CAST (Concept-guided Artifact Suppression Tuning), an SAE-based framework for auditable clinical text classification, 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.
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.
Across application domains, predictive signals often sit in unstructured free text rather than structured fields, yet turning that text into useful and interpretable features is difficult. Running large language models (LLMs) over an entire corpus is costly and hard to reproduce, while end-to-end text representations can rely on surface cues that are difficult to inspect. We present an agentic workflow that takes a prediction task and a raw text corpus as input and produces an auditable feature layer. The first two agents use an LLM to derive a task-specific predictor taxonomy and weakly label a bounded text sample; routed local extractors then process the corpus, and a deterministic builder aggregates the evidence into a dynamic, longitudinal concept bottleneck. We evaluate the framework on medication discontinuation in a longitudinal oncology cohort and 30-day readmission in MIMIC-IV. With gradient boosting, the longitudinal bottleneck increases area under the receiver operating characteristic curve (AUROC) over coarse concept buckets from 0.700 to 0.761 for medication discontinuation and from 0.576 to 0.609 for readmission. The proposed framework achieves predictive performance comparable to direct BioClinicalBERT prediction on both tasks while additionally providing explicit, interpretable, and traceable task-specific concepts. LLM use is confined to a bounded weak-labeling stage costing $24.00 and $23.39, respectively, compared with projected costs of $10,648 and $11,519 for exhaustive sentence-level LLM processing of the full corpora, demonstrating the substantial cost efficiency of the proposed agentic system.
C. Ni, W. Liu, Q. Song et al.· medRxiv· 0 citations
A robust technical framework is established for developing trustworthy, high-efficiency medical AI systems capable of operating entirely within hospital-controlled infrastructure by evaluating the synergistic impact of context window scaling and multi-stage supervised fine-tuning within a localized Retrieval-Augmented Generation framework.
L. Pawlik, Stanisław Deniziak· Scientific Reports· 0 citations
BERT-LER is presented, a BERT-style model for coded EHR timelines pretrained and fine-tuned from a de-identified EHR dataset of 75 million patients, that encodes laboratory test results as discrete tokens while retaining graded information through percentile-based binning, paired with Integrated Gradients for token-level attributions grounded in the input EHR sequence.
Jun-Ni Du, Lukas Adamek, Maxim A Kryukov et al.· 0 citations
Clinical coding agents repeatedly encounter the same failure modes, including unsupported codes, missed documented conditions, specificity errors, and procedure-coding convention mismatches. We introduce Learn-Then-Act, an inference-time adaptation framework that converts errors from a small labeled LEARN batch into a structured Mistake Knowledge Database (MistakeKDB). False-negative lessons are routed to a recall-oriented Coder, while false-positive lessons are routed to a precision-oriented Judge. We instantiate the framework in LearnActCoder, a Coder-Judge clinical coding pipeline with lookup-table grounding where available. On 150 matched MIMIC-III notes, structured MistakeKDB improves CPT F1 by 5.9 percentage points, while raw-example and reflection-style memories remain near the no-memory baseline; the ICD-9 improvement is not significant. On a matched MIMIC-IV cohort, memory shifts ICD-10 coding toward higher precision at a recall cost, leaving F1 statistically unchanged. Applying the same memory to 1,000 held-out MIMIC-III notes maintains a stable ICD operating point, providing scale/stability evidence. Overall, the results are consistent with structured, feedback-derived error memory being useful for adapting clinical coding behavior across cases without weight updates or changes to the underlying workflow. Absolute CPT/HCPCS performance remains low, and the system is evaluated retrospectively rather than in clinical deployment.
Meysam Ghaffari, Bhaskar Sen, Nasim Sabetpour et al.· 0 citations
Findings show that clinical LLM explainability has shifted toward fluent generative rationales, but evidence that such explanations reflect model reasoning remains limited, and three regulatory priorities are highlighted: prioritizing explanations that enable independent verification or logic auditing over plausibility-only rationales; preferring inspectable models where regulatory documentation is required; and prospectively validating explanations in clinical workflows before scaling.
