Hands-On Exploration of Child and Adolescent Mental Health with the Healthy Brain Network
Open Data Week NYC – March 27 2026
VIDEO | AUDIO | RECAP EN / ES / FR | INFO | INDEX
Speakers: Arianna Zuanazzi - Open Science and Research Collaboration Specialist, Child Mind Institute; Adam Santorelli - Software Engineer, Child Mind Institute; Nathalia Bianchini Esper - Software Engineer, Child Mind Institute
Moderator: Arianna Zuanazzi - Child Mind Institute
Introduction to the Child Mind Institute and the Healthy Brain Network
Arianna Zuanazzi introduced the Child Mind Institute as an independent nonprofit organization focused on transforming the lives of children and adolescents facing mental health and learning challenges through care, education, and scientific research. She explained that the presenting team works within the Institute’s science department, which focuses heavily on neuroscience research, open science initiatives, data sharing, and applied technology development.
Zuanazzi emphasized the scale of mental health challenges among children and adolescents:
Globally, approximately one in five children experience mental health or learning challenges
Around 70% of U.S. counties lack a child and adolescent psychiatrist
The average delay between symptom onset and treatment exceeds eight years
To address these gaps, the Child Mind Institute launched the Healthy Brain Network (HBN) initiative in 2015. HBN uses a community-referred recruitment model in which families participate because they have concerns about their child’s mental health. Participants receive free evaluations and recommendations, while researchers gain access to a large open scientific dataset.
The Healthy Brain Network has now collected data from more than 8,000 children and adolescents, primarily in New York City, creating one of the richest openly shared pediatric mental health datasets available.
Characteristics of the Healthy Brain Network Dataset
Zuanazzi reviewed the demographic and diagnostic composition of the HBN sample.
Participants range from:
5 to 22 years old
With the majority between ages 6 and 11
Because the sample is community-referred rather than randomly selected, more than 90% of participants have received at least one diagnosis.
The most common diagnoses include:
ADHD (over 70%)
Anxiety disorders (approximately 50%)
Learning disorders (approximately 40%)
Autism spectrum disorders, language disorders, and depression (approximately 15%)
Zuanazzi emphasized that the dataset is “multimodal,” meaning it combines many different forms of data collection, including:
Questionnaires
Cognitive testing
Language assessments
Emotional and psychological evaluations
Substance use and addiction data
Medical status data
EEG and eye-tracking
MRI imaging
Biological samples
Behavioral monitoring technology
The workshop specifically focused on behavioral monitoring through actigraphy.
Introduction to Actigraphy
Zuanazzi explained that actigraphy is a non-invasive method used to monitor physical activity and sleep continuously over time using wristwatch-like devices equipped with:
Accelerometers
Light sensors
Temperature sensors
Participants in the Healthy Brain Network wore these devices continuously for up to 30 days, producing second-by-second recordings of activity and sleep behavior.
The project used two scientific actigraphy devices:
Actigraph
GENEActiv
Zuanazzi contrasted scientific actigraphy watches with commercial smartwatches.
Scientific actigraphy devices provide:
Raw unprocessed data
Continuous high-frequency recordings
No interactive features or apps
Consumer smartwatches instead provide:
Pre-processed summaries
Steps
Heart rate estimates
Sleep summaries
Notifications and app functionality
Why Actigraphy Matters for Mental Health Research
Zuanazzi described actigraphy as particularly valuable for mental health research because many conditions — including ADHD, anxiety, and depression — are associated with distinct:
Physical activity patterns
Hyperactivity or inactivity
Sleep timing
Sleep interruption
Irregular movement behavior
Actigraphy provides objective real-world measurements rather than relying entirely on self-report questionnaires.
She emphasized several advantages:
Continuous monitoring over long periods
Naturalistic recording in homes and schools
Objective behavioral measurement
Compatibility with ecological momentary assessment data
Scalability and relatively low cost
Introduction to the Hands-On Coding Session
Adam Santorelli then led the technical portion of the workshop.
He explained that the hands-on demonstration used data from two Healthy Brain Network participants:
A 5-year-old girl with ADHD
A 14-year-old boy with ADHD
To keep processing manageable during the workshop, the data were truncated to one week rather than the full month-long recordings.
Santorelli introduced several key actigraphy concepts:
Tri-Axial Accelerometer Data
The devices record acceleration in three dimensions:
X-axis
Y-axis
Z-axis
measured relative to Earth’s gravity.
ENMO (Euclidean Norm Minus One)
ENMO is a common actigraphy metric measuring physical movement intensity.
Santorelli explained:
Values near zero indicate little movement or stillness
Larger values indicate vigorous movement
Angle Z
Angle Z represents the orientation of the wrist relative to the horizontal plane and is heavily used in sleep detection algorithms.
wristpy – Open Source Actigraphy Processing Software
Santorelli introduced wristpy, an open-source Python toolbox developed at the Child Mind Institute for processing raw actigraphy data.
The software:
Reads raw actigraphy files
Processes accelerometer recordings
Extracts movement metrics
Detects sleep
Produces analysis-ready CSV outputs
The workshop used Google Colab notebooks for hands-on demonstrations.
Participants installed wristpy using Python’s pip package manager and downloaded example datasets from cloud storage buckets.
Reading and Inspecting Raw Data
Santorelli demonstrated how wristpy loads raw watch data into Python objects.
The raw data included:
Acceleration measurements
Luminosity values
Temperature
Battery information
Non-wear indicators
The demonstration revealed that even one week of actigraphy data contains tens of millions of measurements.
Participants then visualized the raw accelerometer traces, observing:
High-activity daytime periods
Low-activity nighttime periods likely corresponding to sleep
Running the wristpy Processing Pipeline
Santorelli next demonstrated wristpy’s orchestrator pipeline, which automatically processes raw data into usable behavioral metrics.
The pipeline produced:
ENMO movement metrics
Physical activity classifications
Sleep status predictions
Angle Z measurements
Non-wear detection
Sleep period information
The system also exports:
CSV result files
JSON metadata files documenting processing parameters
Santorelli explained that researchers can customize many parameters, including:
Activity thresholds
Calibration methods
Epoch lengths
Activity metrics
Non-wear algorithms
Physical Activity Analysis
Using interactive Plotly visualizations, Santorelli compared activity patterns between the two participants.
Findings included:
The 5-year-old showed higher peaks of vigorous activity
The 14-year-old remained active later at night
Younger children generally displayed greater movement intensity
He also demonstrated how researchers can categorize activity into:
Sedentary
Light activity
Moderate activity
Vigorous activity
based on ENMO thresholds drawn from published literature.
Sleep Detection and Angle Z
Santorelli then demonstrated sleep detection outputs.
Sleep status strongly correlated with reduced fluctuations in Angle Z, reflecting reduced wrist movement during sleep.
The comparison between participants revealed expected developmental differences:
The 5-year-old generally slept earlier
The 14-year-old stayed awake later into the night
Challenges in Detecting Non-Wear Versus Sleep
A particularly important methodological discussion involved differentiating:
Very still sleep
Device non-wear
Santorelli showed nighttime periods where the participant appeared motionless for extended intervals. These periods could indicate either:
Deep still sleep
Removal of the watch before bed
The algorithm currently relies primarily on wrist-angle changes, making it difficult to distinguish perfectly still sleep from actual device removal.
Santorelli suggested that incorporating additional signals such as:
Temperature
Capacitive sensors
could improve future algorithms.
NotSleepy – Sleep Analysis Toolbox
Santorelli next introduced NotSleepy, another Child Mind Institute software package designed for extracting sleep metrics from wristpy outputs.
NotSleepy computes nightly sleep measures including:
Sleep duration
Time in bed
Wake after sleep onset (WASO)
Sleep efficiency
Number of awakenings
The software also handles daylight savings time adjustments, which Santorelli described as a surprisingly difficult technical problem in sleep analysis.
Computing Daily Physical Activity Metrics
Santorelli demonstrated how processed wristpy outputs can be aggregated into daily activity summaries using the Polars data-processing library.
Metrics included:
Sedentary duration
Light activity duration
Moderate activity duration
Vigorous activity duration
Moderate-to-vigorous physical activity (MVPA)
Total activity counts
The results again showed:
Higher vigorous activity in the 5-year-old participant
Greater total daily movement among younger children
Quality Control Challenges in Actigraphy Research
Nathalia Bianchini Esper then shifted the discussion toward quality control and research methodology.
She stressed that actigraphy analysis is not complete after automated processing because:
Watches may not be worn continuously
Algorithms can misidentify sleep
Manual review remains essential
Esper emphasized that most existing actigraphy pipelines still rely partly on:
Sleep diaries
Human annotation
Manual correction
ActiSleep Tracker
To simplify manual quality control, the Child Mind Institute developed another open-source tool called ActiSleep Tracker.
The web-based application allows researchers to:
Visualize ENMO and Angle Z signals
Inspect nightly sleep predictions
Correct sleep windows manually
Annotate naps
Adjust wake-up times
Export corrected annotations automatically
Esper demonstrated how researchers can move sliders interactively to correct algorithm-generated sleep periods.
For example, one participant showed an unrealistically long still period during the night that appeared suspiciously like non-wear rather than sleep. Researchers could manually adjust the wake-up time and automatically generate corrected CSV annotations.
Diagnoses and Comorbidity in the Dataset
Esper reviewed diagnostic overlap within the Healthy Brain Network.
Many participants have multiple diagnoses simultaneously, including combinations such as:
ADHD and autism
ADHD and anxiety
ADHD and learning disorders
She demonstrated how researchers can select subgroups depending on their study goals.
Importance of Seasonality
Esper highlighted seasonality as a frequently overlooked variable in actigraphy research.
Children wearing watches during:
Winter
Summer camp periods
School sessions
may show dramatically different activity patterns unrelated to diagnosis.
Researchers therefore need to consider:
Enrollment season
School schedules
Environmental context
when analyzing physical activity data.
Compliance Analysis
Esper presented compliance analyses showing how many participants meet different thresholds for valid data collection.
Examples included:
Participants wearing watches continuously for eight days
Participants wearing watches for only partial days
Sleep-only compliance thresholds
Less restrictive thresholds naturally increase usable sample sizes.
Sleep and Activity Findings Across Age Groups
Esper showed several aggregate findings from the HBN actigraphy data.
Sleep Timing
Older children and adolescents:
Tend to go to bed later
Still wake up at similar weekday times due to school schedules
ENMO Activity Trajectories
Average ENMO curves showed:
Younger children move more during the day
Older children show lower activity levels
Weekend activity patterns differ from weekday patterns
Sedentary and Vigorous Activity
Older children spend:
More time in sedentary behavior
Less time in vigorous activity
Girls in the HBN dataset also tended to move less than boys on average.
Open Science and Future Data Release
Esper concluded by emphasizing the broader goals of the project:
Open science
Data sharing
Reproducible research
Low-cost behavioral monitoring tools
The team announced that the Healthy Brain Network actigraphy dataset is expected to become publicly available during the summer of 2026.
Closing Remarks
Arianna Zuanazzi closed the workshop by thanking Open Data Week organizers, BetaNYC, NYC OTI, and Data Through Design. She reminded participants that all workshop resources — including notebooks, software tools, slides, and repositories — were shared publicly and encouraged attendees to contact the Child Mind Institute team with additional questions.
RESOURCES
Child Mind Institute — independent nonprofit behind the workshop, focused on children’s mental health, open science, and open data
Healthy Brain Network (HBN) — community-referred research initiative providing no-cost evaluations and the multimodal open dataset used in the workshop
HBN Impact — overview of the biobank’s scale and the research it has enabled
wristpy — CMI’s open-source Python toolbox for processing raw actigraphy data, demoed by Adam Santorelli
wristpy (JOSS paper) — peer-reviewed Journal of Open Source Software publication describing the wristpy package
ActiSleep Tracker — CMI web app for manually annotating and quality-controlling actigraphy sleep predictions, demoed by Nathalia Bianchini Esper
actfast — fast Rust-based actigraphy file reader underlying the wristpy pipeline
childmindresearch on GitHub — CMI’s open-source repositories, including the actigraphy and sleep-analysis toolboxes
Polars — fast DataFrame library used in the workshop to compute daily physical-activity metrics
Plotly for Python — interactive plotting library used to visualize ENMO and sleep traces