NeuroForecast

A Governance-First Probabilistic Forecasting System for Parkinson’s Clinical Trials

NeuroForecast is a version-controlled probabilistic forecasting system for Phase 2 and Phase 3 Parkinson’s Disease clinical trials.

It is designed as a governance-first decision-support prototype for high-uncertainty healthcare environments.

The system prioritizes calibration discipline, explicit assumptions, version control, and outcome accountability over narrative confidence or model complexity.

The current implementation uses mechanism-stratified base rates, additive structural adjustments, explicit probability caps, and Brier score evaluation.

---

Problem

Clinical trials operate under extreme uncertainty.

Yet most decision systems — whether clinical, financial, or AI-driven — often prioritize:

• Confidence over calibration
• Narrative over base rates
• Complexity over governance

In healthcare, poor calibration is not just a modeling flaw — it is a safety risk.

NeuroForecast explores how to design a structured, auditable forecasting system where:

• Probabilities are locked and immutable
• Assumptions are explicitly defined
• Outcomes are objectively scored
• Model updates are version-controlled
• Errors are visible and measurable

---

What This Project Demonstrates

This repository intentionally showcases:

• Probabilistic reasoning under uncertainty
• Healthcare domain modeling
• Explicit assumption management
• Version-controlled model governance
• Deterministic evaluation via Brier scoring
• Transparent error visibility

Status & Scope

• Research / learning mode
• N = 23 predictions (10 resolved, 13 active)
• Not statistically validated
• Not investment advice
• Not regulatory decision support

This system demonstrates probabilistic reasoning discipline and governance-first AI design.
Statistical calibration assessment requires 50+ outcomes minimum.

Design Philosophy & Governance

NeuroForecast is built on four core governance principles.

1. Immutable Prediction Locking

Once a probability is assigned to a trial, it is never edited.

No retroactive changes.
No silent revisions.
No narrative adjustments.

Predictions are stored in:

data/locked_predictions.csv

This enforces auditability and eliminates hindsight bias.

---

Initial Calibration Validation

Before prospective locking began, the model was stress-tested using a historical backfill set of completed trials.

• 10+ completed Phase 2/3 trials evaluated
• Predictions assigned using model rules
• Outcomes scored using Brier metric
• No retroactive edits permitted

This validation phase surfaced systematic underestimation of Phase 3 symptomatic programs, leading to mechanism-specific baseline separation.

Full details available in:

CALIBRATION_SUMMARY.md

---

2. Explicit Model Versioning

All baselines and structural rules are defined in:

model/WEIGHTS.md
model/VERSION.md

Any structural modification requires:

• A version bump
• A calibration log entry
• A traceable Git commit

This mirrors governance standards used in regulated healthcare AI systems.

---

3. Deterministic Evaluation

Performance is measured using the Brier Score:

Brier = (p_locked − outcome)^2

Lower scores indicate better calibration.

Scoring is reproducible via:

scripts/compute_brier.py

There are no hidden adjustments.

---

4. Safety-Aware Outcome Classification

Trial outcomes are classified conservatively:

• Success = Primary endpoint met AND development continued
• Failure = Primary endpoint not met OR safety/futility termination

Safety signals are treated as failures because development cannot proceed.

---

Current Version (v1.0)

Focus area:

• Parkinson’s Disease
• Phase 2 and Phase 3 interventional trials
• Clinical efficacy primary endpoints
• Enrollment ≥ 30

Mechanism categories:

• Dopaminergic symptomatic
• Non-dopaminergic symptomatic
• Disease-modifying

Scope discipline preserves calibration integrity.

---

Model Structure (v1.0)

Phase 2 Baseline

0.15

Phase 3 Baselines

• Tier 3 Dopaminergic Symptomatic: 0.55
• Non-Dopaminergic Symptomatic: 0.30
• Disease-Modifying: 0.20

Additive adjustments apply for:

• Sample size
• Evidence tier
• Endpoint fragility
• Operational risk
• Mechanism-specific penalties

Probability constraints:

• Floor: 0.10
• Ceiling: 0.60

Full model specification:

model/WEIGHTS.md

---

Repository Structure

Directory structure: └── nickleko-neuroforecast/ ├── README.md ├── CALIBRATION_SUMARRY.md ├── CONTRIBUTING.md ├── MODEL_CARD.md ├── Data/ │ ├── Locked_Predictions.csv │ ├── monitor_log.csv │ └── outcomes.csv ├── docs/ │ └── GLOSSARY.md ├── model/ │ ├── CALIBRATION_LOG.md │ ├── VERSION.md │ └── WEIGHTS.md └── scripts/ ├── README.md └── compute_brier.py

---

What This Is Not

• Not financial advice
• Not an optimized trading system
• Not a machine learning black box
• Not curve-fitted historical modeling
• Not narrative-driven biotech speculation

This is a structured calibration system.

---

Relevance to Healthcare AI

Many healthcare AI failures stem from:

• Poor calibration
• Silent assumption drift
• Unversioned model updates
• Lack of outcome accountability

NeuroForecast embeds:

• Transparent assumptions
• Explicit version control
• Deterministic scoring
• Governance before optimization

These principles transfer directly to:

• Clinical decision support systems
• AI model oversight environments
• Safety-critical ML deployments

---

Future Roadmap

Planned expansions include:

• Logging market-implied probabilities
• Calibration curves and reliability diagrams
• Comparative forecasting (model vs. market)
• Broader neurodegenerative indications
• Structured monitoring dashboard

All expansions will follow versioned governance discipline.

---

Status

Research / personal-use mode.
Public-facing flagship project.
Prospective prediction locking active.
Model version: v1.0

The project prioritizes disciplined iteration over rapid expansion.