Parkinson's Disease Detection from Voice using Machine Learning

Up to 90% of people with Parkinson's disease develop vocal impairment — reduced loudness, breathiness, pitch instability — often years before the classic motor symptoms prompt a diagnosis. Those changes are measurable: jitter (pitch instability), shimmer (loudness instability), harmonics-to-noise ratio and nonlinear measures like PPE capture them numerically from a simple sustained "ahh". The UCI Parkinson's dataset packages 195 such recordings from 31 subjects with 22 precomputed biomarkers, making it an ideal teaching dataset for biomedical ML — small enough to run anywhere, methodologically rich enough to teach the field's central lesson: validate by subject, not by recording. This project trains an RBF SVM with leave-one-subject-out cross-validation, explains each prediction through its biomarker profile, and delivers a voice-test web demo. Screening-only limits are stated plainly throughout.

1. The 195 UCI recordings' 22 precomputed vocal features are loaded with subject IDs and PD/healthy labels.
2. Features are standardized with a scaler fit on the training fold only.
3. An RBF-kernel SVM (C and gamma tuned by nested cross-validation) is trained; logistic regression and random forest serve as baselines.
4. Validation uses leave-one-subject-out: 31 folds, each holding out all recordings of one subject — no speaker appears in both train and test of any fold.
5. Accuracy, sensitivity, specificity and ROC-AUC are aggregated across folds; feature importance is read from the SVM weights and permutation analysis.
6. In the web demo, a voice profile's 22 biomarkers are scaled identically and the saved SVM returns the risk score.
7. Each biomarker is plotted against its healthy range so the prediction is explained feature by feature.

Tech stack:

• Python 3, scikit-learn (SVM)
• UCI Parkinson's dataset (Little et al., 2007)
• librosa / Parselmouth (feature theory)
• NumPy, Matplotlib, pandas
• Jupyter Notebook (training)
• HTML5 canvas + JavaScript (voice-test demo)

Dataset & model details

• Dataset: Parkinson's Disease dataset (UCI ML Repository, Little et al., 2007) — 195 sustained-vowel phonations from 31 subjects (23 with Parkinson's, 8 healthy), 22 vocal features per recording (jitter, shimmer, HNR, NHR, RPDE, DFA, PPE, spread1/2, D2, MDVP statistics).
• Task: Binary classification; input = 22 standardized vocal biomarkers, output = Parkinson's risk score.
• Model: SVM with RBF kernel (C=10, γ=0.01 tuned by nested CV), StandardScaler preprocessing.
• Metrics: Accuracy ~92% (design target for the built-to-order training run), sensitivity, specificity, ROC-AUC — all under leave-one-subject-out CV. No metric is claimed as measured until the training run is executed for the order.

• [22 vocal biomarkers] Jitter, shimmer, HNR/NHR, RPDE, DFA, PPE, spread measures and MDVP statistics — each explained in plain language in the report.
• [Voice-test web demo] Simulated 5-second "ahh" capture with animated waveform, biomarker extraction display, risk score and per-feature healthy-range comparison.
• [Leave-one-subject-out validation] 31-fold CV by subject — the only honest validation when 195 recordings come from 31 people; random splits would leak speakers.
• [RBF SVM classifier] Tuned C and gamma via nested CV, benchmarked against logistic regression and random forest baselines.
• [Per-feature explanation] Every prediction is explained through its biomarker bars against healthy ranges, so the score is inspectable.
• [Full training notebook] Scaling, model selection, LOSO validation, ROC analysis and calibration in one reproducible notebook.
• [Screening report] The demo generates a structured screening summary with the mandatory not-a-diagnosis disclaimer.

• Complete training & evaluation Jupyter notebook
• Trained SVM model file with scaler and feature pipeline
• Voice-test web demo (capture simulation, biomarkers, risk score)
• LOSO metrics, ROC curves and biomarker-importance plots
• Project report PDF (background, vocal-biomarker theory, methodology, ethics)
• PPT presentation for final review
• Viva Q&A preparation document (SVM kernels, LOSO, jitter/shimmer/HNR, screening limits)

• ~92% is a design target for the training run, stated honestly — the report documents the actual achieved figure.
• 31 subjects is a small cohort; generalization claims are bounded and the report says so.
• SCREENING AID ONLY: vocal biomarkers correlate with Parkinson's in research data but cannot diagnose it — not a medical device, never for clinical decisions.
• The demo simulates capture; the shipped build documents microphone integration and the sustained-vowel protocol.
• Age, colds and other voice disorders also shift these biomarkers — confounders the report discusses.