[Paper] QIME: Constructing Interpretable Medical Text Embeddings via Ontology-Grounded Questions

Source: arXiv - 2603.01690v1

Overview

The paper introduces QIME, a new way to turn medical text (e.g., clinical notes, research abstracts) into embeddings that are both high‑performing and human‑readable. Instead of opaque dense vectors, QIME represents each document as a series of yes/no answers to clinically grounded questions derived from biomedical ontologies. This makes the embeddings interpretable for clinicians and developers while still rivaling the accuracy of black‑box models.

Key Contributions

• Ontology‑grounded question generation: Leverages medical concept signatures (e.g., SNOMED‑CT, MeSH) to automatically craft fine‑grained, clinically meaningful binary questions.
• Training‑free embedding construction: Bypasses the need to train a separate classifier for every question; answers are obtained directly from the language model’s masked‑language‑model (MLM) probabilities.
• Interpretability without sacrificing performance: Achieves state‑of‑the‑art results on biomedical similarity, clustering, and retrieval tasks, closing the gap to dense black‑box encoders.
• Scalable and modular design: New medical domains or ontologies can be plugged in with minimal engineering effort.

Methodology

1. Concept Signature Extraction

   • For each medical ontology cluster (e.g., “cardiovascular diseases”), the authors collect a set of representative terms (signatures) using TF‑IDF and ontology hierarchy information.

2. Question Generation

   • A template (“Does the text mention X?”) is instantiated with each signature term, producing a pool of candidate yes/no questions.
   • A lightweight scoring model ranks questions by semantic atomicity (how specific and non‑redundant they are) and clinical relevance.

3. Answering via MLM

   • Instead of training a binary classifier per question, QIME feeds the question and the target text into a pretrained biomedical language model (e.g., BioBERT).
   • The model’s masked token probability for “yes” vs. “no” yields the binary answer, forming one dimension of the final embedding.

4. Embedding Assembly

   • The concatenated vector of binary answers (e.g., 256‑dim) constitutes the interpretable embedding.
   • Because each dimension maps to a concrete question, developers can read out why two documents are similar (e.g., both answer “yes” to “Is hypertension mentioned?”).

Results & Findings

Key takeaways

• The training‑free variant even outperforms the classifier‑based version, showing that MLM‑based answering is a strong signal.
• Qualitative analysis reveals that the top‑ranked questions often correspond to clinically decisive concepts (e.g., “Is the patient on anticoagulants?”).

Practical Implications

• Explainable AI for Clinical Decision Support: Developers can embed patient notes and instantly surface the exact clinical facts driving similarity scores, aiding trust and auditability.
• Rapid Prototyping of Retrieval Systems: Since no per‑question training is required, teams can spin up a searchable knowledge base by simply defining the ontology of interest.
• Regulatory Compliance: Interpretable embeddings satisfy emerging “right‑to‑explain” requirements in healthcare AI deployments.
• Domain Adaptation: Adding a new specialty (e.g., oncology) only needs the corresponding ontology; QIME automatically generates relevant questions without retraining the whole model.

Limitations & Future Work

• Ontology Dependence: The quality of embeddings hinges on the completeness and granularity of the underlying medical ontology; rare or emerging concepts may be missed.
• Binary Question Scope: Complex relations (e.g., temporal or causal) are reduced to yes/no, potentially oversimplifying nuanced clinical narratives.
• Scalability of Question Set: Very large signature pools can inflate embedding dimensionality; future work could explore adaptive pruning or hierarchical question encoding.
• Evaluation on Real‑World Clinical Workflows: The paper focuses on benchmark datasets; deploying QIME in live EHR systems and measuring impact on clinician workflow remains an open step.

Authors

• Yixuan Tang
• Zhenghong Lin
• Yandong Sun
• Anthony K. H. Tung

Paper Information

• arXiv ID: 2603.01690v1
• Categories: cs.CL, cs.AI
• Published: March 2, 2026
• PDF: Download PDF