TimeSynth Framework Improves Health Signal Digital Twin Fidelity

Developing accurate forecasting models for health signal digital twins is crucial, but current evaluation methods often fall short. Standard pointwise metrics, while useful, cannot detect when fundamental properties like oscillations, frequency, phase, and state-transition dynamics of physiological signals are lost during modeling. This blind spot can lead to misranking models, where models with similar pointwise errors can have vastly different phase accuracies, for example. To address this, researchers have introduced TimeSynth, a controlled benchmarking framework designed to ensure temporal fidelity. TimeSynth features a physiologically grounded signal generator that produces signals with known ground-truth dynamics, derived from real EEG, ECG, and PPG data. It also includes diagnostics that precisely quantify amplitude, frequency, phase, and state-transition fidelity. The framework reveals that architectural choices are the dominant determinant of fidelity, with models having localized temporal structures better preserving dynamics than linear or full-sequence attention models. TimeSynth provides a critical preclinical stress test, enabling fidelity-aware model development before deployment with patient data.