Prompt Description

LLM-based drug recommendation + safety (DDI) + long-tail/imbalance + molecular structure alignment + Stage2 frequency-aware fusion. Top Left: LLM output resulting in a safety failure due to drug-drug interaction. Illustrate the input for a patient visit 't' (Dx/Proc/Text) → LLM → Recommended drug set {A, B, C}. Within {A, B, C}, highlight a DDI pair (e.g., A–B) with a red lightning bolt, and label it as "unsafe combination / DDI". Top Right: Why this occurs (lack of structural constraints). Label it as "atomic label embeddings + co-occurrence shortcut". Illustrate an unstructured label space: drug points are disorganized, and unsafe pairs are in close proximity. Bottom Left: Long-tail/imbalance failure. Include a frequency binning histogram from MIMIC-IV (0–50, 50–100…>5000), showing a high head and a long tail. Label it as "tail labels: sparse supervision → poor recall / calibration". Bottom Right: Our solution (structure + frequency-aware). Stage 1: Drug node graph (DDI edges in red, EHR co-occurrence edges in blue, temporal edges as dashed lines), where each ATC3 is a 'multi-prototype' small cluster. Stage 2: Schematic bar weights illustrating LLM + frequency-aware fusion (tail relies more on molecule prior; head relies more on task signal).