CS Paper Radar

Making Clinical Language Models Auditable: Concept-Guided Fine-Tuning for Robust Prediction

Jin Mu, Guanhua Chen

cs.CL cs.AI ·submitted 2026-08-27 17:28 UTC · v1
top venue

Abstract

Clinical language models can achieve strong in-hospital accuracy yet fail under deployment shifts because they exploit note-specific artifacts (e.g., templates, separators, boilerplate) that do not reflect patient state. We propose CAST (Concept-guided Artifact Suppression Tuning), an SAE-based framework for auditable clinical text classification. CAST uses Sparse Autoencoders to expose sparse, human-auditable features from intermediate Transformer activations, labels SAE latents with an LLM-assisted interpretation pipeline and ICD-10 retrieval constraints, suppresses verified artifact latents via residual subtraction during fine-tuning, and provides post-hoc per-concept attributions for auditing model decisions. On MIMIC-IV discharge-note mortality prediction, CAST improves over its corresponding fine-tuned encoder baselines and remains competitive with strong LLM baselines, while producing a feature-level audit trail of the clinical concepts that support each prediction and the artifact concepts suppressed during training.

Closest papers

Nearest neighbours by meaning across the whole index.

Temporal Island Sparse Autoencoders for Interpreting Clinical Time-Series Models

cs.LG ·2026-11-02 ·Muhan Yeo, Bong Gyun Kang, HyunGi Kim, Sungroh Yoon
top venue similarity 0.71
Abstract · Accepted at NeurIPS 2026

Deep learning models for electronic health records achieve strong predictive performance but provide little insight into what clinical concepts they represent internally. Post-hoc attribution methods produce per-cell importance scores that lack temporal coherence, cross-variable structure, and reusability across patients. We introduce the Temporal Island Sparse Autoencoder (TI-SAE), which learns a backbone-specific, patient-shared clinical concept dictionary from a frozen predictive model's representations. Each dictionary entry (latent) is defined by learned temporal Gaussian islands (specifying when the concept is active) and a sparse variable-composition vector (specifying which variables are involved). A decoupled selector then identifies which dictionary entries drive each patient's prediction, separating representation learning from attribution. We evaluate TI-SAE across five backbone architectures spanning different representation strategies on MIMIC-III and MIMIC-IV (three prediction tasks each). TI-SAE produces structured clinical concepts that go beyond transient attribution heatmaps while achieving strong performance on standard faithfulness benchmarks.

0.31

Knowledge-Enriched Structured EHR Features for 30-Day Hospital Readmission Prediction on MIMIC-IV

cs.LG q-bio.QM ·2026-09-14 ·Mohamad Najafi, Hongyun Fu, Mathias Brochhausen, Jian Wu, Yaohang Li
published similarity 0.70
Abstract · 15 pages, 3 figures. Accepted at SDSC 2026 Mid-Atlantic

Recent approaches to 30-day hospital readmission prediction rely on pre-trained language models applied to discharge summaries. Although these methods achieve strong performance, they depend on the availability of clinical notes, incur substantial computational costs, and yield representations that lack interpretability. We propose a knowledge-enriched feature representation that augments structured Electronic Health Record (EHR) data with four medical knowledge sources: disease ontology mapping, procedure classification, drug ingredient vocabulary, and organ system laboratory aggregation, without using clinical notes. Each feature dimension corresponds to a named clinical concept, yielding a sparse and interpretable patient representation. The approach is evaluated with six classifiers on a MIMIC-IV v2.2 cohort. Under 20-fold cross-validation, the best configuration achieves an AUROC of 0.743. This performance is comparable to that of previously reported methods on this dataset, including both those using only structured data and those incorporating clinical notes, while requiring considerably less computational cost. Interpretability analysis shows that demographics, organ system labs, drug ingredient features, and first-level ontology disease categories drive prediction, while deeper hierarchy levels contribute negligibly. These findings indicate that knowledge-enriched structured features offer a competitive and efficient alternative to embeddings from clinical notes for 30-day readmission prediction.

0.25

Evaluating Pretrained Language Semantics for Structured EHR Prediction Tasks

cs.CL cs.LG ·2026-11-12 ·Ashvin Gupta, Brendan Delaney, Alessandra Russo
top venue similarity 0.69
Abstract · Accepted for publication in Proceedings of Machine Learning Research (PMLR), Vol. 346, 2026

Structured Electronic Health Record (EHR) models typically represent clinical concepts as dataset-specific tokens, discarding their associated natural-language semantics. Consequently, relationships between concepts must largely be learned from EHR co-occurrence without directly exploiting pretrained linguistic knowledge. We investigate the contribution of pretrained language semantics to longitudinal EHR prediction through tEHR, a framework that converts clinical events into natural-language descriptions and adapts pretrained autoregressive language models through EHR-specific continued pretraining and task-specific discriminative fine-tuning. We evaluate tEHR across four MIMIC-IV prediction tasks, where Mistral-7B achieves the strongest overall performance against structured-data, code-based, and LLM-based baselines. Controlled ablations demonstrate the importance of lexical grounding, numerical values, and both adaptation stages. A capacity-controlled experiment further shows that replacing clinical descriptions with atomic code tokens substantially reduces performance despite retaining the same LLM backbone. Alternative approaches incorporating pretrained semantic embeddings offer lower computational costs but do not match full textualisation. These results highlight the importance of both semantic representation and downstream adaptation when applying pretrained language models to structured EHR prediction.

0.29

MIMIC-Note-Extract: A Large-Scale Structured Dataset Extracted from MIMIC-IV-Note Discharge Summaries

cs.CL ·2026-11-03 ·Kristian Schwethelm, Jonas Kuntzer, Conrad Ketzer, Eva Wende, Moritz Schlager +4 more
publishedbenchmarkdataset similarity 0.69
Abstract · Accepted at ML4H 2026, Proceedings Track. Dataset: https://physionet.org/content/mimic-iv-ext-note-extract/1.0/

MIMIC is one of the most widely used public hospital datasets for machine learning on electronic health records, yet clinically important information, such as past medical history, allergies, and diagnostic findings, remains confined to free-text discharge summaries. Studies requiring this information must therefore build their own extraction pipelines, duplicating substantial effort across projects. We release MIMIC-Note-Extract (MNE), a large-scale dataset that uses a hybrid extraction pipeline combining deterministic parsing with schema-guided large language model generation to transform 16 sections from 300,720 discharge summaries into structured tables. MNE adds information that has no structured MIMIC counterpart and enriches related tables with broader coverage and greater clinical detail, including 10 times more patient-history entries and medication histories for more than twice as many admissions. Validation by medical doctors confirms extraction accuracy above 92% across sections. Using MNE, we construct a clinical decision-making benchmark for LLM agents, demonstrating how reusable structured representations of clinical narratives reduce data curation effort.

0.24

Toward Better Assessment of LLMs' Performance in Clinical Error Detection

cs.CL cs.AIcs.LG ·2026-08-17 ·Yifan Zhang, Rahmatollah Beheshti
top venuebenchmarkcode/data similarity 0.69
Abstract · Accepted at Machine Learning for Healthcare (MLHC) 2026; to appear in Proceedings of Machine Learning Research (PMLR), Vol. 340

Automated detection of errors in clinical documentation is a promising application of large language models (LLMs), yet decisions to deploy such models rest on benchmarks that evaluate each clinical note in isolation. Error-detection benchmarks are typically constructed by injecting errors into notes, such that each erroneous note has a natural counterpart. Aggregate discriminative metrics (e.g., balanced accuracy or F1) do not exploit this structure. We show that this omission is consequential. In particular, evaluating 15 diverse LLMs on 4 standardized clinical error-detection test sets across 3 languages, we find that 13 of 15 models fall below the level of random pairwise discrimination, even while achieving F1 scores that standard practice would read as moderate. We also observe that the underlying bias patterns differ across languages: the same model can default to "no error" on one language and over-flag errors on another. To diagnose where discrimination breaks down, we further introduce a procedure to score the evidence models cite in their outputs. We find that while models consistently locate error-relevant content, they fail to produce the corresponding correct verdict on the clean counterpart. Finally, we show that F1 and pairwise accuracy are driven in opposite directions by the same underlying bias, so that ranking models by F1 may systematically promote the weakest discriminators. For safety-critical clinical NLP applications, we advocate for supplementing aggregate metrics with paired evaluations in benchmark reporting. Code and analysis scripts are available at https://github.com/healthylaife/paired-clinical-eval.

0.58

An end-to-end-trained vision-language model for native-language prostate pathology report generation

cs.CV ·2026-08-24 ·Christian Grashei, Fabian Gülhan, Maximilian Legnar, Fabian Stögbauer, Cleo-Aron Weis +2 more
similarity 0.68
Abstract

Prostate cancer is among the most frequently diagnosed malignancies worldwide, and structured reporting of each biopsy core burdens pathologists. Existing tools frame this as classification, leaving pathologists to assemble coherent reports, while many slide-level vision-language models rely on English-centric encoders that transfer poorly to other clinical languages. We present a slide-level framework generating prostate biopsy reports that is language-independent by construction: tokenizer and model are trained from scratch, demonstrated here in German. To address paired-data scarcity, an automated pipeline uses a locally deployed large language model to split composite reports into core-specific image-text pairs, yielding 17,344 pairs from 2,402 historical cases without manual annotation. Evaluated for clinical attributes rather than linguistic similarity, the model achieves 96.2% F1 for malignancy detection and 65.2% for Gleason grading, competitive with an FDA-cleared classifier. Grading is further validated on three external cohorts with latent-space augmentation. Institutions can thus train native-language reporting models on their own archives.

0.14

Building Auditable AI Text Detectors with Sparse Autoencoders

cs.CL cs.AI ·2026-10-21 ·Liam Dugan, Callum McDougall, Matthew Denton, Philip Pham, Christine Kaeser-Chen +1 more
top venue similarity 0.68
Abstract · EMNLP 2026

AI-generated text detectors are increasingly deployed in consequential settings---peer review, education, admissions, journalism, and content moderation---where their outputs directly affect individuals. Yet the strongest detectors are fine-tuned neural classifiers whose decisions are opaque, while transparent alternatives (stylometric and linguistic-feature classifiers) underperform them by wide margins. We present a pipeline that converts a fine-tuned neural detector into a classifier that operates on human-readable concepts. We accomplish this by training sparse autoencoders (SAEs) directly on a detector's residual stream, assigning auto-interpretation labels to each SAE feature, and retaining only features whose generated descriptions pass a global fidelity threshold. We then train a lightweight classifier on just the interpretable features. We find that our SAEs recover latent concepts that correspond to highly specific and predictive linguistic "tells" of AI-generated text. Classifiers built on these extracted features are not only highly interpretable but exhibit stronger performance on out-of-distribution texts than either neural or stylometric alternatives.

0.50

ChestPheNoT: Deployable, Auditable Label-Status-Evidence Extraction from Radiology Reports

cs.CL cs.AIcs.LG ·2026-09-09 ·Kai Yu, Chenyu Zhu, Zaifu Zhan, Meijia Song, Min Zeng +3 more
publishedcode/data similarity 0.68
Abstract · Accepted at IEEE Healthcom 2026

Structured phenotype extraction from radiology reports supports cohort construction, quality auditing, and clinical analytics, but practical deployment requires local inference and auditable predictions, while expert annotations remain scarce. Conventional labelers provide structured findings and assertion states but no supporting evidence, while API-hosted large language models may be unsuitable when clinical text cannot leave institutional infrastructure. We present CHESTPHENOT, a compact 0.5-3B language model that jointly extracts finding labels, three-class status (present/absent/uncertain), and verbatim supporting evidence spans. CHESTPHENOT is trained using hybrid CheXbert+72B silver supervision followed by supervised fine-tuning and lightweight GRPO refinement. Across three human-annotated gold sets spanning in-distribution, cross-taxonomy, and cross-institution evaluation, the 3B model remains below its CheXbert silver teacher in distribution but is competitive under distribution shift, significantly surpassing CheXbert on cross-institution detection (+2.0 F1). Task-specific training also enables the 3B model to match or exceed substantially larger prompted models on most detection and status comparisons. For evidence-grounded extraction, over 99% of final evidence spans are locatable in the source report, and the 3B model achieves 47.5 auditable-F1, outperforming Qwen2.5-7B one-shot prompting by 7.6 points and approaching Qwen2.5-72B. These results demonstrate that locally deployable models can provide competitive and directly auditable radiology-report extraction without relying on external inference APIs. Code and the full extraction/judge prompts will be made available at https://github.com/yukkai/ChestPheNoT.

0.43