital Signatures: The data visualizations reveal distinct "signatures" for different pathologies, such as the "round the houses" curved saccades characteristic of Progressive Supranuclear Palsy (PSP) or the segmented "saccadic pursuit" seen in Multiple System Atrophy (MSA).
Key Results
- Usability: The system earned a System Usability Scale (SUS) score of 78.4, significantly higher than the industry benchmark of 68.05 for digital health apps.
- High-Fidelity Capture: The platform successfully captured rare clinical signs digitally, such as downbeat nystagmus in CANVAS syndrome and hypermetric saccades in Gordon-Holmes syndrome.
- Efficiency: The system supports high-throughput recording across multiple connected computers, allowing for a comprehensive behavioral profile to be generated in under 40 minutes.
Limitations and Future Work
- Hardware Constraints: The authors noted occasional Bluetooth connectivity drops with the wearable sensors and reduced eye-tracking quality for some wheelchair users due to head-support limitations.
- Cognitive Load: Some tasks, particularly Multiple Object Tracking (MOT), proved too difficult or time-consuming for patients with significant cognitive impairment, leading the team to implement "early stopping" criteria.
- Scaling Down: While the current booth is a "gold standard" setup, the team is working on a "Neurobooth-lite" version that runs on a single laptop with consumer-grade sensors for home-based monitoring.
- Closing the Loop: Future work involves creating automated, real-time visualizations and reports that clinicians can use during the patient visit to inform treatment decisions.
Reference: Nunes, A. S., et al. (2024). Multimodal Digital Phenotyping of Behavior in a Neurology Clinic: Development of the Neurobooth Platform and the First Two Years of Data Collection. medRxiv preprint.
Summary updated