CPS-Net (Collaborating Physicians in Silico Network): A Decentralized Multi-Agent Transformer Framework for Specialty-Aware Disease Prediction

Abstract Diagnosis in real clinical settings is rarely finalized after a single encounter. Patients with complex conditions move through a referral network from primary care to specialists because physicians cannot maintain expertise across all diseases. Yet most diagnostic prediction tools apply a single generalist model across thousands of medical codes, which dilutes representational capacity for rare conditions and produces outputs poorly aligned with how clinicians actually work and reason. We present CPS-Net, a decentralized agent swarm that mirrors medical specialization hierarchies. The system connects specialized transformers trained on focused patient cohorts to language model agents organized into primary care, specialty, and subspecialty tiers. Agents collaborate without central supervision, communicating through shared case memory to make referral decisions transparent and auditable. We trained 41 transformers on Electronic Health Records (EHR) from 75,000 patients across 10 specialties and 30 subspecialties at Froedtert Hospital. Evaluated on 3,000 test cases, the framework achieved top-1, top-3, and top-5 accuracies of 48.8%, 71.00%, and 77.76% with 94.17% clinical relevance, substantially outperforming a monolithic transformer with 19.20%, 38.60% and 47.20% in top-1, top-3 and top-5 accuracy trained on all patients and a single language model agent baseline which achieved 12.70%, 26.40%, and 37.23% accuracy. This work demonstrates that aligning artificial intelligence architecture with clinical workflows improves both prediction accuracy and interpretability for diagnostic support.