🧬 biology

Explainable Machine-Learning Analysis of Carotid Doppler Waveform Biomarkers for Chronological-Age Estimation

This study demonstrates that carotid Doppler waveforms contain explainable age-related information, particularly through the Velocity Reflection Index, while highlighting that combining multiple feature importance methods with ablation analysis is essential to avoid overstating the unique contribution of highly correlated biomarkers in machine learning models.

Original authors: Azran Azhim

Published 2026-11-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Azran Azhim

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Our bodies are constantly sending out signals, a quiet stream of data that tells the story of how we are aging. For decades, doctors have listened to the heart and measured blood pressure to get a sense of a person's health, but modern technology now allows us to capture these signals with incredible precision. Among the most revealing of these signals is the flow of blood through the arteries. As we grow older, our blood vessels naturally stiffen, much like an old garden hose that loses its flexibility over time. This change alters the way blood moves, creating a unique pattern in the speed and pressure of the flow. The question scientists have long asked is whether we can read these patterns to determine a person's age, and more importantly, whether we can understand exactly which parts of the signal tell us that story.

A researcher at the National Institute of Technology Kitakyushu College in Japan has taken a fresh look at this problem using a technique called carotid Doppler ultrasound. This is a portable device that uses sound waves to measure how fast blood is moving in the neck artery. The goal was not just to build a computer program that guesses a person's age, but to create a clear, step-by-step method for understanding why the computer makes those guesses. In a study involving 227 healthy adults ranging from 20 to 89 years old, the researcher recorded these blood flow waves and fed them into a machine-learning model. The aim was to see if the computer could learn the relationship between the shape of the wave and the person's actual age, and then to explain which specific parts of the wave were doing the heavy lifting.

The results showed that the computer could indeed estimate a person's age with a reasonable degree of accuracy. On average, the model's guess was off by about 8.3 years. While this might sound like a large margin of error, it is a significant achievement for a system that relies solely on the shape of a blood flow wave. The study identified a specific feature of the wave, called the Velocity Reflection Index, as the most important clue. This index measures how much the blood flow bounces back or reflects within the artery, a phenomenon that changes noticeably as arteries stiffen with age. When the researchers looked at the data, this single feature appeared to be the star of the show, accounting for more than half of the computer's decision-making power according to standard testing methods.

However, the true breakthrough of this paper lies not in the age estimate itself, but in how the researcher tested the reliability of that estimate. In the world of artificial intelligence, it is common for a computer to overvalue a single piece of information if that information is closely related to other pieces of data. To ensure the findings were solid, the researcher performed a rigorous check known as an ablation study. This involved removing the top-ranked feature from the computer's training and seeing how much the accuracy dropped. Surprisingly, when the most important feature was taken away, the computer's performance barely changed. The error rate increased by only a tiny fraction.

This discovery revealed a crucial nuance: the feature that looked like the most important single clue was actually sharing its information with other parts of the wave. Because the blood flow patterns are so interconnected, the computer could easily find the answer using other landmarks in the wave if the primary clue was missing. The study demonstrates that while a single number might look like the most powerful predictor, it does not necessarily hold the entire secret on its own. The researchers concluded that relying on just one method to judge importance can be misleading. Instead, a combination of different testing methods is required to get a true picture of what the data is saying.

The study also highlighted a limitation in how the data was collected. The group of people with complete measurements tended to be significantly younger than those with incomplete records, which means the results are most reliable for younger and middle-aged adults. The model struggled more with the oldest participants, where the data was sparsest. Furthermore, the researchers were careful to clarify that this is an estimate of chronological age based on blood flow, not a definitive medical diagnosis of vascular health. They did not have access to other standard measures of artery health to confirm if the blood flow patterns matched the actual condition of the vessels.

Ultimately, this work offers a clear roadmap for how to use artificial intelligence in medical research without getting lost in the complexity. It shows that while machines can find patterns in biological signals that humans might miss, we must be careful not to assume that the most obvious pattern is the only one that matters. By combining different ways of analyzing the data, scientists can build a more honest and trustworthy understanding of how our bodies change over time. The study confirms that the carotid artery wave carries a wealth of information about aging, but it also teaches us that the story is told by the whole wave, not just by its most prominent peak.

Technical Summary: Explainable Machine-Learning Analysis of Carotid Doppler Waveform Biomarkers for Chronological-Age Estimation

Problem Statement
While physiological sensing and machine learning (ML) are increasingly combined for health assessment, a persistent challenge remains: many systems prioritize predictive accuracy over physiologically grounded explanation. This opacity limits clinical trust and adoption. Furthermore, standard ML interpretability tools, such as Random Forest (RF) feature importance scores, are known to be inflated for predictors that are highly correlated with other predictors in the model. This paper addresses the need for a framework that links signal processing, explainable ML, and biomarker interpretation, while empirically testing whether standard importance metrics accurately reflect the unique contribution of correlated features in a physiological context.

Methodology
The study employs a secondary cross-sectional analysis of a dataset comprising carotid Doppler waveforms from 227 healthy adults (aged 20–89 years). The author proposes and apply a five-stage Explainable Biomarker-Discovery Framework:

Signal Acquisition: Portable Doppler ultrasound (2 MHz) recorded blood-flow-velocity waveforms from the left common carotid artery.
Feature Extraction: Fast Fourier Transform (FFT) and envelope detection generated beat-to-beat waveforms. Eight features were extracted: primary systolic peak ( $S_1$ ), secondary systolic peak ( $S_2$ ), incisura velocity ( $I$ ), peak diastolic velocity ( $D$ ), end-diastolic velocity ( $d_{min}$ ), and three derived indices: Resistance Index (RI), Velocity Reflection Index (VRI), and Vascular Elasticity Index (VEI).
Explainable Analysis: A Random Forest regression model (500 trees) was trained to estimate chronological age. Feature importance was assessed using three distinct methods to test for consistency:
- Gini importance (mean decrease in impurity).
- In-sample permutation importance.
- Held-out permutation importance (evaluated only on validation folds).
Ablation Analysis: To quantify the unique contribution of the top-ranked feature, the model was retrained with the top feature (VRI) removed and compared against subsets containing only raw landmarks ( $S_1, S_2$ ) or the index alone.
Biomarker Interpretation: Results were interpreted with physiological context, distinguishing between waveform-derived indices and direct physiological measurements.

Key Results

Correlation: The Velocity Reflection Index (VRI) showed the strongest Pearson correlation with chronological age ( $r = 0.739, p < 0.001$ ). Other features, including $S_1$ and VEI, also showed significant negative correlations.
Predictive Performance: The Random Forest model achieved a Mean Absolute Error (MAE) of 8.27 ± 1.40 years and an $R^2$ of 0.609 ± 0.140 via repeated 10-fold cross-validation.
Feature Importance Discrepancy:
- Gini Importance: VRI dominated, accounting for 50.4% of total importance.
- Permutation Importance: VRI accounted for 61–79% of importance depending on the method (held-out vs. in-sample).
- Ablation Reality: Despite these high importance scores, removing VRI and retraining the model reduced $R^2$ by only 0.043 (from 0.609 to 0.566). This indicates that the information VRI provides is largely recoverable from other correlated features, specifically $S_1$ ( $r = -0.64$ ).
Bias Check: A missingness analysis revealed that participants with complete body-composition data were significantly younger (mean 31.4 years) than those without (mean 45.7 years), highlighting a potential sampling bias in subsets of the cohort.

Key Contributions

Conceptual Framework: The paper introduces the "Explainable Biomarker-Discovery Framework," which explicitly integrates signal processing, multi-method explainable ML, and biomarker-level interpretation into a traceable pipeline.
Worked Illustration: It provides a concrete application of this framework using carotid Doppler waveforms to estimate age, evaluating three established hemodynamic indices (RI, VRI, VEI).
Stress-Testing Explainability: The study demonstrates that relying on a single importance metric (like Gini) can lead to overstated claims about a feature's indispensability. By combining impurity-based scores, permutation importance, and ablation analysis, the author shows that correlated features can mask each other's contributions, leading to divergent interpretations of "importance."
Qualified Claims: The paper explicitly distinguishes between a "Doppler-based chronological-age estimate" and a validated "vascular-age biomarker," noting that the former does not yet meet the criteria for the latter (e.g., independence from chronological age or correlation with vascular outcomes).

Significance and Claims
The paper's primary significance is methodological rather than purely clinical. It argues that in explainable AI pipelines for physiology, a single importance metric is insufficient for determining a biomarker's unique value, particularly when features are correlated. The study concludes that while carotid Doppler waveforms contain substantial age-related information, the apparent dominance of VRI in importance scores is an artifact of correlation with other landmarks (like $S_1$ ).

The author maintains a modest stance: the resulting model is a proof-of-concept for a framework that connects signal processing to interpretable biomarkers. They explicitly state that the current output is a chronological-age estimate derived from Doppler signals, not a validated vascular-age biomarker, and that further validation against independent vascular measurements (e.g., pulse-wave velocity) and clinical outcomes is required before such claims can be made. The study serves as a "stress test" for explainability methods, demonstrating that ablation analysis is necessary to provide a defensible account of feature contributions.

Technical Summary: Explainable Machine-Learning Analysis of Carotid Doppler Waveform Biomarkers for Chronological-Age Estimation

More like this