Table 2 Clinical translation and explainability approaches
Study | Clinical Interpretability & Feature Mapping | Implementation Strategy | Local Explainability Features | Global Explainability Features | Clinical Applications & Validation |
|---|---|---|---|---|---|
Heitz et al. (2024)17 | - Linguistically meaningful features - Clinical relevance emphasis - Linguistic features → cognitive markers | - Direct integration with model pipeline - Real-time processing capability | - Individual prediction explanations - Case-specific feature analysis | - Overall feature importance - Model behavior patterns | - Public dataset validation - Clinical correlation - Individual assessment - Population-level screening |
Ilias & Askounis (2022)18 | - Natural language explanations - Clinical decision support - Language patterns → cognitive status | - Interactive visualization - Clinical workflow integration | - Instance-level explanations - Individual confidence scores | - Feature importance hierarchy - Model interpretation | - Validation against existing research - Clinical testing - Patient-specific diagnosis - General screening |
de Arriba-Pérez et al. (2024)27 | - Domain-adapted explanations - Patient-friendly interpretations - High-level reasoning features | - Web application interface - Real-time analysis | - Individual session analysis - Personal feature importance | - Population-level patterns - Feature relationships | - MMSE score correlation - Clinical validation - Individual monitoring - Group analysis |
Ambrosini et al. (2024)28 | - Multi-language support - Clinical workflow integration - Acoustic-cognitive mapping | - Mobile app integration - Privacy preservation | - Subject-specific analysis - Individual language patterns | - Cross-language patterns - Population trends | - Multi-center validation - Cross-cultural testing - Individual assessment - Population screening |
Tang et al. (2023)19 | - Feature-based explanation - Clinical correlation - Linguistic markers → AD indicators | - Clinical decision support - Real-time analysis | - Case-based explanations - Individual feature impact | - Model-wide patterns - Feature importance | - Clinical dataset validation - Feature verification - Patient diagnosis - General screening |
Chandler et al. (2023)30 | - Telephone-based screening - Clinical accessibility - Language features → cognitive status | - Remote assessment tool - Clinical integration | - Individual call analysis - Personal patterns | - Population trends - Feature relationships | - MMSE correlation - TICS-M validation - Individual screening - Population monitoring |
Iqbal et al. (2024)24 | - Binary classification focus - Clinical screening tool - Clinical feature mapping - POS patterns → cognitive decline | - Clinical screening tool - Feature-based analysis | - Case-specific analysis - Individual thresholds | - Global feature patterns - Model behavior | - ADReSS dataset validation - Clinical testing - Individual diagnosis - General screening |
Han et al. (2025)25 | - Key speech markers identified - MCI-specific patterns - Counterfactual insights | - LLM-based generation - Real-time capable | - Patient-specific counterfactuals - Individual marker analysis | - Population-level patterns - Feature directionality | - Framework validation - Marker verification - Early MCI detection |
Oiza-Zapata & Gallardo-Antolín (2025)20 | - Acoustic biomarkers - Clinical utility focus - Smart city healthcare | - Efficient pipeline - Automated screening | - Patient-level analysis - Personalized features | - Population patterns - Feature rankings | - CUI assessment - Healthcare integration - Screening tool |
Jang et al. (2021)29 | - Multi-modal integration - Clinical feature interpretation - Window features → AD markers | - Testing platform - Multi-sensor setup | - Individual task performance - Personal biomarkers | - Cross-task patterns - Feature correlations | - Expert diagnosis validation - Novel task evaluation - >90% user satisfaction |
Li et al. (2025)26 | - Topic evolution analysis - Cross-modal consistency - Macrostructural markers | - SHAP + attention - Temporal modeling | - Individual narrative patterns - Session-specific features | - Topic variability metrics - Population-level insights | - Two-dataset validation - Severity correlation - Monitoring potential |
Lima et al. (2025)21 | - Risk stratification (3-tier) - Clinical markers - Pronoun/disfluency patterns | - Automated pipeline - Conversational AI ready | - Patient risk profiles - Individual explanations | - Feature importance - Population patterns | - External validation - Real-world pilot (n = 22) - Demographic parity |
Ntampakis et al. (2025)22 | - Literature-grounded explanations - Medical professional design - Evidence-based markers | - RAG architecture - Dual-component system | - Patient-specific explanations - Clinical evidence links | - Model behavior analysis - Feature relationships | - Medical professional evaluation - Low misinterpretation risk (2.38/5) - Clinical utility: 3.70/5 |
