protocol-writing

name	protocol-writing
description	Guide for writing high-quality clinical research protocols for AI-based diagnostic studies, validation trials, and retrospective analyses. Use when drafting or reviewing a research protocol for AI in medicine.

Research Protocol Writing

Overview

This skill guides the generation of high-quality clinical research protocols for AI-based diagnostic studies, validation trials, and retrospective analyses.

Protocol Structure

1. Title Page

• Full protocol title (descriptive, includes study design)
• Protocol version and date
• Principal investigator(s) and affiliations
• Funding sources
• Ethics approval reference (if applicable)

2. Synopsis

One-page summary containing:

• Study title
• Objectives (primary, secondary, exploratory)
• Study design
• Population (inclusion/exclusion)
• Sample size
• Primary endpoint
• Timeline

3. Introduction

Structure:

1. Clinical context (2-3 sentences): Disease burden, current diagnostic standard
2. Knowledge gap (2-3 sentences): What is unknown or suboptimal
3. Proposed solution (1-2 sentences): How AI/technology addresses the gap
4. Study rationale (1-2 sentences): Why this study is needed now

Style:

• No headers within introduction
• Flow naturally from problem to solution
• Cite key statistics (prevalence, mortality, diagnostic accuracy)
• Avoid overpromising

4. Literature Review / Background

Structure:

1. Disease pathophysiology (brief)
2. Current diagnostic approaches and limitations
3. Prior AI/ML work in this domain
4. Foundation models or methods being leveraged
5. Knowledge gaps justifying the study

Style:

• Concise paragraphs, not exhaustive
• Focus on directly relevant prior work
• Identify specific gaps your study addresses
• 1-2 pages maximum

5. Objectives and Hypotheses

Format:

3.1 Primary Objective
To [verb] the [metric] of [intervention/model] for [outcome] using [comparator].

3.2 Secondary Objectives
- To develop/validate [model B, C, etc.]
- To assess performance stratified by [subgroups]
- To evaluate [additional endpoints]

3.3 Exploratory Objectives
- To investigate [mechanistic questions]
- To explore [hypothesis-generating analyses]

Rules:

• One primary objective only
• Objectives are measurable and time-bound
• Match each objective to evaluation criteria

6. Evaluation Criteria

Format as table:

Objective	Evaluation Criteria	Justification
Primary	AUROC, sensitivity, specificity, PPV, NPV	Standard diagnostic metrics
Secondary	Subgroup AUROC, calibration	Generalizability assessment

Standard metrics for diagnostic AI:

• AUROC, AUPRC
• Sensitivity at fixed specificity (80%, 90%)
• Specificity at fixed sensitivity (80%, 90%)
• PPV, NPV
• Calibration (observed vs. predicted)
• Net reclassification improvement (NRI)

7. Study Design and Population

7.1 Design Statement Single sentence: "This is a [retrospective/prospective] [cohort/validation/diagnostic accuracy] study..."

7.2 Population

• Source population and setting
• Timeframe
• Label definitions (positive, negative, extended controls)

7.3 Inclusion Criteria

• Numbered list
• Age, diagnosis, data availability requirements
• Be specific about imaging/ECG quality standards

7.4 Exclusion Criteria

• Conditions that confound the outcome
• Prior treatments affecting the index test
• Data quality exclusions

8. Variables and Data Sources

Format as table:

Category	Variable	Source	In Biobank?
Demographics	Age, sex	EHR	Yes
Clinical	BMI, comorbidities	Clinic DB	No

Categories:

• Demographics/Clinical
• Genetic
• ECG parameters
• Imaging (Echo, CMR)
• Outcomes/Labels

9. Outcomes

9.1 Primary Outcome

• Clear definition with thresholds
• Source of ascertainment
• Adjudication process (if applicable)

9.2 Secondary Outcomes

• List with definitions

9.3 Label Definitions (for AI studies)

Label	Definition
Positive	[Specific criteria]
Negative	[Specific criteria]
Extended Negative	[Controls without disease]

10. Methods

10.1 Data Collection

• Sources, extraction process, quality control
• Reference biobanks/registries used

10.2 Data Preprocessing

• Signal processing (filtering, normalization)
• Image preprocessing (resolution, cropping)
• Missing data handling

10.3 Modeling For each model, specify:

Model [X]: [Name]
- Architecture: [backbone, layers]
- Input: [data type, dimensions]
- Pretraining: [foundation model, if applicable]
- Fine-tuning strategy: [frozen layers, learning rate]
- Output: [prediction type]
- Loss function: [BCE, MSE, etc.]

Hyperparameter Table:

Parameter	Value	Rationale
Optimizer	AdamW	Weight decay regularization
Learning rate	1e-4	Conservative for fine-tuning
Batch size	64	Memory/stability tradeoff
Epochs	100	Early stopping, patience=15

Ablation Experiments Table:

Experiment	Description
A1	[Variation to test]
A2	[Variation to test]

11. Statistical Analysis

11.1 Sample Size

• Power calculation or justification for available sample
• Expected prevalence, desired precision

11.2 Primary Analysis

• Metrics computed with 95% CI (bootstrap or DeLong)
• Comparison to reference standard

11.3 Secondary Analyses

• Subgroup analyses (prespecified)
• Sensitivity analyses

11.4 Software

• Python (scikit-learn, PyTorch), R, SAS
• Version numbers

12. Training and Validation Strategy

Parameter	Value
Data split	70% train / 15% val / 15% test
Cross-validation	5-fold stratified
Stratification	By outcome and key confounders
Random seed	Fixed (e.g., 42)

13. Ethics and Data Governance

• IRB/REB approval status
• Consent requirements (or waiver justification)
• Data security measures
• Reference to data governance framework (e.g., Cadre de Gestion)

14. Timeline

Phase	Activities	Duration
1	Data preparation	Months 1-3
2	Model development	Months 4-8
3	Validation	Months 9-12
4	Prospective pilot	Months 13-18

15. Feasibility

Brief section addressing:

• Data availability and sample size adequacy
• Computational resources
• Team expertise
• Funding

16. References

Standard citation format (Vancouver or APA)

---

Style Guidelines

General:

• Use active voice
• Be specific and quantitative
• Avoid jargon; define abbreviations
• No em dashes; use commas or parentheses
• No excessive hedging

Tables over prose:

• Hyperparameters → table
• Variables → table
• Timeline → table
• Inclusion/exclusion → bulleted list

Scientific precision:

• Report thresholds (e.g., "≥70% stenosis" not "significant stenosis")
• Specify units
• Define all acronyms at first use

Formatting:

• Numbered sections (1, 1.1, 1.1.1)
• Bold for section headers
• Minimal use of italics
• No bullet points in flowing text; reserve for lists

---

Quick Reference: Section Lengths

Section	Target Length
Synopsis	1 page
Introduction	0.5-1 page
Background	1-2 pages
Objectives	0.5 page
Methods	3-5 pages
Statistical Analysis	1-2 pages
Ethics	0.5 page
Timeline/Feasibility	0.5 page each

---

Checklist Before Submission

• Primary objective is single and measurable
• Inclusion/exclusion criteria are specific
• Outcomes have clear definitions with thresholds
• Sample size justified
• All models have architecture and hyperparameters specified
• Ablation experiments defined
• Data sources and biobank coverage clarified
• Ethics/consent addressed
• Timeline realistic
• All abbreviations defined