TLDR: PRISM-Consult is an AI system that uses a “panel-of-experts” architecture to improve diagnostic accuracy in emergency departments. It extends the PRISM sequence model by adding a lightweight router that dispatches patient cases to specialized AI models (e.g., Cardiac, Pulmonary, Gastro–Oesophageal, Musculoskeletal, Psychogenic) based on early symptoms. This approach ensures efficiency, interpretability, and high diagnostic recall across multiple organ systems, leading to significant compute savings and a safety-first policy compared to using a single generalist model.
In the fast-paced environment of the Emergency Department (ED), making quick and accurate diagnoses is crucial. A new research paper introduces PRISM-Consult, an innovative artificial intelligence (AI) architecture designed to enhance clinical decision support by acting as a ‘panel of experts’ for clinician-aligned diagnosis.
The Challenge of Early Diagnosis
Clinical care often involves understanding a sequence of events, from initial complaints to diagnostic orders and lab results, to predict a patient’s trajectory. The goal is to make reliable predictions early on, especially in the ED where time is critical and initial symptoms can be ambiguous. The original PRISM model, a compact sequence model, was developed to address this by tokenizing clinical events and predicting the next event. It proved efficient and interpretable, particularly for chest pain presentations, but its single-domain focus limited its applicability across the diverse range of conditions seen in an ED.
Introducing PRISM-Consult: A Panel of Experts
PRISM-Consult extends the foundational PRISM methodology by transforming it into a routed panel-of-experts system. Instead of a single model trying to cover all bases, PRISM-Consult employs a lightweight ‘router’ that reads the earliest clinical events (like symptoms and first diagnostic cues) and intelligently dispatches the patient’s case to one or more specialized AI models. These specialists are aligned with specific clinical organ systems, such as Cardiac–Vascular, Pulmonary, Gastro–Oesophageal, Musculoskeletal, and Psychogenic.
Each specialist model inherits PRISM’s efficient transformer backbone and token template, ensuring parameter efficiency and interpretability. This design allows each specialist to focus on a ‘local’ subset of tokens relevant to its domain, avoiding the need for a massive, complex global vocabulary that a single generalist model would require. This approach is akin to a clinical referral system: sending the ‘right expert’ to the ‘right case’ at the ‘right time’.
How It Works: Routing and Specialization
The system works by tokenizing episodes as structured clinical events. The router then analyzes the first few tokens to determine the most likely domain(s). For instance, if initial symptoms suggest a cardiac issue, the case is routed to the Cardiac–Vascular specialist. If a life-threatening domain is identified with high confidence, a single high-priority specialist is invoked. Otherwise, the top two relevant domains might be consulted in parallel, with their suggestions merged by a deterministic arbitration layer.
The data used for training PRISM-Consult comes from the MIMIC-IV database, a comprehensive electronic health records repository. Patient episodes are constructed as time-ordered sequences of events, including patient information, admission/discharge events, diagnostics, lab observations, and diagnoses encoded in ICD-9-CM. Crucially, procedures and medications are intentionally excluded to maintain a concise and clinically interpretable vocabulary focused on diagnostic reasoning.
Key Contributions and Benefits
The PRISM-Consult framework offers several significant contributions:
- Clinical Routing Surface: It formalizes a method for routing based on initial ICD-9 symptom codes to specialist targets defined by conclusive diagnostic families.
- Parameter Efficiency: It uses a minimalist router and parameter-efficient specialist adapters, maintaining PRISM’s compact footprint while expanding clinical coverage.
- Cross-Domain Accuracy: The system demonstrates smooth convergence and low development perplexities across various domains, validating that the PRISM backbone generalizes effectively when routed to domain-consistent data.
- Compute Savings: The routing mechanism leads to substantial compute savings compared to a ‘consult-all’ approach, where every specialist would process every case.
- Safety-First Policy: The system incorporates safety measures, such as failing open to all specialists if confidence is uniformly low or if vital signs cross danger thresholds, ensuring robust operation in critical scenarios.
Promising Results
On real-world Emergency Department cohorts, the specialist models showed stable convergence and low perplexities, indicating strong learning dynamics across diverse organ systems. The router achieved high routing quality, with near-perfect recall for identifying any correct domain and high recall for including all correct domains. This suggests that early clinical tokens carry enough discriminative signal for reliable specialist selection, even under strict latency constraints.
Also Read:
- MedMMV: Enhancing Trust and Accuracy in AI for Clinical Decisions
- K-Prism: A Unified AI Model for Versatile Medical Image Segmentation
Future Directions
The researchers outline future steps, including external and temporal validation on different sites and data slices to assess drift, refining safety-first routing policies with asymmetric thresholds for life-threatening conditions, and expanding coverage by adding more specialists. The framework provides a practical path to safe, auditable, and low-latency clinical decision support at scale.
PRISM-Consult represents a significant step forward in leveraging AI for healthcare, offering a model that is not only efficient and accurate but also interpretable and aligned with clinical workflows. For more details, you can refer to the full research paper: PRISM-Consult: A Panel-of-Experts Architecture for Clinician-Aligned Diagnosis.
