enkhtogtokh/synthetic_pathology

Togootogtokh, E., & Klasen, C. (2025). Synthetic Pathology Dataset (1.0.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.14974650 Link: https://zenodo.org/records/14974650

Github code repo: https://github.com/enkhtogtokh/voicegrpo

A synthetic dataset was generated to mimic realistic distributions of voice parameters (e.g., pitch, jitter, shimmer, harmonic-to-noise ratio, age, and a continuous disease severity score). The pathological labels were derived based on domain-inspired thresholds, ensuring a challenging classification task.

we assess the thresholds applied to generate synthetic pathology labels, evaluating their alignment with clinical contexts. • Jitter (> 0.05): Jitter measures frequency variation in voice signals. Healthy voices typically exhibit jitter below 1–2%, while the 0.05 (5%) threshold exceeds clinical norms but may detect pronounced pathology, assuming proper scaling. • Shimmer (> 0.08): Shimmer reflects amplitude variation, normally below 3–5% in healthy voices. The 0.08 (8%) threshold is above typical ranges, suitable for severe cases but potentially missing subtle issues. • HNR (< 15): Harmonic-to-Noise Ratio (HNR) indicates harmonic versus noise balance. Healthy voices often exceed 20 dB, while <15 dB aligns with pathological noisiness, making this threshold clinically plausible. • Age (> 70): Age is a risk factor for voice decline, but >70 as a pathology marker is overly simplistic. It may act as a proxy in synthetic data, though not diagnostic in practice. • Disease Severity (> 0.7): This synthetic parameter, likely on a 0–1 scale, uses a 0.7 cutoff to denote severity. While arbitrary, it is reasonable for synthetic data but lacks direct clinical grounding.