Фізично-керована ланцюгова гаусівська регресія процесу для прогнозування прогресії атеросклеротичної бляшки

Predicting atherosclerotic plaque progression from sparse cross-sectional imaging data is critical for cardiovascular risk assessment. Standard regression methods lack physics-based inductive bias, leading to unreliable extrapolation. A Physics-Guided Chained Gaussian Process Regression (PG-CGPR) framework is presented that incorporates Power-Law Damage Accumulation mechanics as a mean function prior with heteroscedastic noise modeling. Cross-sectional observations are transformed into pseudo-longitudinal training pairs using Glagov-tolerant filtering. Validated on cardiovascular imaging data spanning multiple decades, the model maintains predictive accuracy even when extrapolating beyond the training age range, with confidence intervals covering all test observations. The strong physics prior dominates data-driven learning, demonstrating that in small-sample regimes, mechanistic models outperform flexible kernels. The framework enables projection of future vessel geometries for downstream biomechanical analysis.