← DiviScan GH-OS Home
Technical Validation Package

DiviScan GH-OS Validation Package

System architecture, validation methodology, and transparent capability disclosure for the decision-governance engine.

Version HyperCore v2.5.0 Date April 2026 Patent USPTO Provisional #63/706,309 Classification Pre-submission
Section 1

System Architecture — Honest Capability Map

DiviScan GH-OS operates a 9-layer decision-governance engine. Each layer contributes a distinct analytical mechanism. The table below discloses the exact mechanism type for every layer — distinguishing hand-coded rules, lookup databases, trained ML models, and LLM API calls. No mechanism is overstated.

Layer Component Mechanism Description
4a Clinical Rules Engine Rules 26 hand-coded clinical decision rules for threshold-based alerting
4b ClinVar Integration Lookup 209K variant→disease mappings from NCBI ClinVar database
4c MIMIC-IV ML Models Trained Model 5 disease classifiers with AUC-validated performance scores
4d ICD-10 Code System Lookup 97K diagnostic codes for standardized disease classification
4e Claude LLM Reasoning API Call Open-ended clinical reasoning across any disease via Anthropic Claude
4f HPO Phenotype Ontology Lookup 17,000+ phenotype→disease mappings from Human Phenotype Ontology
4g DisGeNET Lookup 1.1M gene→disease associations for genetic signal detection
4h Hetionet Lookup 2.2M biomedical edges spanning drugs, diseases, genes, and pathways
4i PrimeKG Lookup 8.1M biomedical edges for pathway-level disease reasoning

Transparency note: Only Layer 4c (MIMIC-IV ML Models) uses trained machine learning classifiers. All other layers operate through deterministic rules, database lookups, or LLM API calls. The system's primary capability is knowledge graph inference combined with LLM reasoning across 17,000+ disease phenotypes — not trained ML models on each disease.

Section 2

Validation Methodology — U-Metrics Framework

Traditional validation relies on AUC and sensitivity/specificity — metrics that measure model accuracy but not clinical value. A system can be highly accurate yet clinically useless if it surfaces information the clinician already knows. The u-metrics framework addresses this by evaluating clinical utility through a three-condition test that each alert must pass.

Condition 1
RIGHT

The positive identification is clinically correct. The signal corresponds to a true condition or risk present in the patient.

Condition 2
NEW

The information is clinically novel. It surfaces something the care team did not already know or had not yet identified.

Condition 3
CONVINCING

The alert is actionable by the recipient. Evidence and context are sufficient for a clinician to act without further investigation.

An alert that is RIGHT but not NEW adds noise. An alert that is RIGHT and NEW but not CONVINCING adds cognitive burden without enabling action. Only alerts satisfying all three conditions generate positive clinical utility. The u-metrics framework scores each alert on this three-condition basis, producing a utility score rather than a raw accuracy metric.

Deployment Mode-Specific Utility Scoring

Clinical utility depends on context. A false negative in critical care has different cost than in trial enrollment screening. The u-metrics framework applies different utility weight vectors depending on the deployment mode:

Hospital Mode

Optimized for early deterioration detection in acute care settings.

  • False negative cost: high (missed deterioration)
  • Novelty weight: high (alarm fatigue mitigation)
  • Convincingness threshold: high (clinician must act rapidly)
Pharma Mode

Optimized for trial rescue — identifying patients at risk of dropout, adverse events, or missed enrollment.

  • False negative cost: very high (lost trial subjects)
  • Novelty weight: medium (augments existing monitoring)
  • Convincingness threshold: medium (clinical review follows)
Government Mode

Optimized for population-level surveillance and resource allocation decisions.

  • False negative cost: medium (population-level tolerance)
  • Novelty weight: high (surveillance gaps are the target)
  • Convincingness threshold: medium (policy-level action)

Current status: The u-metrics framework is implemented in shadow mode. Utility weights are initialized with default values and are designed to be co-defined with clinical stakeholders during prospective validation. The framework provides the infrastructure for clinically meaningful evaluation — final weight calibration requires real-world clinical input.

Section 3

MIMIC-IV Trained Classifiers

Five diseases have dedicated machine learning classifiers trained and validated on MIMIC-IV clinical data. These classifiers produce probabilistic risk scores with measurable discriminative performance.

Disease AUC Score Training Data Status
Heart Failure 0.890 MIMIC-IV Validated
Acute Kidney Injury 0.887 MIMIC-IV Validated
Type 2 Diabetes 0.882 MIMIC-IV Validated
Sepsis 0.869 MIMIC-IV Validated
Anemia 0.815 MIMIC-IV Validated

Scope disclosure: 5 diseases have AUC-validated ML scores. All other disease associations are inferred through knowledge graphs and LLM reasoning, not trained models. The MIMIC-IV classifiers represent one layer of nine in the full decision-governance engine. Model expansion beyond these 5 diseases is underway but not yet deployed.

Section 4

Knowledge Base Architecture

The decision-governance engine queries across six federated biomedical knowledge bases, providing 11M+ structured relationships for graph-based inference. Each knowledge base serves a distinct role in the diagnostic reasoning pipeline.

11M+
Total Relationships
6
Integrated Knowledge Bases
17K+
Disease Phenotypes

Not yet integrated: OMIM (mendelian disease catalog), KEGG (metabolic pathways), Reactome (biological pathway data). Integration of additional knowledge bases is planned for future releases and will expand the inference surface for rare disease detection and metabolic pathway analysis.

Section 5

What We Can Honestly Claim

Transparent capability disclosure. Each claim is evaluated against the system's current operational state.

Claim Status Basis
Detects any known disease YES LLM reasoning + knowledge graph inference across 6 federated KBs
Covers 17,000+ disease phenotypes YES HPO phenotype → disease mappings (1,080,402 associations)
Probabilistic scores on 5 diseases YES MIMIC-IV trained classifiers with AUC 0.815–0.890
Trained on hundreds of diseases NO 5 trained classifiers. Expansion underway but not yet deployed.
Validated clinical utility PARTIAL U-metrics framework implemented; utility weights need clinical stakeholder input
FDA cleared or approved NO No regulatory submission filed. Pre-submission stage.
Prospective clinical results NO All validation is retrospective (MIMIC-IV). Prospective validation sought.
Section 6

Trial Rescue Simulation

Retrospective simulation on a 40-patient MIMIC-IV cohort modeling three pharmaceutical trial rescue capabilities: biomarker trajectory modeling, subpopulation responder identification, and endpoint risk detection.

$134,849,686
Recoverable trial value identified on 40-patient MIMIC-IV cohort
Simulation-backed estimate — not a prospective result
Capability Description
Biomarker Trajectory Modeling Continuous monitoring of biomarker trends to detect early deviation from expected therapeutic response curves, enabling intervention before protocol-defined failure thresholds are reached.
Subpopulation Responder Identification Graph-based analysis of patient phenotype clusters to identify subgroups with differential treatment response, supporting adaptive enrichment strategies and rescue protocol amendments.
Endpoint Risk Detection Multi-signal analysis of clinical and laboratory data to flag patients at elevated risk of primary endpoint failure, enabling targeted rescue interventions before data lock.

Important: The $134.8M figure is a retrospective simulation estimate based on MIMIC-IV data applied to published trial cost models. It represents potential recoverable value under modeled conditions — not realized savings. Prospective validation with real trial data is required to confirm these projections.

Section 7

Current Status & Transparency

Honest assessment of what is operational, what is in progress, and what has not yet been achieved.

Operational

Seeking

Not Yet Achieved