Consent Readability Rewriting for Clinical Trials

ISEF Category: Translational Medical Science

Subcategory: Other  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

Many adult consent forms read like legal puzzles, not plain English. That matters, because patients cannot agree to something they do not understand. You can ask whether a model can rewrite these forms without losing the facts that protect patients. If you do this well, you are working on a real problem in trial access and patient rights.

What Is It?

This project studies a simple idea with a hard twist, can a language model shorten informed-consent text while keeping every required idea? Informed consent is the document that explains a clinical trial, including risks, benefits, and what happens to the participant. Think of it like translating a rulebook into a version a younger reader can follow, without deleting any of the rules.

Your job is not just to make the text shorter. You also need to protect the meaning. That means the rewritten version must still include key elements such as purpose, procedures, risks, benefits, privacy, and the right to stop. You can then compare the original and rewritten versions with readability formulas and human scoring.

Why This Is a Good Topic

This is a strong science fair topic because it has a clear input, a clear output, and measurable results. You can test whether one model, prompt style, or fine-tuning method makes consent text easier to read without dropping required content. The project connects to patient understanding, research ethics, and health equity, which gives it real-world weight. You can learn text preprocessing, annotation, model evaluation, and how to judge tradeoffs between clarity and completeness.

Research Questions

• How does fine-tuning a small language model change the reading level of informed-consent text?
• What is the effect of different prompt styles on preserving FDA-required consent elements?
• Does adding a checklist of required elements improve content preservation scores?
• To what extent does model output improve readability without lowering mentor-rated clarity?
• Which consent sections are most likely to lose meaning during compression?
• How does a rule-based rewrite compare with a model-based rewrite for readability and fidelity?

Basic Materials

• Public informed-consent documents from ClinicalTrials.gov or university IRB repositories
• A laptop or desktop computer with internet access
• A spreadsheet program such as Google Sheets or Excel
• A plain-text editor such as Notepad, TextEdit, or VS Code
• A readability calculator or script for grade-level metrics
• A scoring rubric for blinded mentor review
• A small labeled dataset of consent passages and rewrite targets.

Advanced Materials

• A university or school research computer with GPU access
• Python environment with transformers, pandas, numpy, and scikit-learn
• A small open-source language model suitable for fine-tuning
• Annotation tool such as Label Studio for marking required consent elements
• Secure storage for protected text samples and annotations
• Readability metric libraries for automated scoring
• Statistical analysis tools for agreement and significance testing.

Software & Tools

• Python: Runs text cleaning, annotation checks, model training, and evaluation scripts for your rewrite pipeline.
• Google Sheets: Tracks document versions, scoring rubrics, and mentor ratings.
• Excel: Organizes sample texts, summary tables, and comparison charts.
• Readability formulas package: Calculates grade-level scores so you can compare original and rewritten text.
• Label Studio: Helps you tag consent elements and check whether rewrites preserve them.

Experiment Steps

1. Define the consent elements you must preserve, then turn them into a scoring checklist.
2. Collect a small set of public consent passages and split them into training, validation, and test groups.
3. Choose one rewrite strategy first, then decide how you will measure reading level and meaning retention.
4. Build a reference scoring system so a mentor can judge clarity, accuracy, and missing information.
5. Compare rewritten text against the original using both automated readability metrics and human review.
6. Test whether your model fails more often on certain sections, such as risks, privacy, or procedures.

Common Pitfalls

• Lowering the reading level by deleting required consent details, which makes the rewrite unusable for real review.
• Using readability scores alone, which can reward short text that still sounds confusing or incomplete.
• Training on consent text from one source only, which can make the model memorize style instead of learning simplification.
• Ignoring section structure, which can scramble the order of risks, benefits, and participant rights.
• Skipping blinded human scoring, which leaves you unable to tell whether the rewrite actually reads better to people.

What Makes This Competitive

A stronger project goes past simple simplification. You can compare multiple rewrite methods, test several consent subtypes, and measure both readability and factual retention with a clear rubric. Strong entries also report where the model fails, not just where it succeeds. If you add a careful error analysis, you show that you understand the real tradeoff between plain language and legal accuracy.

Project Variations

• Use pediatric assent forms instead of adult consent forms to test whether the same rewrite method works for younger readers.
• Compare a rule-based simplifier with a fine-tuned model to see which one preserves required elements better.
• Test whether adding patient-facing examples or glossaries improves readability without increasing omission errors.

Learn More

• ClinicalTrials.gov: Search for public protocol documents and consent-related materials to build a sample set.
• NIH Plain Language resources: Review plain-language guidance and examples through the NIH website.
• FDA informed consent guidance: Read the FDA guidance documents on consent content and required elements on the FDA site.
• PubMed: Search for review articles on informed consent comprehension, readability, and health literacy.
• MIT OpenCourseWare, Introduction to Machine Learning: Use course notes to review model training, evaluation, and overfitting.
• Label Studio documentation: Learn how to annotate text and track required elements in a structured way.