🤖 AI

Generating Synthetic Health Sensor Data for Privacy-Preserving Wearable Stress Detection

Original authors: Lucas Lange, Nils Wenzlitschke, Erhard Rahm

Published 2026-02-03

📖 5 min read🧠 Deep dive

Original authors: Lucas Lange, Nils Wenzlitschke, Erhard Rahm

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a smartwatch that acts like a personal bodyguard, constantly monitoring your heart rate, skin temperature, and how much you sweat to figure out if you are stressed. Doctors and researchers want to build better "bodyguards" (AI models) to help people, but there's a big problem: privacy.

To teach these AI models, you need real data from real people. But that data is like a diary of someone's deepest secrets. If you share it, you risk exposing private health information. If you try to hide the names (anonymize it), clever hackers can sometimes still figure out who the person is just by looking at their unique heartbeat patterns.

This paper proposes a clever workaround: Instead of using real people's data, let's create fake data that looks exactly like the real thing.

Here is how they did it, explained with some everyday analogies:

1. The Problem: The "Secret Diary" Dilemma

Researchers need a lot of data to train their stress-detection AI. But getting permission to use real people's medical diaries is hard, slow, and risky. It's like trying to teach a chef to cook a perfect steak, but you aren't allowed to let them taste the real meat because it's too expensive or sensitive.

2. The Solution: The "Master Forger" (GANs)

The authors used a technology called Generative Adversarial Networks (GANs). Think of this as a game between two artists:

The Forger (Generator): Tries to create fake stress data (heartbeats, sweat levels) that looks so real it could fool anyone.
The Detective (Discriminator): Tries to spot the difference between the real data and the fake data.

They play this game over and over. The Forger gets better at making fakes, and the Detective gets better at spotting them. Eventually, the Forger becomes so good that the Detective can't tell the difference. The result? A massive library of synthetic (fake) stress data that has all the statistical patterns of real human data but belongs to no actual person.

3. The Extra Safety Net: "The Noise Machine" (Differential Privacy)

Even with fake data, there's a worry: What if the Forger accidentally memorized a specific real person's diary and just copied it?

To stop this, the authors added Differential Privacy (DP). Imagine the Forger is working in a room filled with a thick, swirling fog. When they look at the real data to learn from it, the fog (mathematical noise) blurs the details just enough so they can learn the general shape of the data, but they can't see the specific details of any single person.

This ensures that even if someone steals the fake data, they cannot reverse-engineer it to find out who the original person was. The paper tested different levels of "fog" (privacy budgets):

Light Fog: Less privacy, very accurate fake data.
Thick Fog: Very strong privacy, but the fake data gets a bit "blurry" and less accurate.

4. The Results: Does the Fake Data Work?

The researchers tested if their "Master Forger" could actually help train a stress detector. They used a standard dataset called WESAD (which has data from 15 people) and tried two strategies:

Strategy A: The Full Swap (TSTR)
They threw away the 15 real people and trained the AI only on 15 fake people.
- Result: The AI still worked pretty well, proving you can replace real data with fake data.
Strategy B: The Booster Pack (AUGM)
They kept the 15 real people and added more fake people to the mix (like adding 100 fake students to a class of 15 real ones to help the teacher learn).
- Result: This was the winner. By adding the fake data, the AI got significantly smarter.
- The Big Win: When they used the "Thick Fog" (strong privacy) version, adding the fake data boosted the AI's performance by 11% to 15%. Without the fake data, the privacy rules usually make the AI much dumber. The fake data acted as a bridge, allowing the AI to learn effectively even while keeping strict privacy.

5. The Catch: The Privacy Trade-off

The paper admits that the "Thicker the Fog, the Blurrier the Picture."

When they used a very strict privacy setting (very high security), the fake data started to lose some of its "stress" characteristics. It was still safe, but it wasn't as perfect a copy of reality.
However, even with this blurriness, the fake data was still better than having no data at all.

Summary

The paper shows that we can build a privacy-preserving machine learning pipeline for stress detection. By using a "Forger" (GAN) trained with a "Noise Machine" (Differential Privacy), researchers can generate endless amounts of realistic, safe-to-share health data. This allows them to train better AI models without ever needing to peek at a real person's private medical diary.

Key Takeaway: You don't need to sacrifice privacy to get good medical AI. You can just teach the AI using a very convincing, mathematically safe "hallucination" of real data.

Technical Summary: Generating Synthetic Health Sensor Data for Privacy-Preserving Wearable Stress Detection

Problem Statement

The increasing utilization of smartwatch health sensor data for stress detection and patient monitoring faces a critical bottleneck: the sensitivity of medical data and the difficulty of acquiring sufficient, high-quality datasets for research. While wearable devices generate rich physiological time-series data, strict privacy regulations and the risk of re-identification from anonymized health signals limit data sharing. Furthermore, training robust machine learning models under Differential Privacy (DP) constraints often results in significant utility loss (reduced model performance) due to the noise introduced to protect individual records. The authors address the challenge of optimizing the utility–privacy trade-off by investigating whether synthetic data, generated via Generative Adversarial Networks (GANs) with DP safeguards, can augment or replace real data to improve stress detection models without compromising patient privacy.

Methodology

The study utilizes the publicly available WESAD dataset, which contains multimodal physiological signals from 15 participants recorded via Empatica E4 wristbands. The data includes six modalities: three-axis acceleration (ACC), electrodermal activity (EDA), skin temperature (TEMP), and blood volume pressure (BVP). The task is binary classification (stress vs. non-stress).

Data Preparation

Resampling and Alignment: Signals recorded at different frequencies were downsampled to a unified 1 Hz using Fourier methods to ensure alignment and smoothness.
Labeling: Neutral and amusement states were merged into a single "non-stress" class to create a binary classification problem.
Windowing: Data was segmented into 60-second windows with a 30-second sliding overlap to increase sample size and provide contextual information.
Normalization: Min-max normalization was applied to scale signals to [0, 1].

Generative Models

The authors trained three distinct GAN architectures to generate synthetic time-series data:

Conditional GAN (CGAN): A non-private model where both the generator and discriminator receive class labels (stress/non-stress) as auxiliary input. It utilizes LSTM layers to capture temporal dependencies and an FCN discriminator.
DoppelGANger (DGAN): A non-private model that decouples the generation of metadata (labels) from continuous measurements using separate generators and discriminators to handle multimodal correlations.
Differentially Private CGAN (DP-CGAN): A privacy-preserving variant of the CGAN. It employs Differentially Private Stochastic Gradient Descent (DP-SGD) specifically on the discriminator during training. Noise is injected into the gradients of the discriminator when processing real data to ensure the generated samples do not leak information about specific individuals.

Evaluation Strategies

The synthetic data was evaluated using two primary strategies within a Leave-One-Subject-Out (LOSO) cross-validation framework:

Train Synthetic, Test Real (TSTR): Models were trained exclusively on synthetic data generated by the GANs and tested on real WESAD data.
Augmentation (AUGM): Synthetic data was combined with real training data to augment the dataset size.

The study tested three privacy budgets ( $\epsilon$ ) for the DP-CGAN: 10 (relaxed), 1 (strong), and 0.1 (very strict).

Quality Assessment

Synthetic data quality was assessed via:

Visual/Statistical Analysis: PCA and t-SNE to visualize data distribution and clustering.
Correlation Analysis: Pearson correlation coefficients and p-values to verify if synthetic data preserved the relationships between signals and labels (e.g., the known positive correlation between EDA and stress).
Classifier Two-Sample Test (C2ST): A classifier trained to distinguish between real and synthetic data; lower accuracy indicates higher similarity between the two distributions.

Key Results

Synthetic Data Quality

Non-Private Models: Both CGAN and DGAN produced synthetic data that closely mimicked the statistical distribution and correlations of the original WESAD data. The CGAN showed slightly better alignment in t-SNE visualizations.
Privacy Impact: As the privacy budget tightened (lower $\epsilon$ $ϵ$ ), the quality of the synthetic data degraded.
- At $\epsilon = 10$ and $\epsilon = 1$ , the DP-CGAN maintained reasonable correlation structures, particularly for EDA and TEMP.
- At $\epsilon = 0.1$ , the model struggled to preserve the critical correlation between EDA and stress labels, shifting focus to other signals (ACC, TEMP) and introducing over-correlations in BVP. The distribution density of EDA also deviated significantly from the original data.

Stress Detection Performance

The synthetic data was used to train CNN, CNN-LSTM, and TSCT classifiers.

Non-Private Augmentation: Using non-private synthetic data (CGAN) provided a marginal improvement in F1-score (0.45%) over baseline models trained on real data alone.
Private Augmentation: The most significant findings were observed in the DP-constrained scenarios.
- Training classifiers with DP-CGAN synthetic data resulted in substantial F1-score improvements compared to training directly on real data under the same DP constraints.
- Specifically, in private training scenarios, the use of synthetic data led to an 11.90% to 15.48% increase in F1-score across different $\epsilon$ values.
- For example, at the strict $\epsilon = 0.1$ budget, the DP-CGAN augmented model achieved an F1-score of 84.19%, significantly outperforming the related work's DP-TSCT model (68.71%) and the non-augmented DP baseline.

Significance and Claims

The paper claims that differentially private synthetic data generation is a viable strategy to overcome the utility–privacy trade-off in wearable health monitoring.

Enhanced Utility under Privacy Constraints: The primary contribution is demonstrating that generating synthetic data under DP constraints allows researchers to train stress detection models with significantly higher performance than training directly on real data under the same privacy guarantees. The synthetic data effectively "dilutes" the privacy cost, allowing the model to learn robust patterns without exposing individual records.
Privacy-Preserving Data Availability: The approach enables the creation of large, labeled datasets for stress detection that can be shared freely without the ethical and legal hurdles of handling real patient data, provided the generation process adheres to DP.
Feasibility of Multimodal Generation: The study confirms that GANs can successfully generate complex, multimodal time-series data (ACC, EDA, TEMP, BVP) that retain the physiological correlations necessary for accurate stress classification, even when privacy noise is applied.

The authors conclude that while increasing privacy requirements ( $\epsilon \to 0.1$ ) inevitably impacts the fidelity of the synthetic data, the resulting utility gain for downstream classification tasks in privacy-preserving settings is substantial, offering a practical path forward for privacy-aware smart health research.