Composite Biofidelity: Addressing Metric Degeneracy in Biomechanical Model Validation and Machine Learning Loss...

Area of Science:

• Computational Biomechanics
• Signal Processing
• Machine Learning

Background:

• Single spectral similarity metrics (e.g., RMSE) can be misleading in computational biomechanics.
• This limitation impacts model validation and the design of machine learning loss functions.
• Objective spectral biofidelity assessment is critical for accurate simulation and analysis.

Purpose of the Study:

• To develop and evaluate a multi-metric framework for objective spectral biofidelity.
• To determine if this framework better captures meaningful disagreements in complex frequency-domain responses.
• To assess the framework's performance in validating computational models and informing machine learning.

Main Methods:

• Evaluated 12 complementary spectral similarity metrics, including CORA and ISO/TS 18571.
• Used controlled spectral perturbations (resonance shifts, spikes, tilts) to mimic real-world deviations.
• Applied the framework to a finite-element middle-ear model, assessing convergence and noise robustness.

Main Results:

• No single metric reliably performed across all distortion types.
• Shape-based metrics captured morphology but missed scaling; MaxError was key for narrowband anomalies.
• CORA and ISO/TS 18571 did not consistently outperform simpler metrics.
• Rank aggregation (Borda count) provided a robust consensus, identifying data saturation and noise thresholds.

Conclusions:

• Spectral biofidelity cannot be accurately assessed using a single metric.
• A multi-metric consensus offers a clearer, more physically meaningful basis for comparing experimental and simulated spectra.
• This framework provides a robust foundation for data-fidelity terms in physics-informed and simulation-based machine learning.