Nonlinear Metrics

The study compared recurrence biomarkers with static acoustic summaries, entropy biomarkers, forecastability features, Hurst exponent features, a Lyapunov-like instability proxy, and a determinism proxy. The Hurst exponent performed poorly, indicating that long-memory scaling alone did not characterize depression-related speech changes in the dataset. The determinism proxy was weak, suggesting diagonal recurrence structure was not the primary signal. Combining Lyapunov-like features with recurrence features reportedly did not improve performance beyond recurrence alone. The Lyapunov-like instability proxy was intended to capture local fluctuation magnitude and was not treated as a formal largest Lyapunov exponent.