🤖 AI

A Comparative Explainability Framework for DeBERTa-v3 in Zero-Shot Medical Abstract Classification

This paper presents a comparative explainability framework that audits DeBERTa-v3's zero-shot classification of medical abstracts by evaluating the agreement of five attribution methods, revealing that explanatory stability correlates with predictive certainty and identifying systemic failure modes like lexical hypersensitivity to guide future model improvements.

Original authors: Javier Diaz Esteban-Herreros, David Muñoz-Valero, Raquel Martínez-España, Jose M. Juarez, Juan Moreno-Garcia

Published 2026-10-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Javier Diaz Esteban-Herreros, David Muñoz-Valero, Raquel Martínez-España, Jose M. Juarez, Juan Moreno-Garcia

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the modern world of artificial intelligence, computers have become remarkably skilled at reading and understanding human language. They can scan thousands of medical reports in seconds, spotting patterns that might take a human doctor hours to find. These systems, often built on complex structures called transformers, act like powerful engines that process text by looking at how words relate to one another. However, a significant problem remains: while these engines are accurate, they are often opaque. They function as black boxes, delivering a diagnosis or a classification without explaining which specific words led them to that conclusion. In the high-stakes field of medicine, where a wrong guess could affect a patient's safety, a doctor cannot simply trust a machine's output without understanding its reasoning. If a computer says a patient has a specific disease, the doctor needs to know if the machine noticed the right symptoms or if it was tricked by a single, misleading word. This need for transparency has given rise to a field dedicated to opening the black box, trying to make the internal logic of these machines visible and understandable to humans.

A team of researchers set out to test how well we can currently see inside one of these advanced language models when it is asked to classify medical abstracts without any prior training on specific medical data. They focused on a model called DeBERTa-v3, which is known for its ability to understand language nuances. The researchers did not teach the model to recognize diseases; instead, they asked it to guess the category of a medical text based on general knowledge it had already learned, a method known as zero-shot learning. To do this, they treated the task like a logic puzzle: for each medical abstract, the model was asked to decide if the text supported a specific description of a disease category. They tested this on five different groups of diseases: cancers, digestive issues, nervous system disorders, heart and blood vessel problems, and a broad catch-all category for general or systemic conditions.

The core of their investigation was to see if different methods used to explain the model's decisions would agree with one another. Imagine asking five different experts to point out the most important words in a sentence that led to a conclusion. If the experts are reliable, they should all point to roughly the same words. The researchers used five distinct techniques to generate these explanations, ranging from methods that treat the model as a mystery to be probed from the outside, to methods that look directly at the model's internal mathematical pathways. They then compared the lists of top twenty words each method highlighted for the same text. They measured how much these lists overlapped, essentially asking, "Do these different ways of looking at the model see the same thing?"

The results revealed a clear and telling pattern. When the model was classifying specific diseases like cancer or heart conditions, the different explanation methods largely agreed. They all pointed to the same key medical terms, such as "metastatic" for cancer or "liver" for digestive issues. In these cases, the model was confident, and the explanations were stable, suggesting that the system was indeed using the correct clinical logic. However, the situation changed dramatically when the model faced the broad category of general medical conditions. Here, the model's accuracy dropped significantly, and the explanation methods stopped agreeing. Instead of pointing to a shared set of medical terms, the different methods highlighted completely different words, often drifting toward common grammatical particles or unrelated terms. The lack of agreement among the explanation tools served as a warning signal that the model was struggling and that its decision was not based on a solid clinical foundation.

Through a detailed look at the errors the model made in this general category, the researchers identified three specific ways the system failed. First, the model showed a hypersensitivity to specific words. If a text contained a single strong word like "biopsy" or "malignancy," the model would immediately jump to a cancer diagnosis, even if the rest of the text described a routine procedure with no cancer context. Second, the model struggled with semantic overlap. When a text described a systemic issue that involved blood vessels, the model often confused it with a specific heart condition, unable to weigh the fact that the problem was actually affecting the whole body rather than just the heart. Finally, when the text lacked a clear, dominant medical clue, the model's explanation fell apart completely. The importance of the words scattered randomly across the sentence, with the model seemingly latching onto random grammatical words like "it" or "diagnose" rather than forming a coherent clinical reason.

The study concludes that relying on a single method to explain an AI's decision in medicine is insufficient. Because different methods can produce such different results, especially when the model is unsure, doctors and regulators need to look at the agreement between multiple explanation tools. If the tools agree, the decision is likely trustworthy. If they disagree, it is a sign that the model is confused or that the category it is trying to classify is too vague. The researchers suggest that for medical AI to be safe and auditable, we should focus on specific, well-defined disease categories rather than broad, general labels. By narrowing the scope to clear clinical definitions, we can ensure that the AI's reasoning remains stable and that the explanations it provides are actually useful for human experts.

Technical Summary: A Comparative Explainability Framework for DeBERTa-v3 in Zero-Shot Medical Abstract Classification

Problem Statement
The paper addresses the "disagreement problem" in Explainable Artificial Intelligence (XAI), where different attribution methods yield divergent or contradictory explanations for the same model prediction. This issue is particularly critical in high-stakes clinical settings, where regulatory frameworks (e.g., the European AI Act, GDPR) demand auditability and where spurious lexical cues must not be confused with legitimate clinical descriptors. While transformer models like DeBERTa-v3 have achieved state-of-the-art performance in medical text classification, their internal decision processes remain opaque. Furthermore, standard XAI toolkits often lack native support for DeBERTa-v3's specific architecture (disentangled attention), and existing comparative studies frequently rely on qualitative visual inspection rather than quantitative metrics to measure explanation stability.

Methodology
The authors propose a comparative explainability framework to audit DeBERTa-v3 under a zero-shot classification paradigm using the Medical Abstracts corpus. The methodology consists of five sequential stages:

Corpus and Sampling: A balanced sample of 1,000 texts per category was drawn from the Medical Abstracts corpus (Schopf et al., 2022), covering four specific diagnostic categories (Neoplasms, Digestive, Nervous, Cardiovascular) and one broad, semantically overlapping category (General Pathological Conditions).
Inference Engine: Classification was formulated as a Natural Language Inference (NLI) task. The model, deberta-v3-large-zeroshot-v2.0, was used without task-specific fine-tuning. For each abstract, five enriched hypotheses were generated per diagnostic category to reduce token sensitivity. The final prediction was derived by averaging entailment probabilities across these hypotheses.
Explanation Methods: Five distinct explanation methods were implemented, spanning three algorithmic families:
- Model-agnostic: SHAP and LIME (configured with 1,000 perturbations per sample).
- Perturbation-based: Occlusion.
- Gradient-based: Input $\times$ Gradient.
- Transformer-specific: Attention $\times$ Gradient.
- Note: Native implementations were created for Occlusion, Input $\times$ Gradient, and Attention $\times$ Gradient to accommodate DeBERTa-v3's disentangled attention mechanism, which standard libraries do not fully support.
Consistency Analysis: Explanations were standardized by selecting the Top-20 tokens with the highest absolute attribution values. Pairwise agreement between methods was quantified using the Jaccard index ( $J = |A \cap B| / |A \cup B|$ ) for each clinical category.
Error Analysis: A qualitative audit was performed on misclassified instances within the "General Pathological" category using SHAP saliency maps and waterfall plots to identify systemic failure patterns.

Key Contributions
The paper makes four primary contributions:

Zero-Shot NLI Engine: An operationalized zero-shot inference engine for medical abstracts using DeBERTa-v3 with multivariate prompt engineering (five enriched hypotheses per category).
Comparative Framework: A framework integrating agnostic, deep-learning-specific, and transformer-specific explanation methods adapted specifically for DeBERTa-v3, addressing a gap in the literature where this architecture is less studied than BERT or RoBERTa.
Quantitative Validation Protocol: A protocol using Top-20 token attribution and the Jaccard index to directly measure the disagreement problem across method families and clinical categories, moving beyond visual inspection.
Error Analysis: A diagnostic protocol identifying three systemic failure mechanisms (lexical hypersensitivity, semantic overlap, and loss of attribution coherence) that link predictive collapse to explanation structure in ambiguous clinical categories.

Results

Predictive Performance: The model achieved high recall in specific clinical domains (Neoplasms: 85.70%, Cardiovascular: 84.20%, Digestive: 76.70%, Nervous: 74.50%) but collapsed in the General Pathological category (6.00% recall), where errors were distributed across the specific categories.
Explanatory Agreement: There is a strong positive correlation between predictive certainty and explanatory stability.
- In specific categories, methods showed moderate convergence (e.g., Neoplasms mean Jaccard: 0.33; Cardiovascular: 0.28).
- In the General category, agreement dropped significantly (mean Jaccard: 0.17), with some method pairs (e.g., LIME vs. Occlusion) reaching a minimum of 0.036.
Method Divergence: Agreement was highest within algorithmic families (e.g., SHAP–LIME mean: 0.31; Input $\times$ Gradient–Attention $\times$ Gradient mean: 0.30) and lowest between agnostic and attention-based methods (e.g., SHAP–Attention $\times$ Gradient mean: 0.19).
Failure Patterns: The error analysis identified three recurring mechanisms in the General category:
- Lexical Hypersensitivity: Over-reliance on single high-salience tokens (e.g., "biopsy," "malignancy") that trigger premature category assignment regardless of context.
- Semantic Overlap: Inability to weigh negative evidence when shared descriptors (e.g., vascular terms) overlap with specific organ specialties.
- Loss of Attribution Coherence: When no dominant clinical feature exists, attribution mass disperses onto grammatical particles (e.g., "it," "diagnose"), resulting in erratic explanations.

Significance and Claims
The paper claims that explanatory agreement is not a static property of the model but is dynamically conditioned by the clinical specificity of the category and the family of the explainer. The authors argue that:

Single-explainer audits are insufficient for transformer-based medical classifiers because attribution paths vary significantly by method family.
Quantitative agreement metrics (like the Jaccard index) serve as a diagnostic signal for model reliability; low agreement correlates with high uncertainty and predictive failure.
In clinical NLP, broad diagnostic labels (like "General Pathological Conditions") are inherently harder to audit and less safe for automated decision-making than specific ontologies due to semantic overlap and lexical ambiguity.
The study supports the recommendation to prioritize specific clinical ontologies over broad diagnostic labels when auditability is a regulatory requirement.

The authors maintain a modest scope, noting that their comparative analysis used a single-case-per-category design ( $N=1$ ) to isolate inter-method variance, and that future work is needed to validate these findings across larger populations and with ranking-aware metrics.

Technical Summary: A Comparative Explainability Framework for DeBERTa-v3 in Zero-Shot Medical Abstract Classification

More like this