Data from: Aim high, stay private: Differentially private synthetic data enables public release of behavioral health information with high utility
Data files
Apr 17, 2026 version files 1.38 MB
-
README.md
19.88 KB
-
survey_codebook.xlsx
14.15 KB
-
survey_syn_5.csv
1.34 MB
Abstract
Sharing behavioral health and wearable data poses privacy challenges, as traditional de-identification remains vulnerable to re-identification. Differential privacy (DP) provides mathematical guarantees through a tunable privacy budget, ϵ. This study evaluates the feasibility of generating and releasing DP synthetic behavioral health data with high analytical utility, identifying practical ϵ values for public data sharing. We analyzed physiological data from wearable devices and self-reported data from Phase 1 of the Lived Experiences Measured Using Rings Study (LEMURS), which tracked sleep, stress, and well-being among first-year college students. Three DP synthetic data generators: AIM, MST, and PATECTGAN, were evaluated across privacy budgets ranging from ϵ = 1 to 100. Utility was assessed using L1/L2 errors, correlation, regression, UMAP, and assessed vulnerability via privacy attacks. Results: AIM outperformed MST and PATECTGAN in preserving both statistical and analytical properties of the original data. For the Survey dataset, the lowest marginal errors occurred at ϵ = 5 and 10. Correlation, regression, and UMAP analyses confirmed that AIM generated data closely replicated original relationships at moderate ϵ values. Choice of privacy budget is still an open question, and it is task-agnostic and dataset-specific. Moderate privacy budgets (5 ≤ ϵ ≤ 10) maintained key associations between physiological and psychological measures while ensuring privacy. AIM’s workload-aware design effectively allocated noise toward relevant features, enhancing performance. A privacy budget of ϵ = 5 offers a practical balance between data utility and participant privacy for LEMURS behavioral health data sharing.
Title: Aim High, Stay Private: Differentially Private Synthetic Data Enables Public Release of Behavioral Health Information with High Utility
Authors: Mohsen Ghasemizade, Juniper Lovato, Chris Danforth, Peter Sheridan Dodds, Laura S.P. Bloomfield, Matthew Price, Joseph Near
Institution: University of Vermont, Burlington, VT, USA
Corresponding Author: Mohsen Ghasemizade (mghasemi@uvm.edu)
Related Publication: Journal of the American Medical Informatics Association, 2025
Dataset Overview
This dataset is a differentially private (DP) synthetic version of the Survey dataset collected during Phase 1 of the Lived Experiences Measured Using Rings Study (LEMURS). LEMURS is a longitudinal study that recruited approximately 600 first-year college students at the University of Vermont beginning in Fall 2022. Participants completed weekly surveys and wore Oura rings to track sleep, physiological signals, stress, and general well-being.
The original Survey dataset contains 108 columns capturing behavioral, psychological, and lifestyle data. Because the original data cannot be shared publicly due to re-identification risks, this release provides a DP synthetic version generated using the AIM (Adaptive and Iterative Mechanism) algorithm at a privacy budget of ε (epsilon) = 5, which was identified as the optimal balance between data utility and participant privacy.
Differential Privacy (DP) is a mathematical framework that protects individual privacy by adding controlled random noise to the data. The privacy budget ε controls the trade-off: smaller ε = stronger privacy but less fidelity to the original data. A value of ε = 5 was selected because it maintained key statistical and analytical properties of the original dataset while resisting linkage and membership inference attacks.
Files Included
| Filename | Description |
|---|---|
survey_syn_5.csv |
Differentially private synthetic Survey dataset generated with AIM at ε = 5 |
survey_codebook.xlsx |
Codebook describing all column names, survey questions, and response encodings |
README.md |
This documentation file |
Dataset Structure
- File:
survey_syn_5.csv - Format: Comma-Separated Values (CSV)
- Unit of observation: One row = one weekly survey response for one participant
- Participant identifier:
record_id— a pseudo-anonymized numerical ID. Each participant may contribute multiple rows (one per study week). - Note: Because this is synthetic data,
record_idvalues do not correspond to real individuals. - Missing values: Represented as
N/A. These reflect questions that were not applicable or not answered for a given week (e.g., follow-up questions shown only if a prior question was answered "Yes"). - Qualitative (text-based) columns from the original dataset have been removed, as the DP algorithm operates only on quantitative values.
Variable Descriptions
All column names follow the format F1_[Domain]_[Variable], where F1 refers to Form 1 (the weekly survey instrument). Full question text and response codes for every column are provided in survey_codebook-2.xlsx.
A summary of the thematic domains is provided below.
Oura Ring App Usage (F1_App_*)
Questions about how participants used the Oura ring app during the week.
| Column | Description | Values |
|---|---|---|
F1_App_numtakeoff |
Number of times the Oura ring was removed during the week | Continuous (count) |
F1_App_sleeprecovery |
Used Sleep Recovery feature | 1 = Yes, 0 = No |
F1_App_meditation |
Used Meditation feature | 1 = Yes, 0 = No |
F1_App_breathwork |
Used Breathwork feature | 1 = Yes, 0 = No |
F1_App_periodpredictor |
Used Period Predictor feature | 1 = Yes, 0 = No |
F1_App_otherring |
Used another wearable ring | 1 = Yes, 0 = No |
F1_App_encourgesleep |
Ring encouraged earlier sleep | 1 = Yes, 0 = No |
F1_App_encouragemove |
Ring encouraged more movement | 1 = Yes, 0 = No |
F1_App_encourageworkout |
Ring encouraged more workouts | 1 = Yes, 0 = No |
F1_App_encouragetakeeasy |
Ring encouraged taking it easier | 1 = Yes, 0 = No |
F1_App_encourageother |
Ring encouraged other behavior | 1 = Yes, 0 = No |
Physical Activity (F1_Activity_*)
Based on the International Physical Activity Questionnaire (IPAQ).
| Column | Description | Values |
|---|---|---|
F1_Activity_numvigorous |
Days of vigorous activity in past 7 days | 0–7 (days/week) |
F1_Activity_nummoderate |
Days of moderate activity in past 7 days | 0–7 (days/week) |
F1_Activity_hoursminutes1 |
Hours of vigorous activity per day | Continuous (hours) |
F1_Activity_hoursminutes4 |
Minutes of vigorous activity per day | Continuous (minutes) |
F1_Activity_numwalking |
Days of walking ≥10 minutes in past 7 days | 0–7 (days/week) |
F1_Activity_hoursminutes2 |
Hours of walking per day | Continuous (hours) |
F1_Activity_hoursminutes3 |
Minutes of walking per day | Continuous (minutes) |
F1_Activity_hoursminutes5 |
Minutes of sitting on a weekday | Continuous (minutes) |
F1_Activity_hoursminutes6 |
Hours of sitting on a weekday | Continuous (hours) |
Caffeine Consumption (F1_Caffeine_*)
| Column | Description | Values |
|---|---|---|
F1_Caffeine_caffeinethisweek |
Consumed caffeine this week | 1 = Yes, 0 = No |
F1_Caffeine_coffee |
Drank coffee | 1 = Yes, 0 = No |
F1_Caffeine_tea |
Drank tea | 1 = Yes, 0 = No |
F1_Caffeine_soda |
Drank soda | 1 = Yes, 0 = No |
F1_Caffeine_energydrink |
Drank energy drink | 1 = Yes, 0 = No |
F1_Caffeine_workoutdrink |
Drank workout/pre-workout drink | 1 = Yes, 0 = No |
F1_Caffeine_otherdrink |
Drank other caffeinated drink | 1 = Yes, 0 = No |
F1_Caffeine_whencaffeine* |
Time of day caffeine was consumed | 1 = Yes, 0 = No (see codebook for time windows) |
F1_Caffeine_whycaffeine* |
Reason for caffeine consumption | 1 = Yes, 0 = No (see codebook for reason categories) |
Time Outdoors (F1_TimeOut_*)
| Column | Description | Values |
|---|---|---|
F1_TimeOut_Days |
Days spent in nature last week | 0–7 (days/week) |
F1_TimeOut_Hours |
Hours spent outdoors | Continuous (hours) |
F1_TimeOut_Minutes |
Minutes spent outdoors | Continuous (minutes) |
F1_TimeOut_When |
Time of day outdoors | 1=Early Morning, 2=Mid-Morning, 3=Early Afternoon, 4=Late Afternoon, 5=Early Evening, 6=Night |
F1_TimeOut_Cycling through F1_TimeOut_Other |
Type of outdoor activity | 1 = Yes, 0 = No (see codebook) |
F1_TimeOut_Location |
Type of outdoor environment | 1=Suburban, 2=Urban, 3=Rural |
Social Engagement (F1_Social_*)
| Column | Description | Values |
|---|---|---|
F1_Social_numbersocial |
Number of social engagements this week | Continuous (count) |
F1_Social_howmanyfriends |
Number of friends spent time with | Continuous (count) |
F1_Social_howmmanyinperson |
Number of in-person interactions | Continuous (count) |
F1_Social_videocall |
Number of video call interactions | Continuous (count) |
F1_Social_phone |
Number of phone interactions | Continuous (count) |
F1_Social_talkfamily |
Times saw or talked to family | Continuous (count) |
F1_Social_moreorless |
Satisfaction with social frequency | 1=More than desired, 2=Less than desired, 3=Just right |
Sleep (F1_Sleep_*)
| Column | Description | Values |
|---|---|---|
F1_Sleep_sleephours |
Average nightly sleep hours this week | Continuous (hours) |
F1_Sleep_sleepquality |
Self-rated sleep quality | 1=Very good, 2=Fairly good, 3=Fairly bad, 4=Very bad |
F1_Sleep_sleepamount |
Perceived adequacy of sleep | 1=Too little, 2=Adequate, 3=Ideal, 4=Too much |
Stress Events (F1_Stress_*)
| Column | Description | Values |
|---|---|---|
F1_Stress_stressevent |
Experienced a stressful event this week | 1 = Yes, 0 = No |
F1_Stress_describeevent |
Nature of the stressful event | 1=Social, 2=Romantic, 3=Familial, 4=Financial, 5=Academic, 6=Physical, 7=Mental |
F1_Stress_stressacademic |
Had an academic test this week | 1 = Yes, 0 = No |
F1_Stress_stresspaper |
Had a paper/project due this week | 1 = Yes, 0 = No |
Perceived Stress Scale (F1_PSS_*)
The PSS is a validated 10-item psychological scale measuring perceived stress over the past week. Higher total scores indicate greater perceived stress.
| Column | Description | Values |
|---|---|---|
F1_PSS_stressupset |
Upset by unexpected events | 0=Never, 1=Almost Never, 2=Sometimes, 3=Fairly Often, 4=Very Often |
F1_PSS_stresscontrol |
Felt unable to control important things | 0–4 (same scale) |
F1_PSS_stressnervous |
Felt nervous and stressed | 0–4 |
F1_PSS_stressconfident |
Felt confident handling problems | 0–4 |
F1_PSS_stressthings |
Felt things were going your way | 0–4 |
F1_PSS_stresscope |
Could not cope with required tasks | 0–4 |
F1_PSS_stressirritations |
Able to control irritations | 0–4 |
F1_PSS_stressontop |
Felt on top of things | 0–4 |
F1_PSS_stressangered |
Angered by things outside control | 0–4 |
F1_PSS_stressdifficulties |
Difficulties piling up uncontrollably | 0–4 |
Generalized Anxiety Disorder Scale (F1_GAD_*)
The GAD-7 is a validated 7-item scale measuring anxiety symptoms over the past two weeks.
| Column | Description | Values |
|---|---|---|
F1_GAD_nervous |
Feeling nervous, anxious, or on edge | 0=Not at all, 1=Several days, 2=More than half the days, 3=Nearly every day |
F1_GAD_worry |
Unable to stop or control worrying | 0–3 (same scale) |
F1_GAD_worrydifferent |
Worrying too much about different things | 0–3 |
F1_GAD_troublerelaxing |
Trouble relaxing | 0–3 |
F1_GAD_restless |
Too restless to sit still | 0–3 |
F1_GAD_annoyed |
Easily annoyed or irritable | 0–3 |
F1_GAD_afraid |
Feeling afraid something awful might happen | 0–3 |
F1_GAD_totalscore |
GAD-7 total score (sum of 7 items) | 0–21; 0–4=Minimal, 5–9=Mild, 10–14=Moderate, 15–21=Severe anxiety |
F1_GAD_howdifficult |
Difficulty functioning due to anxiety symptoms | 0=Not at all, 1=Several days, 2=More than half days, 3=Nearly every day |
Study Week
| Column | Description | Values |
|---|---|---|
week |
Study week number | Continuous (integer); week 1 = first week of the study semester |
Abbreviations
| Abbreviation | Definition |
|---|---|
| DP | Differential Privacy |
| ε (epsilon) | Privacy budget — controls the privacy-utility trade-off in DP |
| AIM | Adaptive and Iterative Mechanism — the DP synthetic data generator used |
| LEMURS | Lived Experiences Measured Using Rings Study |
| PSS | Perceived Stress Scale |
| GAD | Generalized Anxiety Disorder (scale) |
| GAD-7 | 7-item Generalized Anxiety Disorder scale |
| TST | Total Sleep Time |
| HR | Heart Rate |
| HRV | Heart Rate Variability |
| UMAP | Uniform Manifold Approximation and Projection |
| IPAQ | International Physical Activity Questionnaire |
| UVM | University of Vermont |
Data Generation Method
The synthetic data was generated using the AIM (Adaptive and Iterative Mechanism) algorithm, a state-of-the-art workload-aware DP synthetic data generator. AIM allocates noise strategically across the most informative marginal distributions of the data. The privacy budget was set to ε = 5, which was identified in the associated publication as the lowest ε at which:
- Marginal L1/L2 errors stabilized
- Spearman correlation structure was preserved
- Random forest regression R² scores plateaued (~0.64)
- UMAP cluster geometry closely matched the original
- Membership inference attacks (via TAPAS framework) showed no success
10 synthetic datasets were generated for each ε value tested (ε = 1, 2, 5, 10, 20, 50, 100), and the dataset with the lowest L1/L2 errors was selected for this release.
Privacy and Ethics
The original LEMURS dataset contains sensitive behavioral health and physiological data from college students and cannot be publicly released due to re-identification risks. This synthetic release was designed to enable open science while protecting participant privacy. Participants provided informed consent for data collection under IRB oversight at the University of Vermont. The DP guarantee at ε = 5 ensures that no individual record can be reliably re-identified, even when combined with auxiliary data.
Software and Reproducibility
- DP generator: AIM (via
smartnoise-sdkPython library) - Analysis environment: Python 3.x
- Key libraries:
pandas,numpy,scikit-learn,umap-learn,statsmodels,matplotlib,seaborn - Code for reproducing the analyses is available in the associated publication.
Citation
If you use this dataset, please cite the associated publication:
Ghasemizade M, Lovato J, Danforth C, Dodds PS, Bloomfield LSP, Price M, Near J. Aim High, Stay Private: Differentially Private Synthetic Data Enables Public Release of Behavioral Health Information with High Utility. Journal of the American Medical Informatics Association, 2026.
Human subjects data
This dataset has been de-identified using differential privacy, an additional layer of privacy after normal de-identification techniques.