Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PRISM-Consult: A Panel-of-Experts Architecture for Clinician-Aligned Diagnosis

Lionel Levine, John Santerre, Alexander S. Young, T. Barry Levine, Francis Campion, Majid Sarrafzadeh

TL;DR

PRISM-Consult extends a compact ED decision-support backbone to a routed panel of domain specialists, dispatching episodes to Cardiac-Vascular, Pulmonary, Gastro-Oesophageal, Musculoskeletal, and Psychogenic experts based on early token signals. The approach preserves a shared tokenization template and embeddings, while training lightweight domain adapters (LoRA) and using a calibrated router to balance safety, coverage, and compute savings versus consult-all. Across five domains, specialists converge with low development perplexities near 2.0 and the router achieves high life-threat recall with substantial reductions in consulted experts, enabling auditable, real-time consult at scale. The work outlines data, calibration, and validation steps to support prospective deployment, and discusses plans for external validation, thresholding refinements, and expanded domain coverage to enhance clinical safety and scalability.

Abstract

We present PRISM-Consult, a clinician-aligned panel-of-experts architecture that extends the compact PRISM sequence model into a routed family of domain specialists. Episodes are tokenized as structured clinical events; a light-weight router reads the first few tokens and dispatches to specialist models (Cardiac-Vascular, Pulmonary, Gastro-Oesophageal, Musculoskeletal, Psychogenic). Each specialist inherits PRISM's small transformer backbone and token template, enabling parameter efficiency and interpretability. On real-world Emergency Department cohorts, specialists exhibit smooth convergence with low development perplexities across domains, while the router achieves high routing quality and large compute savings versus consult-all under a safety-first policy. We detail the data methodology (initial vs. conclusive ICD-9 families), routing thresholds and calibration, and report per-domain results to avoid dominance by common events. The framework provides a practical path to safe, auditable, and low-latency consult at scale, and we outline validation steps-external/temporal replication, asymmetric life-threat thresholds, and multi-label arbitration-to meet prospective clinical deployment standards.

PRISM-Consult: A Panel-of-Experts Architecture for Clinician-Aligned Diagnosis

TL;DR

PRISM-Consult extends a compact ED decision-support backbone to a routed panel of domain specialists, dispatching episodes to Cardiac-Vascular, Pulmonary, Gastro-Oesophageal, Musculoskeletal, and Psychogenic experts based on early token signals. The approach preserves a shared tokenization template and embeddings, while training lightweight domain adapters (LoRA) and using a calibrated router to balance safety, coverage, and compute savings versus consult-all. Across five domains, specialists converge with low development perplexities near 2.0 and the router achieves high life-threat recall with substantial reductions in consulted experts, enabling auditable, real-time consult at scale. The work outlines data, calibration, and validation steps to support prospective deployment, and discusses plans for external validation, thresholding refinements, and expanded domain coverage to enhance clinical safety and scalability.

Abstract

We present PRISM-Consult, a clinician-aligned panel-of-experts architecture that extends the compact PRISM sequence model into a routed family of domain specialists. Episodes are tokenized as structured clinical events; a light-weight router reads the first few tokens and dispatches to specialist models (Cardiac-Vascular, Pulmonary, Gastro-Oesophageal, Musculoskeletal, Psychogenic). Each specialist inherits PRISM's small transformer backbone and token template, enabling parameter efficiency and interpretability. On real-world Emergency Department cohorts, specialists exhibit smooth convergence with low development perplexities across domains, while the router achieves high routing quality and large compute savings versus consult-all under a safety-first policy. We detail the data methodology (initial vs. conclusive ICD-9 families), routing thresholds and calibration, and report per-domain results to avoid dominance by common events. The framework provides a practical path to safe, auditable, and low-latency consult at scale, and we outline validation steps-external/temporal replication, asymmetric life-threat thresholds, and multi-label arbitration-to meet prospective clinical deployment standards.

Paper Structure

This paper contains 59 sections, 4 equations, 6 figures.

Table of Contents

  1. Background
  2. Introduction
  3. Why a Panel of Experts? Tokenization Trade-offs and Clinical Coverage
  4. Related Work
  5. Clinical foundation models and domain adaptation
  6. Mixture-of-experts and routing among specialists
  7. Positioning PRISM-Consult
  8. Methods
  9. Overview.
  10. Data and Preprocessing
  11. Data Source
  12. Data structure and cohort window.
  13. Inclusion/exclusion Criteria.
  14. Initial (Pre-Diagnosis) ICD-9-CM Code Set — Router Inputs
  15. Final (Conclusive) ICD-9-CM Code Set — Specialist Gold Labels
  16. ...and 44 more sections

Figures (6)

  • Figure 1: Initial set of diagnoses, notionally tied to associated diagnostic grouping
  • Figure 2: Set of final 'Gold Label' diagnostic codes, representing an ultimate diagnostic determination
  • Figure 3: Development-set performance by specialist model. Perplexity is $\exp(\text{Val Loss})$. $\Delta$ Val Loss is the percentage reduction from epoch 1 to the best epoch. Scoped domains used capped training sets for balance.
  • Figure 4: Cross-specialist Training Loss over epochs for both training and validation cohorts
  • Figure 5: Light-Weight Router Training Cohort and Overall Results Summary
  • ...and 1 more figures