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

Abstract Background Physiological sensing and machine learning are increasingly combined for health assessment, but many systems emphasize predictive accuracy over physiologically grounded explanation. Objective To describe a small explainable biomarker-discovery framework linking signal processing, explainable machine learning, and physiological interpretation, illustrated and stress-tested with carotid Doppler waveforms. Methods In 227 healthy adults (20–89 years) with portable carotid Doppler recordings, eight waveform features were related to chronological age using Pearson correlation and Random Forest (RF) regression with repeated 10-fold cross-validation. Feature importance was assessed by Gini and permutation methods (in-sample and held-out), and ablation quantified the unique contribution of the top-ranked feature, the Velocity Reflection Index (VRI). Results VRI correlated most strongly with age (r = 0.739, p < 0.001). The RF model estimated age with MAE 8.27 ± 1.40 years and R² = 0.609 ± 0.140. VRI dominated both Gini importance (50.4%) and held-out permutation importance (an even larger share), yet removing VRI and retraining cost only 0.043 R² (0.609 to 0.566), since VRI correlates strongly with other retained landmarks (r = -0.64). A missingness-bias check showed complete-body-composition participants were significantly younger than others (31.4 vs. 45.7 years, p < 0.001). Conclusion Carotid Doppler waveforms carry substantial, explainably identifiable, age-related information, but a single importance metric can overstate a correlated feature's unique contribution; combining importance methods with ablation gives a more defensible account.