Behind the [MASK]: Disentangling Representation and Faithfulness in DAPF-Based Dementia Detection
DAPF (Domain-Adapted models via Prompt-based Fine-tuning) achieves the best dementia detection performance among tested models, with accuracy=0.83 and macro-F1=0.83 Diagnosis information is most recoverable from the [MASK] token representation in DAPF, confirming its representational strength Token-level explanation faithfulness does not match representational performance; attributions primarily reflect language task vocabulary, discourse markers, and transcription artifacts rather than clinical
Analysis
TL;DR
- DAPF (Domain-Adapted models via Prompt-based Fine-tuning) achieves the best dementia detection performance among tested models, with accuracy=0.83 and macro-F1=0.83
- Diagnosis information is most recoverable from the [MASK] token representation in DAPF, confirming its representational strength
- Token-level explanation faithfulness does not match representational performance; attributions primarily reflect language task vocabulary, discourse markers, and transcription artifacts rather than clinically meaningful signals
- Perturbation tests reveal weak or even negative effects, indicating that the masked-token interface determines diagnosis without producing faithful token-level explanations
- The study highlights a critical gap between predictive performance and interpretability in prompt-based domain-adaptive models for low-resource medical screening
Why It Matters
This work is directly relevant to AI practitioners building explainable systems for healthcare applications, as it demonstrates that strong predictive performance does not guarantee faithful explanations—a key concern for clinical deployment. For researchers, it provides empirical evidence that [MASK]-based prompt interfaces may encode diagnostic information in ways that are not accessible through standard attribution methods, urging caution when relying on token-level explanations for high-stakes decisions.
Technical Details
- The DAPF framework reformulates dementia detection as a diagnosis-related masked-token prediction task, leveraging prompt-based fine-tuning on domain-adapted language models for spoken-language analysis
- Interpretability was evaluated using a variety of probing and analysis techniques, including perturbation-based faithfulness tests, to assess whether model attributions align with actual diagnostic signals
- DAPF outperformed strong baselines on accuracy (0.83) and macro-F1 (0.83), with the [MASK] representation showing the highest recoverability of diagnosis information among all analyzed representations
- Attribution analysis revealed that DAPF's token-level explanations were dominated by non-clinical factors: language task vocabulary, discourse markers, and ASR transcription artifacts, with perturbation tests yielding weak or negative faithfulness effects
Industry Insight
- Practitioners should not assume that high-performing prompt-based models produce trustworthy explanations; representational capacity and explanation faithfulness are decoupled, requiring explicit interpretability validation before clinical deployment
- For low-resource medical screening applications, investing in faithfulness-aware evaluation (e.g., perturbation tests) alongside standard accuracy metrics is essential to avoid deploying models whose explanations mislead clinicians
- The findings suggest that alternative interpretability approaches beyond token-level attribution—such as representation-level probing or latent-space analysis—may be needed to unlock the diagnostic reasoning embedded in masked-token interfaces
Disclaimer: The above content is generated by AI and is for reference only.