EXPLAINABLE AI FOR PARKINSON’S DISEASE DETECTION THROUGH SPEECH BIOMARKER FOR CLINICAL CONSTRUCT ALIGNMENT AND CROSS-LINGUAL CONSISTENCY

Abstract

Parkinson’s disease (PD) afflicts over 8.5 million people globally and is projected to exceed 12 million by 2040, rendering it the most rapidly expanding neurological dis order of our time. A central complication in disease management is that the motor signs used for clinical diagnosis become apparent only after 60%-80% of nigrostri atal dopaminergic neurons have undergone irreversible degeneration, leaving a critical window for neuroprotective treatment permanently inaccessible under current prac tice. Because the basal ganglia circuits that coordinate laryngeal and articulatory mus culature are anatomically inseparable from those governing limb control, the acous tic consequences of this neurodegeneration collectively characterised as hypokinetic dysarthria (HD) are detectable in sustained phonation years before diagnosis. Pertur bation measures reflecting aperiodic vocal fold vibration, reduced glottal closure effi ciency, elevated spectral noise energy, and compressed fundamental frequency range can be extracted from seconds of speech with commodity microphones, positioning voice analysis as a pragmatically deployable screening modality. Despite two decades of published research producing an accuracy range of 72%-99.6%, no voice-based PD detector has reached clinical practice. Three inter-related evalu ation confounds explain the implausible breadth of this range. First, leave-one-out cross-validation (LOOCV) on the 195-recording UCI benchmark inflates accuracy by roughly 6.6 percentage points against stratified k-fold evaluation, because near-total training-set overlap across folds lets classifiers exploit corpus-level artefacts rather than generalisable pathophysiological signal. Second, splitting individual recordings rather than participants introduces speaker identity leakage, adding a further 3–8 points. Third, and mostconsequentially, evaluating on an independently collected external cor pus, the only design with direct clinical relevance consistently depresses performance by 5-15 points relative to within-corpus estimates. Two structural gaps persist beyond these methodological issues: no validated bridge links machine learning (ML) feature attribution to the clinical constructs of speech pathology, and the literature is concen trated in five European language communities, leaving fundamental cross-linguistic generalisability unresolved. This thesis addresses these gaps through two coordinated contributions. The first is a critical systematic review extending the 147-study corpus of Ngo et al. (2022) to De cember 2024 via ten-database search with dual-reviewer screening (Cohen’s κ =0.83), identifying approximately 220 works and synthesising findings across five original comparison tables covering dataset characteristics, acoustic feature taxonomies, clas sical ML and deep learning (DL) architectures, self-supervised learning (SSL) mod els, and specialised detection objectives. A three-tier gap analysis ranks ten barriers to clinical translation by urgency, with absent prospective trials, single-corpus evaluation, and unstandardised recording protocols in the highest tier. The second contribution is a multi-corpus experimental study covering the UCI Parkinson’s dataset (195 record ings, 22 Multi-Dimensional Voice Program (MDVP) features), the UPDRS Telemon itoring corpus (5,875 recordings, 42 patients, continuous motor severity labels), and BenSParX (900 conversational Bengali recordings, 71 Praat features plus raw audio), v the first Bengali PD speech corpus. Six classifiers- Random Forest (RF), XGBoost (XGB), SVM with radial basis function kernel (SVM-RBF), logistic regression (LR), K-nearest neighbours (KNN), and a deep multi-layer perceptron (MLP) were evalu ated under 10-fold stratified cross-validation with strict participant-level partitioning, and SHapley Additive exPlanations (SHAP) attribution was applied using architecture matched explainers throughout. SVM-RBF achieved the highest AUC-ROC on BenSParX (0.985±0.014) and KNN led on UCI (0.982±0.017). Severity regression under subject-independent Group KFold produced negative R2 across all regressors, a theoretically instructive finding quantifying precisely how far inter-subject vocal variability exceeds the disease sig nal for unseen patients, while leaving within-patient longitudinal monitoring clinically viable. The Clinical Construct Alignment Table maps each of the 22 UCI acous tic features to a DAB dysarthria dimension and its neurophysiological substrate; all 16 UPDRS corpus features correlated significantly with motor UPDRS severity after Bonferroni correction (p < 0.001), and SHAP importance rankings were concordant with severity correlation rankings, confirming that classifier decisions reflect genuine pathophysiological signal rather than dataset confounds. Cross-lingual SHAP con sistency across 15 acoustically matched English-Bengali feature pairs was weak and non-significant (ensemble ρ = +0.118, p = 0.676), attributed to task and vocabu lary mismatch rather than language-family differences. WavLM-Base-Plus probing on BenSParX audio identified the second encoder layer as most PD-discriminative (AUC = 0.871±0.124), trailing the handcrafted SVM baseline by 11 AUC points at this dataset scale, with lower-layer dominance corroborating the phonation-quality construct hierarchy independently. Collectively, this thesis delivers the first numeri cally grounded evaluation-confound analysis for PD voice detection, the first formally validated ML-to-speech-pathology construct mapping, and the first SHAP consistency and WavLM probing study on a non-European PD speech corpus.