Large language models (LLMs) can reason over structured measurements and free‑text records, making them attractive for emergency department (ED) prediction. Yet on most EHR‑based tasks they lag behind conventional machine‑learning methods such as tree models. Controlled experiments trace the gap to two weaknesses: LLMs do not actively select informative features, and even with the right features they misinterpret the risk implied by numeric values. To address both we introduce TIR‑EC (Tool‑Integrated Reinforcement Learning for Emergency Care). The framework exposes groups of EHR features as tools the LLM can call, and trains the LLM with rewards from a gradient‑boosted‑tree (GBT) teacher that supplies a reference set of useful tools and a graded risk target. During training a reverse curriculum gradually removes task‑specific clinical skills from the prompt, forcing the model weights to internalize those skills. Evaluation on three clinical prediction tasks shows that the 4B‑parameter agent achieves the highest macro F1, outperforms all conventional baselines, and is competitive with much larger frontier LLMs. Further analysis reveals that TIR‑EC selects more useful evidence with fewer tool calls and combines measurements more effectively.
Review