A four-step AI pipeline produced more clinically grounded explanations of intensive-care mortality predictions than a standalone language model in a small feasibility study. The pipeline separated data interpretation, guideline checking and final explanation instead of asking one prompt to perform the whole task.
Researchers used records from 2,353 stays in the eICU Demo dataset, where mortality was 8.1%. An XGBoost risk model reached an area under the receiver operating characteristic curve of 0.855. On a stratified set of 38 explanations, the standalone model explicitly leaked the known outcome once, while the agent pipeline did not.
The trade-off appeared in 14 cases also reviewed with SHAP, a method for estimating how individual features influence a prediction. Standalone explanations aligned more closely with the SHAP features and their direction. The pipeline scored higher for citing clinical guidance, including specific patient values and overall plausibility. Neither approach dominated every measure. The authors conclude that agent decomposition may improve safety-relevant context but should be paired with attribution checks before high-stakes use. The evaluation is too small for clinical deployment claims, and it explains a model’s output rather than proving the underlying prediction is correct. Larger datasets and clinician testing are needed before such narratives could support bedside decisions.