The paper introduces a bounded updating framework for ICU prediction models that freezes the physiological encoder while allowing updates only to the treatment pathway and fusion head. Experiments on 84,792 MIMIC-IV ICU stays across four temporal shifts show that selective adaptation yields more stable explanations—higher rank correlation, better top‑5 feature agreement, and improved retrieval stability—compared to full model adaptation. Predictive performance varies by outcome, but the results demonstrate that explanation stability is governed by the structural boundaries of allowed adaptation rather than merely by freezing components.
arXiv:2607. 19020v2 Announce Type: replace-cross Abstract: Clinical decision support degrades as treatment protocols evolve, but the obstacle to updating a deployed model is governance as much as accuracy: once retraining touches every parameter, no one can say afterwards where the update acted.
The paper investigates how to evaluate bounded model updates for ICU prediction models using the MIMIC‑IV database. After auditing the data for recording artifacts and cleaning it, the authors compare a bounded update that retrains only treatment‑context and fusion parameters with full fine‑tuning across two model families and two targets, finding that bounded updates are non‑inferior or superior in most cases but depend on target, model family, and learning‑rate tuning. The study also shows that audits should precede update evaluation and that the update boundary should be chosen per target.
arXiv:2607. 19020v1 Announce Type: cross Abstract: Background: Clinical decision support systems degrade silently as treatment protocols evolve, yet standard adaptation methods treat models as monolithic blocks, unable to distinguish stable patient physiology from shifting institutional practice.
arXiv:2607. 26752v1 Announce Type: new Abstract: Medical world models aim to learn a latent state of patient or organ physiology and a transition function that forecasts how that state evolves under interventions, supporting downstream tasks from imaging-based diagnosis to digital-twin treatment planning.
arXiv:2605. 10840v3 Announce Type: replace-cross Abstract: We present Clin-JEPA, a multi-phase co-training framework for joint-embedding predictive (JEPA) pretraining on EHR patient trajectories.
The paper introduces a response‑aware patient dynamics model that explicitly captures how respiratory support alters a patient’s short‑term physiological trajectory. By decomposing predicted physiological change into baseline dynamics and support‑associated deviations anchored to room‑air reference, the model updates latent patient states across a forecast horizon. Experiments on ICU cohorts from two institutions show that the model matches baseline trajectory predictions overall and consistently improves performance when physiological states are changing.
The study examined how physiological trajectories and observation patterns influence ICU mortality prediction across different datasets (MIMIC‑III, MIMIC‑IV, eICU). Models incorporating longitudinal physiological summaries (R2) consistently improved external prediction, while observation‑pattern features (R3) mainly benefited locally developed models and were highly sensitive to measurement timing and availability. The findings highlight that the transportability of clinical prediction models depends not only on patient populations but also on the conditions under which measurements are collected.
The paper investigates how the granularity of interventions—single components versus full treatment bundles—affects counterfactual simulations using a clinical world model. By analyzing 945,707 patient-hours from MIMIC-IV, the authors show that interventions are naturally bundled, and editing a single component often represents an unreal scenario. Experiments with Clin‑JEPA on 1,019 ventilation onsets demonstrate that editing the complete bundle produces larger predicted state changes than editing individual settings, indicating that bundle-aware editing better captures treatment sensitivity.
arXiv:2608.29301v1 Announce Type: new Abstract: Predicting future organ dysfunction in Intensive Care Unit (ICU) patients is critical for early clinical intervention, yet existing machine learning ap...
PulseBound is a photoplethysmography (PPG) representation learner that enforces an explicit stored‑window information boundary, ensuring that encoder inputs depend only on a visible prefix and a content‑independent cutoff. By combining physiologically structured future‑beat prediction with prefix‑only normalization, suffix replacement, and aligned masking, PulseBound achieves stored‑suffix invariance, preventing changes in the stored suffix from affecting forecast context or predictions. Evaluated on MIMIC and VitalDB datasets, PulseBound reduces mean absolute error by over 22% compared to last‑visible‑beat persistence and outperforms seven other models on nine frozen linear‑probe and seven full‑fine‑tuning downstream tasks.
The paper introduces CELLAUDIT, a method for auditing whether inputs claimed to influence predictive models actually do so. By testing if an input can enter the computation, whether predictions depend on it, and if that dependence improves observed responses, the authors evaluate agent-generated predictors on a morphology‑transcriptomics benchmark (BBBC047). Their findings show that many models claim compound contributions that are not supported by the data, and that falsification‑guided revisions can recover genuine input effects while improving performance.
arXiv:2606. 23487v2 Announce Type: replace Abstract: Medical vision-language models (VLMs) such as BiomedCLIP generalize broadly, but adapting them to a clinical service is as much a safety problem as an accuracy one.