TIBECG: Temporal Information Bottleneck for Lead-Specific Demographic Privacy in ECG Waveform Release

Tsai-Min Chen, Jun SeitaarXiv 2026 · arXiv

TL;DR

TIBECG introduces a "temporal information bottleneck" to protect demographic privacy in ECG waveforms released for remote diagnostic classification. By architecturally restricting internal time resolution, it successfully suppresses age and sex inference by a median of 22 percentage points over conventional methods while retaining at least 95% of diagnostic utility.

What Problem Does This Paper Address?

Electrocardiogram (ECG) waveforms are rich data sources. While they are primarily used to diagnose heart conditions, they also inadvertently leak sensitive demographic information like age and sex. This creates a privacy risk in modern "cloud-service" deployment models: if a proprietary diagnostic AI lives in the cloud, a consumer device must transmit the raw (or processed) waveform to get a diagnosis.

Existing methods like adding noise or downsampling often destroy the very features needed for accurate diagnosis. The authors tackle the challenge of creating a "protector" that can strip away demographic cues while preserving the specific signal features required by a pre-trained, frozen diagnostic classifier.

Methods

The authors propose TIBECG, an architectural temporal information bottleneck. Unlike standard Variational Autoencoders (VAEs) that might bottleneck the entire representation, TIBECG specifically targets temporal resolution.

The system consists of:

  1. A "Protector" Processor: A CNN-based encoder that processes a single-lead ECG, an adaptive average pooling layer that reduces the sequence to $L$ latent positions (the bottleneck), and a decoder that reconstructs the waveform to its original length.
  2. Utility-Only Training: The protector is trained to minimize the difference between the diagnostic scores of a frozen "teacher" classifier (HMC) on the clean waveform versus the protected waveform. Crucially, no age or sex labels are used during training; the model learns to preserve what the classifier needs, not explicitly what the adversary wants.
  3. Empirical Auditing: To measure privacy, the authors separately train "auditor" models (adversaries) to predict age and sex from the released waveforms.

The TIBECG architecture (a) and the two-stage training/auditing process (b). The training and auditing supervision flow.

Datasets

The study uses the CPSC2018 public training set, which contains 6,877 clinical twelve-lead ECGs. After filtering for records with complete demographic data and sufficient length, 6,386 records were used. The data includes various conditions such as atrial fibrillation, bundle branch blocks, and ST-segment changes.

Key Insights and Ideas

  • Lead-Specific Privacy: Privacy is not uniform across the heart. Leads V2, V3, and V4 (the anterior leads) naturally leak significantly more sex information than others. This suggests that privacy-preserving measures must be evaluated on a lead-by-lead basis.
  • Utility Preservation over Privacy Maximization: Compared to direct downsampling, TIBECG is "gentler." While downsampling can suppress demographics more aggressively, it does so at a catastrophic cost to diagnostic accuracy. TIBECG finds a better "Pareto frontier" for utility.
  • KL Divergence is a Poor Proxy: In many information bottleneck studies, Kullback-Leibler (KL) divergence is used to estimate information leakage. This paper finds that at a fixed bottleneck length, lower KL divergence does not reliably correlate with less sex leakage, highlighting the need for direct empirical auditing.
  • Architectural vs. Conventional Transforms: TIBECG consistently outperformed eight conventional transformations (like Gaussian/Laplace noise, quantization, and block permutation) in maintaining the utility-privacy balance.

Key Results

The authors evaluated 21 different bottleneck lengths ($L$) across all 12 clinical leads.

  • Utility Retention: For every lead, there was a TIBECG setting that retained at least 95% of the clean macro-F1 score.
  • Privacy Gain: At this 95% utility floor, TIBECG provided a median 22.08 percentage point advantage in privacy gain over the best eligible conventional transformation.
  • Lead V1 and V6: These leads were particularly robust, maintaining high diagnostic utility even at very short bottleneck lengths.
Fig. 2. Native-scale results by lead. Each point is a median of ten test-fold scores. The gray segment spans the eight conventional medians on that lead. Arrows indicate the preferred direction. Clean and output-only coincide in diagnostic macro-F1. Table S2 gives exact values for each method and each tested length.

Limitations and Future Work

  • Clinical vs. Consumer Data: The study used high-quality hospital recordings. Real-world consumer devices face challenges like motion artifacts, baseline drift, and poor electrode contact, which were not modeled here.
  • Human Interpretability: The "protected" waveforms were not evaluated by cardiologists. It is unknown if a human could still accurately interpret a TIBECG-processed trace or if the process introduces "unreadable" artifacts.
  • Retrospective Selection: The optimal bottleneck length $L^*$ was selected retrospectively. Future work is needed to determine if these selections generalize to entirely new cohorts or different hardware.
  • No Formal Guarantee: The authors emphasize that these are empirical audits, not formal differential privacy guarantees. A stronger adversary or repeated releases might still reveal sensitive data.

Summary updated

Abstract

Transmitting an electrocardiogram (ECG) waveform can expose demographic information beyond an intended diagnostic task. We evaluate TIBECG, an architectural temporal information bottleneck, on individual leads extracted from hospital twelve-lead recordings. A 2,302,337-parameter processor varies its latent length while retaining the same nine-score consistency objective, without age or sex labels in protector fitting. Across 6,386 recordings, ten folds, 12 leads and 21 lengths, retrospective selections retained at least 95% of each lead's clean macro-F1. Their weaker age/sex interval gain exceeded the best eligible conventional transformation on all leads, by a median 22.08 percentage points. The implemented sampling-rate comparison shows that direct downsampling usually suppresses demographic inference more strongly but at a larger loss of diagnostic utility, whereas TIBECG favors utility preservation. Selection stability also varies by lead. Lower latent Kullback-Leibler (KL) divergence does not reliably rank residual sex leakage at a fixed length. These findings support lead-specific evaluation of waveform processing, not validated consumer deployment or expert-readable reconstruction. When the waveform-input diagnostic classifier is hosted as part of a proprietary cloud service rather than deployed on the device, waveform transmission is required to obtain its scores.
BibTeX
@misc{chen2026tibecg,
  author = {Tsai-Min Chen and Jun Seita},
  title = {{TIBECG: Temporal Information Bottleneck for Lead-Specific Demographic Privacy in ECG Waveform Release}},
  year = {2026},
  eprint = {2611.21694},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.21694},
}
Figure