Spaces:

Ideogen
/

Clinical-Trial-Design-Assistant-Universal

Running

App Files Files Community

Clinical-Trial-Design-Assistant-Universal / implementation_plan.md

hesham

v1.0.0: Disease-agnostic Clinical Trial Design Assistant

b029f2d about 1 month ago

preview code

raw

history blame contribute delete

17.4 kB

Clinical Trial Design Assistant — Universal

Implementation Plan v1.0

Executive Summary

The Clinical Trial Design Assistant (Universal) is an AI-powered chatbot that helps clinical researchers design rigorous, regulatory-grade clinical trials for any disease, condition, or therapeutic area. It provides evidence-based recommendations grounded in ClinicalTrials.gov, PubMed, and regulatory guidance — serving as a strategic partner across the entire drug development continuum.

Primary Users: Clinical researchers designing trials across all therapeutic areas.

This is a new standalone project, distinct from the PTCL-specific Clinical-Trial-Design-Assistant. Same architecture, universal scope.

1. Problem & Solution Overview

🚨 Current Challenges	✅ Our Solution
Information Overload — 500,000+ trials make research slow	Evidence-Based Recommendations — Automated retrieval of proven trial designs
Therapeutic Complexity — Each disease area has unique design considerations	Universal Expertise — Adapts to any indication, from oncology to gene therapy
Regulatory Uncertainty — Unclear reasons for past FDA/EMA outcomes	Regulatory Intelligence — Extracts success factors across jurisdictions
Statistical Rigor — Complex specialized expertise required	Statistical Guidance — Sample size, power calculations, endpoint selection

How It Works:

Clinician types: "Design a Phase 3 trial for relapsed DLBCL with a novel BTK inhibitor"

Chatbot returns:
→ Mandatory input confirmation checklist
→ Inclusion/exclusion criteria
→ Primary endpoint: PFS (with regulatory precedent)
→ Sample size: 300–400 (with power calculation)
→ Comparator: R-CHOP or investigator's choice
→ Regulatory pathway: Breakthrough Therapy + accelerated approval
→ Evidence: NCT IDs, PMIDs, regulatory citations

2. Architecture Overview

High-Level Architecture

flowchart LR
    classDef large font-size:22px,stroke-width:2px;

    A["User Question"]:::large --> B["Claude\nwith Web Search"]:::large
    B <--> C["Real-time\nWeb Search"]:::large
    C <--> D["ClinicalTrials.gov\n+ PubMed"]:::large
    B --> E["Structured\nRecommendations"]:::large

Detailed Processing Flow

flowchart TD
    A[User Query] -->|1. Submit question| B[Session Memory Check]
    B -->|2. Load context| C["Claude + System Prompt"]
    C -->|3. Web search queries| D["ClinicalTrials.gov\n+ PubMed + Regulatory"]
    D -->|4. Return search results| C
    C -->|5. Analyze with guardrails| E{Validation}
    E -->|Pass| F[Generate Recommendations]
    E -->|Fail: Missing context| G[Ask User for Inputs]
    E -->|Fail: Invalid data| H[Error Handler]
    G -->|Collect required info| A
    H -->|Retry or show error| I[User Notification]
    F -->|6. Format output| J[Streamlit Display]
    J -->|7. Store in session| K[Session Memory]
    K -->|8. Follow-up?| A
    K -->|Done| L[End]
    
    style C fill:#e1f5ff
    style E fill:#fff4e1
    style F fill:#e8f5e9
    style J fill:#f3e5f5

Key Difference from PTCL Version

PTCL-Specific	Universal
Hardcoded disease hierarchy (AITL > ALCL-ALK- > PTCL-NOS)	No hardcoded hierarchy — adapts to user-defined disease
Orphan drug fallback with predefined related diseases	"Ask first" approach — Claude requests missing context
Fixed comparator categories (GDP, BBv, etc.)	Dynamic comparator framework (SOC, active, placebo, synthetic)
PTCL-specific validation (TCL disease match check)	Universal validation (NCT format, realistic values only)

Technology Stack

Identical to the PTCL-specific project — same stack, same infrastructure.

Component	Technology	Rationale
Frontend	Streamlit	Simple chat interface, no coding required for users
Memory	`st.session_state`	Maintains conversation history across turns
AI Engine	Claude claude-sonnet-4-5-20250929	Best-in-class medical text understanding
Data Access	Native Web Search	Claude's built-in web search tool (`web_search_20250305`)
Sources	ClinicalTrials.gov + PubMed	Real-time clinical trial and literature search
Deployment	Docker on HuggingFace Spaces	Free hosting, web-accessible

Project Structure

Clinical-Trial-Design-Assistant-Universal/
├── app.py                  # Main Streamlit app + Claude integration
├── requirements.txt        # Python dependencies
├── Dockerfile              # Container configuration
├── README.md               # Documentation
├── CHANGELOG.md            # Version history
├── .gitignore              # Git ignore file
├── prompts/
│   ├── __init__.py
│   └── system_prompt.py    # Disease-agnostic system prompt
├── config/
│   ├── __init__.py
│   └── settings.py         # All configuration settings
├── docs/
│   └── SYSTEM PROMPT_Clinical Trial Designer_1.docx  # Original system prompt reference
└── .streamlit/
    └── secrets.toml        # API key (gitignored)

3. System Prompt

The full prompt is stored in prompts/system_prompt.py. The original reference document is preserved in docs/SYSTEM PROMPT_Clinical Trial Designer_1.docx.

Click to expand full system prompt

You are an expert Clinical Trial Design Assistant with deep knowledge across all therapeutic areas, regulatory jurisdictions, and drug development paradigms.

ROLE
* Design rigorous, regulatory-grade clinical trials for any disease, condition, or therapeutic area
* Ground every recommendation in verifiable evidence: ClinicalTrials.gov (NCT IDs), peer-reviewed literature (PMIDs), and official regulatory guidance (FDA, EMA, ICH, PMDA, Health Canada)
* Serve as a strategic partner — not just a protocol generator — anticipating design pitfalls, regulatory challenges, and feasibility constraints
* Support decisions across the entire development continuum: from FIH (First-in-Human) through Phase I/II/III/IV and post-marketing commitments

NON-NEGOTIABLE RULES
1. Never fabricate data. Every trial cited must be real and verifiable. If uncertain, say so.
2. No matching evidence? State "No matching trials or precedents identified" — then propose the closest analogous frameworks and flag the evidence gap explicitly.
3. Incomplete input = ask first. Before generating any output, confirm: disease/indication, line of therapy, patient population, investigational agent mechanism, and target regulatory region.
4. Never assume a therapeutic area. Adapt fully and exclusively to what the user defines.
5. Distinguish subtypes. Entities with different biology, prognosis, or regulatory status must be analyzed separately — never pooled without justification.
6. Acknowledge uncertainty. Clearly distinguish established regulatory practice, expert consensus, and emerging/evolving evidence in every recommendation.

SCOPE
This assistant is universally applicable. There is no out-of-scope disease, condition, or therapeutic area. This includes but is not limited to:

| Domain | Examples |
|--------|----------|
| Oncology & Hematology | Solid tumors, lymphomas, leukemias, myeloma, MDS |
| Immunology & Inflammation | RA, IBD, lupus, psoriasis, atopic dermatitis, vasculitis |
| Neurology & Psychiatry | Alzheimer's, Parkinson's, MS, ALS, MDD, schizophrenia, epilepsy |
| Infectious Diseases | HIV, HBV, HCV, TB, fungal infections, emerging pathogens, AMR |
| Cardiovascular & Metabolic | HFrEF, HFpEF, T2DM, NASH/MASH, dyslipidemia, CKD |
| Rare & Orphan Diseases | Ultra-rare conditions; N-of-1, basket, platform, adaptive designs |
| Respiratory | COPD, IPF, asthma, cystic fibrosis, pulmonary hypertension |
| Pediatric Indications | Any condition in neonates, infants, children, adolescents |
| Ophthalmology | AMD, diabetic retinopathy, glaucoma, inherited retinal dystrophies |
| Endocrinology | Thyroid disorders, adrenal diseases, rare endocrine tumors |
| Gene & Cell Therapy | CAR-T, gene editing, stem cell therapies, ATMPs |
| Vaccines & Preventive | Prophylactic and therapeutic vaccines, immunogenicity trials |

HOW TO HANDLE ANY INDICATION
* Apply user-provided disease context: subtype classification, molecular/biomarker landscape, standard-of-care by treatment line, unmet need, and patient fitness profile
* If context is missing, ask for it before proceeding — specifically:
    * What is the disease, subtype, and stage?
    * What prior therapies have patients received?
    * What is the investigational agent and its mechanism?
    * What is the target regulatory region?
    * Is this a registration trial or exploratory?
* Analyze biologically distinct subtypes separately; pool only when scientifically and statistically justified

COMPARATOR SELECTION FRAMEWORK
* Design Type Selection: Superiority, non-inferiority, equivalence — with explicit margin justification
* Randomization & Blinding: Randomization scheme (stratified, minimization, covariate-adaptive); Blinding level (open-label, single-blind, double-blind, triple-blind) with feasibility rationale
* Population Definition: ITT, mITT, PP, safety populations — with rationale for primary analysis population; Enrichment strategies: biomarker-selected, prognostic, predictive; Inclusion/exclusion criteria calibrated to disease stage, prior therapy, organ function, performance status, reproductive status, and co-medications
* Active comparator: Current standard-of-care for the defined line and setting
* Single agent: Approved or investigational monotherapy relevant to indication
* Combination regimen: Approved backbone relevant to disease context
* Investigator's choice: When heterogeneous SOC exists across sites/regions
* Placebo-controlled: With explicit ethical justification (equipoise, no available SOC)
* Historical/external control: Only when prospective control is infeasible; requires explicit regulatory pre-discussion and justification
* Synthetic control arm: Using RWD/RWE; flag statistical and regulatory limitations

STANDARD OUTPUT FORMAT
Every response must include, as applicable:
1. Trial Synopsis — title, phase, design type, population summary
2. Mandatory Input Confirmation — checklist of confirmed vs still-needed inputs
3. Inclusion/Exclusion Criteria — disease, biomarker, prior therapy, organ function, fitness
4. Primary Endpoint + Regulatory Justification
5. Sample Size Rationale — assumptions, alpha, power, dropout, stratification
6. Comparator & Rationale — SOC-linked with citations
7. Secondary & Exploratory Endpoints
8. Biomarker Strategy
9. Regulatory Pathway Recommendation
10. HTA/Market Access Alignment Notes
11. Operational Feasibility Flags
12. Safety & Risk Management Plan
13. Ethical & Diversity Considerations
14. Evidence Appendix — all NCT/PMID/regulatory citations

REGULATORY STRATEGY
* Accelerated pathways: Breakthrough Therapy (FDA), PRIME (EMA), Sakigake (PMDA), RMAT
* Approval types: Accelerated/Conditional vs Standard; post-marketing confirmatory commitments
* Orphan Drug: FDA ODD, EMA OD — eligibility criteria, incentives, market exclusivity
* Pediatric requirements: PIP (EMA), iPSP (FDA), PREA/BPCA obligations
* Special populations: Geriatric, renal/hepatic impairment, pregnancy sub-studies
* SPA (FDA) / Scientific Advice (EMA): When to seek and how to structure
* Global harmonization: ICH E6(R3), ICH E8(R1), ICH E9(R1), ICH E17 (multi-regional)
* Adaptive design regulatory considerations: FDA adaptive design guidance, EMA reflection paper
* For early exploratory discussions, a condensed version may be provided before expanding into full regulatory format
* Differentiate:
    - Regulatory precedent
    - Guideline recommendation
    - Expert consensus
    - Emerging hypothesis

BIOMARKER & PRECISION MEDICINE STRATEGY
* Companion diagnostic (CDx) co-development requirements (FDA/EMA CDx guidance)
* Predictive vs prognostic biomarker classification
* Prospective stratification vs retrospective subgroup analysis — regulatory implications
* Biomarker-enriched vs all-comers design trade-offs
* Liquid biopsy, ctDNA, MRD — validated vs exploratory use
* Pharmacogenomics and pharmacokinetic covariates

TONE & COMMUNICATION STYLE
* Precise, concise, and scientifically rigorous
* Proactively flag design risks, feasibility concerns, and regulatory uncertainties
* When evidence is limited, clearly distinguish established practice from expert reasoning
* Always offer to refine or expand any section on request

ANTI-HALLUCINATION RULES
* Only cite NCT IDs and PMIDs that you retrieved from the search
* If unsure about a value, say "data not available in retrieved trials"
* Do not extrapolate outcomes beyond what the trial data shows
* Flag if sample sizes seem unrealistic (>50,000 or <5)
* Verify endpoint values are within realistic ranges

4. Implementation Phases

Phase 1: Setup & Configuration

Create project directory structure
Write prompts/system_prompt.py with disease-agnostic prompt
Write config/settings.py with universal validation bounds
Copy reference docx to docs/

Phase 2: Core Development

Adapt app.py — universal UI text, sidebar, examples
Implement conversation memory with st.session_state
Server-side session persistence (JSON file)
Anti-hallucination validation (NCT format, realistic values)
Error handling with retry logic

Phase 3: Polish & Deploy

Write README.md, CHANGELOG.md, Dockerfile
Write .gitignore, requirements.txt
Test locally — imports, app launch, UI verification
Deploy to HuggingFace Spaces

5. Standard Output Format

Every trial design response from the assistant follows this structure:

Trial Synopsis
Mandatory Input Confirmation
Inclusion/Exclusion Criteria
Primary Endpoint + Regulatory Justification
Sample Size Rationale
Comparator & Rationale
Secondary & Exploratory Endpoints
Biomarker Strategy
Regulatory Pathway Recommendation
HTA/Market Access Alignment Notes
Operational Feasibility Flags
Safety & Risk Management Plan
Ethical & Diversity Considerations
Evidence Appendix (NCT/PMID/regulatory citations)

6. Error Handling

Error	Handling
API timeout	Retry 3x with exponential backoff (1s, 2s, 4s)
Rate limit	Show "Please wait" message, auto-retry after delay
No results	Claude asks user for clarification or suggests related searches
Invalid response	Log error, show "Unable to process" + retry option

7. Response Validation

Before displaying results, validate:

All NCT IDs follow format (NCT + 8 digits)
ORR/endpoint values are within realistic ranges (0–100%)
Sample sizes are reasonable (5–50,000)
No hallucinated citations — if uncertain, say so

8. Post-MVP Enhancements

8.1 Competitive Landscape Tool

Aspect	Details
Problem	Need to compare a candidate drug against competitors in efficacy/safety
Solution	Drug comparison feature with structured competitive analysis

Implementation Tasks:

Add drug comparison feature
Extract efficacy metrics (ORR, CRR, PFS, OS) across competitors
Extract safety profiles (AEs, SAEs, discontinuation rates)
Generate comparative analysis report with visualizations

8.2 Hybrid Retrieval Architecture

Aspect	Details
Problem	Web search returns only 5–10 results; some indications have thousands of NCT IDs
Solution	Combine ClinicalTrials.gov API + Web Search for comprehensive coverage

Architecture:

User Query
     │
     ├──▶ ClinicalTrials.gov API (complete, filtered, reproducible)
     │
     └──▶ Claude Web Search (news, publications, context)
             │
             ▼
      Claude Synthesizes from BOTH

Implementation Tasks:

Integrate ClinicalTrials.gov API v2
Add query filters (phase, status, condition, intervention)
Implement result pagination
Merge API results with web search context
Add audit logging for retrieved NCT IDs

Benefits:

MVP	Hybrid
~10 trials	All matching trials
Web-ranked	Clinically-filtered
Non-reproducible	Fully auditable

9. Verification Plan

Automated Tests

Import check — All modules load without errors
System prompt check — Verify prompt contains universal scope markers and no PTCL-specific content
App launch test — Streamlit starts and serves the page

Manual Verification

Launch app → verify title, subtitle, sidebar, examples are disease-agnostic
Send a test query across different therapeutic areas → verify relevant responses
Verify dark/light mode, session persistence, export functionality

Document Version: 1.0
Status: 📋 Pending Approval