Feature Selection for Nonstationary Data: Application to Human Recognition Using Medical Biometrics

Majid Komeili, Wael Louis, Narges Armanfard, Dimitrios HatzinakosIEEE Trans. Cybernetics 2017 · DOI

Introduction

Traditional biometric systems, such as fingerprint and facial recognition, are increasingly susceptible to sophisticated spoofing attacks using silicone molds or high-resolution imagery. In response, researchers have turned to medical biometrics—physiological signals like the Electrocardiogram (ECG) and Transient Evoked Otoacoustic Emission (TEOAE). These signals are inherently resistant to spoofing because they require "liveness" and are difficult to capture surreptitiously. However, medical biometrics face a significant hurdle: non-stationarity. Physiological signals fluctuate based on time, emotional state, and physical posture, making consistent recognition difficult over long periods.

The Challenge of Non-Stationary Biometrics

In a real-world deployment, a biometric system typically operates under a "single-session enrollment" model. A user provides a brief sample once, and the system must recognize them days or months later. Because the system only sees a snapshot of the user's signal, it cannot account for how that signal might change over time. Standard machine learning classifiers often overfit to the specific conditions of the enrollment session, leading to high error rates when the user returns in a different state.

Visualization of feature non-stationarity. Left: Features f1 and f2 show significant drift between sessions. Right: Features f3 and f4 remain stable across sessions, making them ideal for biometric recognition.

As shown in the figure above, certain features are highly sensitive to temporal changes (non-stationary), causing class clusters to shift significantly between sessions. Identifying the subset of features that remain stable (stationary) is the key to robust recognition.

Multisession Feature Selection (MSFS)

To address this, the authors propose a novel Multisession Feature Selection (MSFS) algorithm. The core innovation is the use of an auxiliary dataset—a separate pool of data from general subjects recorded across multiple sessions. While the system may only have one session for a new user, it uses the auxiliary data to learn which physiological features are generally stable across time and postures for everyone.

The Core Algorithm

MSFS evaluates thousands of potential features using a modified margin criterion. In standard learning, a "nearest hit" is the closest sample from the same class. MSFS redefines this: a "nearest hit" must be a sample from the same class but from a different session. By minimizing the distance to these cross-session hits while maximizing the distance to "nearest misses" (samples from different classes), the algorithm effectively filters out features that fluctuate wildly over time.

Scaling with Random Projection (MSFS-RP)

Evaluating over 7,000 features for nearest neighbors is computationally expensive. To make the system practical, the authors introduced MSFS-RP, which utilizes Random Projection. By projecting high-dimensional data into a lower-dimensional space, the algorithm can rapidly identify candidate nearest neighbors, speeding up the selection process by approximately 4x with negligible loss in accuracy.

Experimental Validation

The framework was tested on three distinct datasets, focusing on realistic, non-clinical collection environments:

  • BioSec ECG Database: 82 subjects recorded over six months using dry electrodes on the fingertips (see Figure 1). This method is more practical for consumer electronics than clinical chest gels but introduces significantly more noise.
  • BioSec TEOAE Database: 54 subjects recorded in a standard office environment, rather than a soundproof booth, to simulate real-world acoustic interference.
ECG data collection using dry electrodes on the fingertips, facilitating practical, non-clinical biometric enrollment.

Results and Impact

The MSFS approach demonstrated a dramatic improvement over existing state-of-the-art feature selection methods (such as Simba and LLFS):

  • ECG Stability: In across-session testing, MSFS achieved an Equal Error Rate (EER) of 6.9%, compared to over 14% for traditional methods. When testing across different body postures (sitting vs. standing), the EER dropped to 3.7%.
  • TEOAE Accuracy: The improvement in TEOAE was even more pronounced, with MSFS reaching a 3.3% EER, while prior methods struggled with error rates exceeding 22%.
  • Efficiency: The MSFS-RP variant reduced processing time from over an hour to roughly 13 minutes for the ECG dataset, proving the feasibility of the approach for large-scale biometric databases.

Conclusion

By leveraging auxiliary data to identify stationary physiological traits, MSFS overcomes the primary barrier to using medical biometrics in consumer applications. While the study focused on ECG and TEOAE, the underlying mathematical framework is applicable to any non-stationary time-series data, offering a path forward for more secure and reliable human recognition systems.

Summary updated

Abstract

Electrocardiogram (ECG) and transient evoked otoacoustic emission (TEOAE) are among the physiological signals that have attracted significant interest in biometric community due to their inherent robustness to replay and falsification attacks. However, they are time-dependent signals and this makes them hard to deal with in across-session human recognition scenario where only one session is available for enrollment. This paper presents a novel feature selection method to address this issue. It is based on an auxiliary dataset with multiple sessions where it selects a subset of features that are more persistent across different sessions. It uses local information in terms of sample margins while enforcing an across-session measure. This makes it a perfect fit for aforementioned biometric recognition problem. Comprehensive experiments on ECG and TEOAE variability due to time lapse and body posture are done. Performance of the proposed method is compared against seven state-of-the-art feature selection algorithms as well as another six approaches in the area of ECG and TEOAE biometric recognition. Experimental results demonstrate that the proposed method performs noticeably better than other algorithms.
BibTeX
@article{komeili2017feature,
  author = {Majid Komeili and Wael Louis and Narges Armanfard and Dimitrios Hatzinakos},
  title = {{Feature Selection for Nonstationary Data: Application to Human Recognition Using Medical Biometrics}},
  journal = {IEEE Transactions on Cybernetics},
  year = {2017},
  doi = {10.1109/tcyb.2017.2702059},
}
Figure