PRISM

Appearance

PRISM Document Summaries ​

This document provides concise summaries of all PRISM documentation for quick reference and navigation. Each summary captures the core message and key takeaways to help you quickly identify which documents contain the information you're seeking.

How to Use This Guide ​

  • Scan section headers to find your area of interest
  • Read summaries to determine relevance before diving into full documents
  • Follow links to jump directly to detailed documentation
  • Cross-reference with the Table of Contents for section-level navigation

Introduction & Overview ​

PRISM Initiative Contact & Resources ​

Essential contact information, project resources, and founding team details. Provides links to code repositories, HuggingFace organization, email contact, and LinkedIn profiles for the three founders (Brian Jorden, Timothy Collins, and Ryan Herchig). Start here if you're looking to get involved, connect with the team, or access public resources.

PRISM Initiative: Executive Summary ​

Comprehensive overview of the entire PRISM initiative in approximately 3,000 words. Covers the problem PRISM solves, the technical solution, core innovations, business model, implementation strategy, safeguards, evidence base, collaborative framework, team composition, future vision, and investment opportunity. Read this first if you're new to PRISM and want a complete understanding of the initiative before diving into technical details.

Table of Contents ​

Hierarchical navigation hub with direct links to all documents and their major sections. Provides comprehensive section-level linking across the entire documentation collection. Use this for quick jumps to specific topics when you know what you're looking for but need the exact document location.

Foundation & Data Architecture ​

PRISM Data Format ​

Detailed specification of PRISM's eight-field data structure that transforms medical billing records into pattern recognition input. Explains POOL, PATIENT, ASL, and the Five Ws (WHO, WHAT, WHEN, WHERE, WHY), why markdown tables are used, and how the format enables pure sequence completion. Essential reading for anyone implementing PRISM or preparing data for processing.

Completely Anonymous Data Only ​

Explains PRISM's privacy-by-architecture approach using truly anonymous data rather than de-identified records. Covers the critical distinction between anonymization and de-identification, why MD5 hashing is used for verification, how complete anonymity enables better pattern recognition, computational efficiency benefits, and legal advantages. Crucial for understanding PRISM's privacy guarantees and regulatory positioning.

Insurance Company Unique Vantage Point ​

Explains why insurance companies are uniquely positioned to enable PRISM's pattern recognition capabilities. Covers cross-provider visibility, longitudinal patient journeys, healthcare delivery fragmentation, claims data completeness, and why hospitals or health systems can't replicate this comprehensive view. Important for stakeholders evaluating why PRISM partners with insurance companies rather than other healthcare entities.

Standardized and Finite Codesets ​

Comprehensive explanation of the medical coding systems PRISM leverages: ICD-10-CM, ICD-10-PCS, CPT, NDC, HCPCS, CMS place of service codes, and NUCC provider taxonomy. Details how these finite, documented coding systems create ideal input for language model pattern recognition and why LLM pre-training on medical codes provides significant advantages. Essential for understanding the semantic foundation of PRISM's pattern recognition.

Three-Pattern Learning ​

PRISM's foundational training methodology using GOOD (early detection), BAD (delayed diagnosis), and NOPE (alternative diagnosis) examples automatically extracted from historical data. Explains how TEST, EARLY, and LATE codes enable automated pattern extraction, demonstrates the approach with primary aldosteronism, and shows how the framework generalizes to any condition with clear diagnostic and treatment patterns. Core reading for understanding PRISM's training approach.

AI Architecture & Philosophy ​

Pure Sequence Completion Without Complexity ​

Explains PRISM's radical simplification approach: models trained exclusively to continue medical billing sequences without chat, reasoning, or complex prompting. Covers supervised fine-tuning as continued pre-training, intentional destruction of general capabilities, training to exhaustion on medical sequences, condition-specific nudges through BCO (Binary Classification Objective), and why simplicity works better than sophisticated AI techniques. Essential for understanding PRISM's deliberately "broken" models and why this design choice provides superior reliability.

Ensemble of 100 Specialized Models ​

Details PRISM's ensemble architecture with complete data isolation between 100 independent models. Explains the pool assignment methodology using last two digits, why exactly 100 models, how natural cross-validation emerges from independent training, consensus as evidence, and model diversity benefits. Critical for understanding how PRISM achieves robust pattern recognition through consensus rather than sophisticated individual model capabilities.

Retrospective-Prospective Validation ​

Describes PRISM's unique validation approach achieving perfect equivalence between retrospective and prospective analysis through truncation methodology. Explains how truncating historical patient data at specific points creates scenarios functionally equivalent to evaluating current patients, enabling comprehensive validation across millions of historical cases with known outcomes. Important for understanding PRISM's evidence base without requiring years of prospective studies.

"All Models are Wrong" Philosophy ​

Explores PRISM's embrace of statistician George Box's insight that "all models are wrong, but some are useful." Explains why being wrong 99% of the time can represent extraordinary success when the 1% of correct suggestions prevent serious complications, how this aligns with medical investigation reality, the value of rare disease detection, and the statistical framework supporting low positive predictive values. Essential for understanding PRISM's success metrics and philosophical foundation.

Constructive-Only Architecture ​

Explains how PRISM's architecture prevents care denial or restriction by design. Details how sequence completion can only generate additive suggestions never negative recommendations, why consensus mechanisms count positive suggestions only, what silence means when no consensus is reached, and how structural safeguards prevent misuse. Critical for understanding PRISM's ethical constraints and trust guarantees that cannot be reversed regardless of future ownership or business pressures.

Examples & Demonstrations ​

Synthetic Case Examples and Training Framework ​

Introduction to the GOOD, BAD, and NOPE synthetic examples demonstrating how primary aldosteronism patterns appear in PRISM's data format. Explains the visual framework, clarifies that SIGN markers are demonstration-only (not part of actual training), and outlines how the replicable framework applies to multiple conditions. Read this before diving into the specific example documents to understand the structure and purpose.

GOOD Example: Early Detection of Primary Aldosteronism ​

Detailed synthetic case showing optimal early detection where a patient receives appropriate screening after clear patterns emerge, leading to timely diagnosis and simple treatment that prevents complications. Demonstrates the healthcare utilization phenotype that PRISM learns to recognize, showing escalating blood pressure medications, metabolic abnormalities, and specialist referrals that trigger successful early intervention. Illustrates what PRISM aims to replicate at population scale.

BAD Example: Delayed Diagnosis of Primary Aldosteronism ​

Synthetic case illustrating the tragedy of delayed diagnosis where warning signs persist for years before complications force investigation. Shows the same patterns visible in the GOOD example but demonstrates what happens when screening opportunities are missed—emergency visits, medication escalation, organ damage, and eventual late-stage diagnosis requiring intensive treatment. Highlights the missed opportunities PRISM aims to prevent.

NOPE Example: Sleep Apnea Presenting as Resistant Hypertension ​

Synthetic case demonstrating differential diagnosis where resistant hypertension patterns lead to appropriate testing that reveals an alternative condition (sleep apnea) rather than primary aldosteronism. Shows how NOPE examples teach models the boundaries of pattern recognition, preventing overfitting by demonstrating that similar symptom patterns can have different underlying causes. Essential for understanding how PRISM learns specificity alongside sensitivity.

Sequence Generation Demonstration ​

Placeholder document for future demonstrations of how PRISM's models generate sequence continuations. Will show the actual model inference process, token-by-token generation, and how consensus emerges from multiple independent predictions.

Audio & Video Resources ​

Placeholder document for multimedia resources including presentations, demonstrations, and explanatory videos about PRISM's approach and capabilities.

Implementation & Operations ​

Consensus Voting Mechanism ​

Explains how PRISM aggregates 100 independent model predictions into screening suggestions through calibrated thresholds. Covers empirical threshold calibration using historical performance, condition-specific requirements, medical oversight integration, sensitivity versus specificity tradeoffs, and dynamic adjustment capabilities. Essential for understanding how PRISM ensures reliable suggestions through consensus rather than individual model outputs.

Clinical Decision Support Positioning ​

Details PRISM's positioning within FDA regulatory frameworks as a Clinical Decision Support (CDS) tool. Explains physician autonomy preservation, how physician-friendly explanations are generated to accompany suggestions, human-in-the-loop requirements, why communication goes to primary care physicians only, liability considerations, and integration with existing clinical workflows. Important for medical professionals and compliance teams evaluating PRISM's regulatory status.

Self-Aligning Incentive Structure ​

Describes PRISM's results-based compensation model where payment occurs only after documented successful early detection. Explains pre-agreed percentage of savings, PRISM-specific ICD tracking codes for outcome validation, patient cost coverage goals ensuring no financial barriers to screening, why fee-for-service would misalign incentives, and long-term sustainability. Critical for understanding PRISM's business model alignment with patient benefit.

Platform Potential with Constraints ​

Outlines PRISM's expansion potential beyond primary aldosteronism while maintaining ethical constraints. Explains focus on non-invasive tests, routine visit integration, condition selection criteria (clear diagnostic tests, significant early vs. late intervention differences, sufficient prevalence), rapid expansion capability through reusable infrastructure, and explicit limitations on what PRISM won't target. Important for understanding PRISM's scalability and boundaries.

Healthcare Utilization Phenotype Recognition ​

Explores how medical conditions create distinctive patterns in billing data as they develop—the healthcare utilization phenotypes PRISM recognizes. Covers how conditions express themselves through provider visits, procedures, medications, and diagnoses over time, temporal evolution of patterns, multi-provider manifestations, differences between observable billing phenotypes and clinical phenotypes, and research implications. Essential for understanding what PRISM "sees" in claims data.

Zero Integration Burden ​

Explains how PRISM integrates into existing insurance operations without requiring workflow changes from providers or patients. Covers use of existing data flows, no workflow changes required, existing communication channels, on-premises hardware management, remote administration model, and true background operation. Important for implementation teams evaluating deployment complexity and operational impact.

Infrastructure & Processing ​

Cluster Architecture Approach ​

Details PRISM's on-premises hardware infrastructure using clusters of consumer GPUs in modular server rack enclosures. Explains server rack design, consumer GPU strategy providing cost-effective processing power, network isolation for security, graceful degradation when hardware fails, and linear scaling properties enabling growth by adding enclosures. Essential reading for technical teams planning physical infrastructure and deployment.

Continuous Model Retraining ​

Describes PRISM's approach to maintaining current pattern recognition through continuous model updates. Explains the one-model-per-day retraining target, 100-day complete cycle ensuring each model refreshes quarterly, data freshness strategy, model replacement process, feedback incorporation from real-world outcomes, and quality assurance testing before deployment. Important for understanding how PRISM evolves and improves over time.

Continuous Batch Inference Process ​

Explains PRISM's background processing design that evaluates patients continuously without physician initiation. Covers patient prioritization logic using vector similarity to identify high-value screening candidates, no physician initiation required for pattern recognition, and computational efficiency through batch processing. Essential for understanding PRISM's operational workflow and how suggestions get generated.

Collaboration & Growth ​

Open Collaboration with Privacy Protection ​

Comprehensive explanation of PRISM's collaborative framework where insurance companies share trained models but never patient data. Covers model sharing mechanics, collective intelligence benefits, model merging process, secure contribution framework, early adopter advantages, dual contribution requirements (trained models plus technical improvements), code transparency with protected models, exponential early adopter benefits, implementation agreement framework, competitive dynamics transformation, and network effects. Essential reading for organizations considering participation in the collaborative network.

Organization & Research ​

Large and Rich Dataset ​

Quantifies PRISM's data scale with token calculations showing hundreds of billions of tokens per million patient population. Explains benefits of maximizing temporal context, rare pattern detection at population scale, gradual progression capture, statistical power advantages, data density around significant medical events, and semantic richness of medical codes. Important for understanding the data foundations enabling PRISM's pattern recognition capabilities.

PBC Structure and Team ​

Details PRISM's Public Benefit Corporation (PBC) legal structure and founding team composition. Explains PBC rationale legally enshrining patient benefit as primary objective, patient outcome priority, investment framework compatibility allowing traditional venture capital within mission-aligned structure, founding team expertise (Brian Jorden, Timothy Collins, Ryan Herchig), advisory relationships, and growth strategy. Essential for investors, partners, and stakeholders evaluating organizational structure and team capabilities.

Research and Publication ​

Outlines PRISM's commitment to peer-reviewed publication and academic collaboration. Covers peer review commitment for validation, pattern discovery potential revealing previously unknown disease progression relationships, academic credibility building, open research philosophy, contribution to medical knowledge beyond just screening suggestions, and collaborative opportunities with research institutions. Important for academic partners and researchers interested in PRISM's scientific contributions.

Technical Deep Dives ​

Vector Representations & Mathematical Foundations ​

Explores how PRISM transforms medical sequences into mathematical vector spaces for similarity analysis. Explains the geometry of medical similarity, trajectory analysis through time, mathematical similarity for patient prioritization, embedding space properties, dimensionality and information density, and practical implications for pattern recognition. Advanced technical content for data scientists and ML engineers working on PRISM's implementation.

Technology Stack & Core Libraries ​

Placeholder document for detailed specification of PRISM's complete technology stack including Python libraries (TRL, Transformers, Accelerate, DeepSpeed, Liger Kernel, Trackio, uv, Gradio), training frameworks, infrastructure components, and development tools. Will provide technical teams with the exact tools and versions used in PRISM's implementation.

Processing Pipeline Specifications ​

Detailed technical specification of PRISM's five-stage processing pipeline: patient selection and prioritization, parallel model inference across 100 models, consensus aggregation, explanation generation for physician-friendly narratives, and delivery to insurance company systems. Provides implementation teams with precise workflow specifications for replication and deployment.

Reference Architecture & Implementation Standards ​

Placeholder document for comprehensive reference architecture diagrams, implementation standards, best practices, and technical specifications. Will serve as the authoritative guide for teams implementing PRISM, covering system architecture, data flows, security standards, and integration patterns.

References & Resources ​

Glossary of Terms ​

Comprehensive glossary defining 80+ terms, acronyms, and technical concepts used throughout PRISM documentation. Organized into 10 categories: PRISM-Specific Terminology, Medical Coding & Healthcare Standards, Regulatory & Legal, AI Architecture & Training, Training Tools & Libraries, Technical & Data Formats, Dataset Formats, Medical Concepts & Conditions, Methodologies & Concepts. Each entry includes clear definitions, relevance to PRISM, and external resource links. Essential reference for readers encountering unfamiliar terms.

Document Organization ​

PRISM documentation uses a numbered prefix system (00-99) to organize content into thematic sections:

  • 00-09: Introduction & Overview
  • 10-19: Foundation & Data Architecture
  • 20-29: AI Architecture & Philosophy
  • 30-39: Examples & Demonstrations
  • 40-49: Implementation & Operations
  • 50-59: Infrastructure & Processing
  • 60-69: Collaboration & Growth
  • 70-79: Organization & Research
  • 85-88: Technical Deep Dives
  • 99: References & Resources

This numbering system provides clear ordering and makes it easy to identify which section a document belongs to while maintaining flexibility for future additions.

These summaries provide quick-reference overviews. For comprehensive details, click through to the full documentation. Start with the Executive Summary for a complete introduction, or use the Table of Contents for section-level navigation.