AI Doctor Swarm Mirrors Hospital Referrals to Sharpen Disease Prediction

Diagnosis in real clinical practice is rarely a single-step affair. A patient with a confusing set of symptoms typically starts with a primary care physician, who weighs the history, orders initial tests, and—if the problem falls outside generalist expertise—sends the patient onward to a specialist. For particularly complex conditions, care may pass again to a subspecialist whose experience is concentrated in a narrow disease domain. Most artificial intelligence diagnostic tools ignore this layered reality entirely, applying one enormous generalist model across thousands of medical codes at once. A new study argues that this mismatch is not just cosmetic: it measurably hurts accuracy and leaves clinicians without a reasoning trail they can trust.

Researchers have now introduced CPS-Net, short for Collaborating Physicians in Silico Network, a decentralized multi-agent framework that deliberately mirrors the referral structure of a hospital. Published in the Journal of Medical Systems, the work describes a system in which 41 specialized transformer models are connected to language model agents organized into primary care, specialty, and subspecialty tiers. Instead of a central supervisor deciding who speaks to whom, the agents collaborate directly through a shared case memory, making every referral and its justification explicit and auditable. The result, evaluated on 3,000 test cases drawn from real hospital records, substantially outperformed both a monolithic transformer and a standalone large language model.

The scale of the gap is striking. CPS-Net achieved top-1, top-3, and top-5 accuracies of 48.8 percent, 71.00 percent, and 77.76 percent, respectively, with a clinical relevance score of 94.17 percent. By comparison, a single monolithic transformer trained on all patients reached only 19.20, 38.60, and 47.20 percent on the same metrics, while a single large language model agent managed just 12.70, 26.40, and 37.23 percent. The study was led by Mohammad Assadi Shalmani, Masoud Khani, Michael S. Harris, Qiang Lu, and Jake Luo, spanning the University of Wisconsin-Milwaukee, the Medical College of Wisconsin, and the China University of Petroleum.

To build the system, the team drew on electronic health records from 75,000 patients treated at Froedtert Hospital in Wisconsin, covering 10 medical specialties and 30 subspecialties. Each patient’s record—demographics, diagnosis history, procedures, laboratory results, and medications—was converted into a chronological sequence of discrete tokens, each encoding both the type of clinical event and its specific medical code, paired with a time index marking days elapsed since the first recorded event. Patient sequences averaged 1,881 events, with a median of 725, giving the models long, temporally rich trajectories to learn from. The cohort was 56.2 percent female, with a mean age of 64.99 years; cardiology patients were the oldest on average at 74.96 years, while ear, nose, and throat patients were the youngest at 50.55.

The 41 transformers share a single decoder-only architecture: four layers of causal multi-head self-attention with eight heads each, a gated feed-forward network using the SwiGLU activation, and a hidden dimension of 512. Each token’s embedding combines three learnable components capturing what the event was, what kind of event it was, and when it occurred. For diagnosis tokens, the model additionally blends in embeddings of the code’s ancestors in the ICD-10 hierarchy, weighted by a learnable scaling factor—a mathematical way of teaching the model that a specific diagnosis also inherits properties from its broader disease category. Training used a next-token prediction objective with label smoothing, the AdamW optimizer, and a sliding window of 64 tokens with a four-token overlap to make long histories computationally tractable.

The agents themselves are deliberately lightweight. Every agent runs on the same base language model, DeepSeek-flash, with no fine-tuning and no specialty-specific training data; behavior differs only through assigned role, system prompt, and which transformer the agent can query. Specialty-level prediction is supplied entirely by the transformers. Agent judgment is bounded rather than open-ended: each agent sees the transformer’s ranked candidate codes with probabilities and retains the top-ranked code by default, departing only in two defined situations—when the record documents an active, progressing problem in a different organ system, prompting a cross-specialty redirection, or when a lower-ranked code is the exact variant the record actually uses, prompting a reordering within the subspecialty.

The consultation protocol enforces quality through simple rules. An agent must contribute new clinical information or reasoning before referring; the system tracks consultation history to prevent circular referrals; and every referral must name its target and state a clear clinical rationale. A case concludes when a subspecialty agent issues a final answer containing the primary diagnosis, a ranked differential, and an explanation tying the conclusion to the patient’s history. In a representative case described in the paper, a 67-year-old woman on dialysis with prior myocardial infarction and a 3,580-event record was routed from primary care to cardiology, then to an ischemic-cardiology subspecialist, which correctly identified atherosclerotic heart disease without angina as the leading diagnosis while flagging differential diagnoses and safety warnings along the way.

Comparisons against four baselines isolated the value of each design choice. A hierarchical transformer baseline using the same 41 models but no consultation reached 44.2, 52.23, and 52.67 percent in top-1, top-3, and top-5 accuracy—meaning agent consultation added 4.60, 18.77, and 25.10 percentage points respectively. Across the test set, agents intervened in 1,011 cases, or 33.7 percent, most often by re-ranking diagnosis codes at the subspecialty stage. The interventions promoted the correct code into the top prediction in 222 cases and displaced it in 84, a net gain of 138 cases. As the ranking window widened, benefits grew and harms shrank, indicating the agents corrected far more transformer errors than they introduced.

Decentralization itself proved important. When the researchers built a centrally orchestrated version of the same system, with a supervisor agent routing every message, an independent language model judge, GPT o1, strongly preferred the decentralized framework in blinded comparisons—choosing it for justification quality 71 percent of the time, clinical reasoning coherence 84 percent, and natural language interpretation 93 percent. The study also quantified an average 8 percent loss of information when messages passed through the central supervisor’s filtering. Performance varied by domain, with oncology-hematology achieving the highest top-5 accuracy at 95 percent, but even apparent errors carried signal: among top-1 predictions judged wrong for the immediate next diagnosis, 53.08 percent of predicted codes appeared later in the patient’s history, rising to 76.16 percent for top-5 predictions—suggesting the framework captures latent disease trajectories rather than merely failing to predict the next event.

The authors are candid about limitations. The system was trained and tested on data from a single tertiary care hospital, retrospective coded diagnoses served as ground truth, and practicing physicians have not yet evaluated whether the referral logic matches real clinical reasoning. Future work will pursue multi-site prospective validation, integration of imaging summaries, genomic data, and clinical notes, and human-in-the-loop mechanisms that let the system learn from clinician feedback. Still, the central lesson is likely to echo beyond this one hospital: artificial intelligence that mimics the division of labor in medicine—and shows its work at every handoff—may be both more accurate and more trustworthy than any single model, however large.

Subject of Research: A decentralized multi-agent transformer framework that simulates physician referral hierarchies for disease prediction from electronic health records.

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

Article References: Shalmani, M. A., Khani, M., Harris, M. S., Lu, Q., & Luo, J. (2026). CPS-Net (Collaborating Physicians in Silico Network): A Decentralized Multi-Agent Transformer Framework for Specialty-Aware Disease Prediction. Journal of Medical Systems, 50(1), Article 131. https://doi.org/10.1007/s10916-026-02460-8

Image Credits: AI Generated

DOI: 10.1007/s10916-026-02460-8

Keywords: artificial intelligence, multi-agent systems, transformers, electronic health records, disease prediction, large language models, clinical decision support, machine learning, diagnosis, predictive medicine, CPS-Net, Collaborating