Simple models beat fancy AI at predicting hospital discharge from surgery notes

Researchers tested 13 different models to predict whether spine surgery patients would be ready to go home the next day, using clinical notes written after surgery. A basic statistical model (TF-IDF with LGBM) worked better than smaller AI language models, achieving 80% accuracy with half the computational cost.

Hospitals need to know who's going home tomorrow so they can allocate beds and staff efficiently. The finding is straightforward: when you have messy, real clinical data and an imbalanced prediction problem (most patients aren't discharged the next day), the fancier AI doesn't help. This matters because it pushes back against the assumption that specialized language models are always the right tool for medical prediction tasks. If this pattern holds across other clinical prediction problems, hospitals could save money by using simpler, faster models instead of investing in fine-tuning smaller AI systems.