Cross-corpus consensus feature selection for language-invariant Parkinson’s disease detection: a dual-stream acoustic and spectrogram-attention framework

Speech-based biomarkers of Parkinson’s disease (PD) routinely achieve high within-corpus accuracy yet degrade sharply when transferred to an unseen dataset, language, or recording protocol. We argue that this fragility is primarily a feature-stability problem rather than a modelling problem, and we propose a dual-stream cross-corpus consensus feature-selection framework that explicitly enforces inter-corpus agreement. The first stream operates on 156 interpretable handcrafted acoustic descriptors: within each training corpus a heterogeneous set of classifiers performs recursive feature elimination, features are retained by intra-dataset model agreement, and a strict cross-dataset intersection isolates a compact set of cross-corpus stable candidate biomarkers, a subset of which additionally exhibit consistent effect-size direction across derivation languages. The second stream operates on log-Mel spectrogram images, where the spatial attention of a CBAM-augmented ResNet, the attention rollout of a Vision Transformer, and Grad-CAM saliency are combined into a consensus saliency mask whose agreed-upon salient regions yield pooled, file-level image features that are likewise stabilised across corpora. The two consensus streams are fused and classified under monolingual, multilingual, and cross-lingual paradigms. Crucially, three corpora (English, Italian, Bengali) are used only for feature derivation, while three disjoint corpora (Spanish, English, Spanish) are withheld for external validation, so that feature-level invariance is measured independently of model-specific adaptation. The framework advances the transition from language-specific PD models toward scalable, language-agnostic, and interpretable speech-based screening. Cross-corpus stability, per-feature effect-size consistency, and leave-one-language-out validation are reported as the evidentiary basis for these claims, and we explicitly quantify the corpora and margins by which dataset-specific optimisation still outperforms the consensus representation.