Agentic Clinical Intelligence

01npj Health Systems · 2026

Eyal Klang et al.

Orchestrated multi agents sustain accuracy under clinical-scale workloads compared to a single agent

Read the publication↗

Multi-agent systemsOrchestrationClinical workloads

Multi-agent orchestration

Can agents preserve performance under clinical-scale workloads?

Three plots comparing multi-agent and single-agent accuracy, latency, and token use as batch size increases — **Figure 2**Accuracy, latency, and total token use across increasing task batches for the multi-agent and single-agent designs. Figure reproduced from the publication.

Question

Does lightweight orchestration outperform a single agent when simultaneous clinical tasks become numerous and heterogeneous?

Approach

Single-agent and orchestrated multi-agent designs completed retrieval, extraction, and dosing tasks in batches ranging from 5 to 80.

Finding

Across models, multi-agent accuracy remained 65.3% at 80 tasks versus 16.6% for a single agent, while using up to 65-fold fewer tokens.

Why it matters

At clinical scale, coordinating specialized agents with bounded responsibilities may matter more than building ever-larger prompts.

02JAMIA · 2024

Benjamin S. Glicksberg et al.

Evaluating the accuracy of a state-of-the-art large language model for prediction of admissions from the emergency room

Read the publication↗

Emergency medicineAdmission predictionRAG

Clinical prediction

Can an LLM predict emergency admission from real-world clinical data?

Receiver operating characteristic curves comparing machine learning and GPT-4 admission prediction strategies — **Figure 2**Receiver operating characteristic curves comparing conventional machine learning and multiple GPT-4 prompting and grounding strategies. Figure reproduced from the publication.

Question

How accurately can GPT-4 predict hospital admission from emergency-department encounters, and what clinical context improves it?

Approach

Using EHR data from seven New York City hospitals, GPT-4 strategies were compared with Bio-Clinical-BERT, XGBoost, and an ensemble model.

Finding

Naïve GPT-4 achieved 0.79 AUC. Adding relevant examples and machine-learning probabilities increased performance to approximately 0.87, close to the 0.88 ensemble.

Why it matters

General-purpose reasoning becomes more clinically useful when it is grounded in relevant examples and established predictive signals.

03Int. J. Medical Informatics · 2026

Eyal Klang et al.

Clinical agents fail silently on patient identity

Read the publication↗

Patient safetyIdentity integrityEHR agents

Patient safety

Will an agent notice that the record belongs to the wrong patient?

Study design showing MIMIC-IV records, identity tampering conditions, and evaluation of six agent models — **Figure 1**Study design using MIMIC-IV records, controlled identity perturbations, and six agent models completing approximately 1.2 million tool calls. Figure reproduced from the publication.

Question

Do clinical agents detect patient-identity inconsistencies before copying information into an electronic record?

Approach

Visits from 500 patients were subjected to header swaps, one-digit medical-record-number changes, and age shifts. Six models completed approximately 1.2 million tool calls.

Finding

The best model withheld writing in only 17.4% of obvious header swaps; detection of subtler MRN and age changes was near zero.

Why it matters

A correct clinical action attached to the wrong patient remains unsafe. Identity verification and abstention must be explicit system functions.

04PLOS Digital Health · 2026

Eyal Klang et al.

Large language models are poor clinical administrators: An evaluation of structured queries in real-world electronic health records

Read the publication↗

Structured EHRTool useCode execution

Tool-enabled analysis

Should a language model answer—or execute the analysis?

Bar chart showing mean coding-task accuracy across six language models — **Figure 5**Mean accuracy by model when the model generated executable code for structured EHR administration tasks. Figure reproduced from the publication.

Question

Can LLMs reliably count and filter structured EHR tables using direct answers, chain-of-thought reasoning, or executable code?

Approach

Nine models were tested across 32,950 queries and 25 table-size combinations drawn from 50,000 emergency-department visits.

Finding

Direct prompting was consistently weak and reasoning helped only modestly. Tool-generated and executed code substantially improved accuracy.

Why it matters

Language models should coordinate reliable computation rather than substitute fluent reasoning for arithmetic, filtering, or database operations.

05npj Digital Medicine · 2024

Eyal Klang et al.

A strategy for cost-effective large language model use at health system-scale

Read the publication↗

Query strategyCost efficiencyContext scaling

Health-system scale

How much clinical work can one query safely carry?

Four plots showing formatting failures, omission failures, and accuracy as the number of clinical notes increases — **Figure 5**Formatting failures, omissions, and accuracy as prompt size increases while total task count remains fixed at 50. Figure reproduced from the publication.

Question

Can clinical notes and questions be grouped to reduce cost without overwhelming the model or degrading useful performance?

Approach

Ten LLMs completed more than 300,000 experiments under progressively larger combinations of real-world clinical notes and questions.

Finding

High-capacity models effectively handled up to 50 grouped tasks after mitigable failures were addressed, with cost reductions reaching 17-fold.

Why it matters

At health-system scale, query design, failure handling, and context reuse become part of the model’s practical performance.

We study what happens when models enter clinical workflows.

Selected Works

Can agents preserve performance under clinical-scale workloads?

Can an LLM predict emergency admission from real-world clinical data?

Will an agent notice that the record belongs to the wrong patient?

Should a language model answer—or execute the analysis?

How much clinical work can one query safely carry?