⚡ electrical engineering

Interpretable machine learning for identifying individual-specific cardiogram signatures

This study demonstrates that interpretable machine learning models can identify robust, individual-specific cardiogram signatures primarily through stable ECG QRS features, effectively distinguishing personal identity from transient emotional states like anger while optimizing feature sets for high accuracy and computational efficiency.

Original authors: Ilija Tanasković, Ljiljana B. Lazarević, Goran Knežević, Nikola Milosavljević, Olga Dubljević, Bojana Bjegojević, Nadica Miljković

Published 2026-07-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ilija Tanasković, Ljiljana B. Lazarević, Goran Knežević, Nikola Milosavljević, Olga Dubljević, Bojana Bjegojević, Nadica Miljković

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 your body is like a unique musical instrument that never stops playing. Even when you are sitting still, your heart is drumming a rhythm, and your chest is expanding and contracting with every beat. Scientists have long known that this rhythm isn't just a timer; it's a signature. Just as no two people have the exact same fingerprints or iris patterns, no two hearts beat with the exact same electrical and mechanical "fingerprint." This is the world of biometric identification, where we try to prove who you are by listening to your body's song. But here is the tricky part: your song changes when you feel different emotions. If you get angry, scared, or excited, your heart rate speeds up, and the shape of the beat changes. This creates a puzzle for computers: How can a machine tell if a heart signal belongs to "You" or if it's just "You" having a bad day? Researchers need to figure out which parts of the heartbeat are your permanent ID card and which parts are just temporary mood swings.

This paper dives into that puzzle by treating the heartbeat like a complex code made of 29 different clues. The researchers used a smart computer program called a "Random Forest" (think of it as a team of 100 detectives, each looking at the clues from a slightly different angle) to see which clues are the best at identifying a person. They looked at two types of heart signals: the electrical spark (ECG) and the mechanical push of blood (ICG). They tested these signals while people were calm and while they were angry, trying to find the "super clues" that stay the same no matter how you feel.

The team discovered that the most reliable clues are hidden in a specific, sharp spike in the heartbeat called the QRS complex. It's like the loudest, most distinct note in a song. The computer found that the height, slope, and shape of this spike are incredibly stable. Even when the person got angry, these specific features barely changed. The study showed that if you focus only on these 12 best clues (mostly from the electrical signal), the computer can still identify the person with 99% accuracy—almost as good as using all 29 clues. This is a big deal because it means we can build faster, simpler security systems that don't get confused by your mood.

However, the paper also ruled out some ideas. It found that many of the other clues, like the time between beats or the shape of the later parts of the wave, are very sensitive to anger. When the researchers tried to use only these mood-sensitive clues to identify people, the system failed miserably, dropping in accuracy to around 70%. This proves that relying on the parts of the heartbeat that change with emotion is a bad idea for security. The study also showed that many of the clues are actually "best friends" that tell the same story; if you know one, you mostly know the other. The computer naturally figured this out, spreading its attention across these similar clues, but the researchers confirmed that you can cut the list of clues in half without losing any power.

In the end, the paper suggests that your heart's electrical "spike" is your true, unchangeable ID, while the rest of the signal is a mix of your identity and your current feelings. By focusing on the stable parts and ignoring the emotional noise, we can create better ways to prove who we are, even when we are stressed or angry. The researchers are careful to say this works for healthy young adults in their study, and while the results are very strong for this group, more work is needed to see if it holds up for everyone else. But for now, we have a clearer map of where the "you" lives inside your heartbeat.

Technical Summary: Interpretable Machine Learning for Identifying Individual-Specific Cardiogram Signatures

Problem Statement
While electrocardiogram (ECG) and impedance cardiogram (ICG) signals are established as promising modalities for biometric identification due to their liveness detection capabilities, existing research has predominantly focused on predictive performance rather than interpretability. A critical gap remains in understanding which specific physiological features encode individual identity and how these features behave under emotional variability. Emotional states, particularly high-arousal conditions like anger, can alter cardiac morphology, potentially degrading authentication accuracy. Furthermore, the field lacks clarity on distinguishing stable, identity-specific physiological carriers from transient, affective signals. This study addresses the need to transition from "black-box" models to transparent systems that identify the specific discriminative contributions of ECG and ICG features and their robustness to emotional states.

Methodology
The study utilizes a publicly available dataset of simultaneously recorded ECG and ICG signals from 202 healthy young adults, recorded during baseline and anger conditions. The methodology employs a comprehensive, multi-path analytical pipeline:

Feature Extraction: 29 features were extracted across four domains: temporal (intervals), amplitude, slope, and morphological. These include fiducial points from the QRS complex, T-wave, and the BCX complex of the ICG.
Baseline Modeling: A Random Forest (RF) classifier was trained on all 29 features to establish a performance benchmark (99.17% accuracy).
Feature Importance Analysis: Three distinct interpretability methods were applied to rank features: Gini importance, permutation feature importance, and Shapley additive explanations (SHAP). The top 10 features from each method were intersected to form a consensus subset.
Feature Selection: Two data-driven selection techniques—Recursive Feature Elimination with Cross-Validation (RFECV) and a Genetic Algorithm (GA)—were employed to identify optimal feature subsets. The intersection of these subsets defined a compact feature set.
Correlation and Cluster Analysis: Pearson and Spearman correlation coefficients were calculated to identify highly correlated feature clusters ( $|r| > 0.7$ ). The impact of multicollinearity on feature importance was assessed by shuffling correlated pairs and observing changes in importance scores.
Emotional Sensitivity Analysis: Statistical testing compared features between baseline and anger segments to identify emotion-sensitive (significant) versus emotion-insensitive (non-significant) features. RF models were retrained on three specific subsets: all features, only emotion-insensitive features, and only emotion-sensitive features to evaluate cross-segment generalization.

Key Results

Feature Importance: ECG-derived features, specifically those describing the QRS complex (amplitude, slope, and morphology), consistently ranked as the most important across all interpretability methods. ICG-derived BCX features provided complementary information but showed lower stability in rankings.
Multicollinearity: The analysis revealed substantial multicollinearity, with seven distinct clusters of highly correlated features identified. The RF model distributed importance scores across these correlated pairs, confirming that highly correlated features have reduced independent contributions.
Feature Selection Efficiency: The intersection of RFECV and GA selections yielded a compact subset of 12 features. Retraining the RF model on this subset achieved 98.17% accuracy, within 1% of the full 29-feature model (99.17%), demonstrating that dimensionality reduction can significantly improve storage and computational efficiency without sacrificing discriminative power.
Emotional Stability: Statistical analysis identified 14 features that differed significantly between baseline and anger states. Crucially, the most discriminative QRS-derived features (e.g., RQ_amp, RS_amp, QR_slope) remained stable across emotional states. In contrast, features related to the T-wave and certain ICG intervals showed significant variability.
Generalization Performance: Models trained exclusively on emotion-sensitive features performed poorly in cross-segment generalization (66–72% accuracy). Conversely, models trained on emotion-insensitive features maintained high within-segment accuracy but did not significantly improve cross-emotion generalization compared to the full model, suggesting that removing sensitive features alone does not fully mitigate the impact of emotional variability on identification.

Significance and Contributions
The paper claims to provide a blueprint for linking machine learning decisions to physiologically meaningful components in biometric systems. Its primary contributions include:

Identification of Stable Identity Carriers: The study establishes that the QRS complex of the ECG, particularly amplitude and slope descriptors, serves as the primary stable carrier of individual-specific information, remaining robust even under the influence of anger.
Interpretability Framework: By integrating multiple interpretability methods, the research moves beyond predictive accuracy to explain why models work, distinguishing between stable anatomical/electrophysiological traits and transient affective modulations.
Efficient Feature Subsets: The identification of a 12-feature subset that retains near-complete performance offers a practical pathway for developing efficient, real-time biometric authentication systems.
Emotional Robustness Insights: The findings clarify that while emotional states alter specific cardiographic features, the core identity information encoded in the QRS complex remains largely intact. This supports the development of authentication systems that rely on stable physiological invariants rather than transient states.

The study concludes that cardiographic signals contain multiple layers of information: stable identity traits concentrated in the QRS complex, complementary but less stable cues from the T-wave and ICG, and transient affective signals. The proposed framework aids in distinguishing these layers, supporting both more robust physiological authentication and future research into stable person-related characteristics.

Technical Summary: Interpretable Machine Learning for Identifying Individual-Specific Cardiogram Signatures

More like this