On August 28, 2026, two new papers published on arXiv (cs.AI) introduced solutions leveraging large language models (LLMs) and agentic workflows to process complex data in intensive care units (ICUs) and electroencephalogram (EEG) diagnostics.
Agentic Workflows for ICU Mortality Risk Interpretation
In the ICU study based on the eICU Demo dataset—comprising 2,353 admissions with an 8.1% mortality rate—an XGBoost model achieved an AUROC of 0.855 (95% CI: 0.796–0.906) and an AUPRC of 0.332 for predicting in-hospital mortality. To translate these technical risk metrics into clinical explanations for physicians, the researchers compared a standalone LLM with a four-step agentic pipeline that decoupled data ingestion, clinical guideline alignment, and final narrative generation. In a test across 38 cases, the standalone LLM suffered one instance of outcome leakage, whereas the agentic pipeline completely avoided this error.
A deeper comparative evaluation against SHAP (SHapley Additive exPlanations) values revealed distinct trade-offs. The standalone LLM adhered more closely to raw SHAP attribution rankings, achieving an average Jaccard index of 0.171 (versus 0.077 for the agentic pipeline) and a 92.9% directional consistency (compared to 78.6%). Conversely, the agentic pipeline showed clear superiority in clinical guideline adherence (0.762 vs. 0.143), patient-specific medical detail (0.236 vs. 0.143), and overall clinical plausibility (0.700 vs. 0.671). The authors concluded that while agentic architectures enhance safety and granular patient context, integrating attribution verification remains critical before deployment in high-stakes environments.
Automated EEG Clinical Reporting via Multimodal Frameworks
Addressing neurodiagnostics, the second paper presented 'EEG-to-Report', a browser-based system designed to overcome the severe scarcity of paired signal-text datasets required to train clinical AI models. The framework standardizes raw EEG signal channels and integrates a multimodal annotation interface combining keystroke data with synchronized voice dictations.
The system simultaneously extracts spectral, temporal, entropy, Hjorth, and sharp-wave features, storing them alongside clinician notes in structured JSON format. 'EEG-to-Report' combines an ensemble of convolutional neural networks (CNNs) with an LLM to automatically generate draft diagnostic reports for neurologist review.
Both publications underscore an emerging consensus in healthcare AI: rather than relying on standalone LLMs, optimal clinical integration requires multi-step agentic pipelines, structured intermediate data representations, and systematic oversight mechanisms.