Laya-Med: Efficient Clinical Decision Support with a Compact Language Model

Clinical decision support requires models that can interpret patient records with short response times and modest computational demands. We present Laya-Med, a clinical adaptation of a 322-million-parameter bidirectional language model that assigns probabilities to predefined outcomes in a single forward pass. We evaluated the approach in two retrospective studies. For acute abdominal diagnosis, we fine-tuned the model on 1,680 MIMIC-IV-Ext Clinical Decision-Making admissions and compared it with three zero-shot large language models and a term frequency–inverse document frequency (TF-IDF) classifier on 360 held-out admissions. Laya-Med achieved macro-F1 of 81.1% (95% confidence interval 76.1–85.7) and accuracy of 85.6%, compared with 84.6% and 86.9% for Qwen3.8-27B. Mean processing time under batched inference was 10.8 ms per case, versus 0.60 s for Qwen3.8-27B on the same GPU. In a second study, we independently fine-tuned Laya-Dengue on Portuguese narratives generated from dengue surveillance records. On 2,157 test cases, its macro-averaged area under the receiver operating characteristic curve was 95.2%, compared with 94.0% for TF-IDF and 91.1% for MedGemma 27B. For non-fatal severe dengue versus death, the corresponding values were 73.9%, 67.7% and 54.7%. Local interfaces updated predictions in approximately 50 ms. These results support compact, task-adapted encoders as a computationally efficient approach to clinical decision support. Prospective evaluation is needed to establish their performance on clinical workflows and outcomes.