Patient-independent validation · AUC 0.79
Adjust the acoustic measures below to see how a logistic regression model — trained on sustained vowel recordings and validated so no patient's voice ever appears in both training and test — scores the likelihood of Parkinson's-associated vocal impairment.
Record your voice
Click Start, then sustain the vowel “ahh” at a comfortable pitch and loudness for about 5 seconds. We'll estimate as many of the sliders below as we reliably can from the recording and refresh the gauge automatically.
About these estimates. Everything in this panel is computed client-side in your browser from an uncalibrated microphone using basic pitch-detection and amplitude-analysis code — it is not the clinical MDVP acoustic analysis software the underlying model was actually trained on. Populated sliders (especially those marked approx.) are rough, uncalibrated estimates for exploring the model, not a measurement to draw any real conclusion from.
Frequency instability (nonlinear)
Pitch range (Hz)
Amplitude & noise
0%
estimated probability of PD-associated vocal impairment
Feature contributions
How to read this. The dial shows the model's output probability. The lower tick marks the 0.5 screening threshold (89% sensitivity, 35% specificity in patient-independent testing); the upper tick marks the 0.91 confirmatory threshold (50% sensitivity, 94% specificity). Feature contribution bars show how far each measure pushes the score toward PD (right, red) or away from it (left, teal), based on the standardized value relative to the full dataset.
This is a research and educational demonstration, not a diagnostic tool. The underlying model is an 8-feature logistic regression fit to the public Oxford Parkinson's Disease Detection dataset (195 recordings, 32 individuals), evaluated with patient-independent cross-validation (AUC = 0.79). It should not be used to assess any real individual.
External validation. This exact model's 8 features don't have clean equivalents in other public voice datasets, so it can't be directly re-scored elsewhere. As a proxy for external validity, a fresh model trained with the same patient-independent methodology on an independent, larger dataset (756 recordings, 252 patients — Sakar et al., 2018) reaches a comparable AUC (≈0.83), suggesting the acoustic approach generalizes to a new population even though this specific trained model could not be ported feature-for-feature.