manuscript-writing

name	manuscript-writing
description	Full guide for writing a medical AI manuscript from scratch, covering all sections (Introduction, Methods, Results, Discussion, Abstract) with structure, style, and reporting standards. Use when drafting a new AI manuscript or when you need section-by-section writing instructions beyond TRIPOD-AI results reporting.

AI Manuscript Writing Guide

A section-by-section guide for writing medical AI manuscripts, from blank page to submission-ready draft.

---

General Style Rules

• No em dashes — use commas, parentheses, or separate sentences
• Write Results in past tense, Discussion in present tense
• Report all metrics with 95% confidence intervals
• No interpretation in Results — keep interpretation in Discussion only
• Maximum one acronym introduced per paragraph; spell out on first use
• Never use F1-score, recall, or accuracy as primary performance metrics

---

1. Title & Authors

Title formula: [Model/Approach] for [Task] using [Data Source]: [Design]

Example: "Deep Learning Detection of Left Ventricular Dysfunction from 12-Lead ECG: A Multicenter Validation Study"

• Keep under 15 words
• Avoid starting with "A" or "The"
• Include study design (Prospective, Multicenter, Randomized)

Authorship order:

1. First author: primary contributor to experiments + writing
2. Middle authors: data, methods, clinical input
3. Penultimate: senior statistician or domain expert
4. Last: senior/corresponding author (lab PI)

---

2. Abstract

Write this last, after all sections are complete.

Structured format (250–300 words):

Subsection	Content	Sentences
Background	Clinical problem + gap	2
Objective	What this study does	1
Methods	Design, cohort, model, metrics	3–4
Results	Primary outcome with metric + CI	2–3
Conclusion	Key clinical implication	1–2

Rules:

• No citations in abstract
• Include primary metric with 95% CI (e.g., AUROC 0.91, 95% CI 0.88–0.94)
• For classification models: include sensitivity and specificity at reported threshold
• State the threshold used (e.g., Youden Index, prespecified operating point)

---

3. Introduction / Background

Structure (4–5 paragraphs):

Paragraph 1 — Clinical Problem

• Open with the disease burden (incidence, mortality, cost)
• Quantify the diagnostic challenge (missed diagnoses, diagnostic delay, resource burden)
• No AI mention yet

Paragraph 2 — Current Standard of Care & Limitations

• Describe existing diagnostic approach and its limitations
• Cite reproducibility, access, cost, or inter-observer variability data

Paragraph 3 — AI in This Space (Literature)

• Trace AI approaches from traditional ML → deep learning
• Cite key prior works with their best-reported metrics
• Identify specific gaps: external validation, prospective design, subgroup fairness, real-world deployment

Paragraph 4 — What This Study Does

• State your contribution directly: "In this study, we developed and externally validated..."
• Each gap from paragraph 3 maps to one contribution here

Paragraph 5 — Roadmap (optional but recommended)

• "This paper is organized as follows..."
• Only include if journal allows or paper is complex

Forbidden in Introduction:

• Presenting results
• Vague gaps ("more work is needed")
• Overclaiming novelty without citation support

---

4. Methods

Required subsections, in this order:

4.1 Ethics Statement (FIRST, always)

This study was approved by [Institution] Institutional Review Board (IRB #XXXX).
Informed consent was [obtained / waived] given the [prospective/retrospective] design.
If large language models were used in manuscript preparation, disclose here.

4.2 Study Design & Data Sources

• Study type: retrospective / prospective / randomized
• Institutions involved, years of data collection
• Internal vs. external validation datasets

Flowchart: Include a CONSORT/PRISMA-style flowchart as Figure 1 showing:

• Total records screened
• Exclusions with reasons
• Final training / validation / test set sizes
• Event counts in each split

4.3 Inclusion / Exclusion Criteria

• List all criteria as bullet points
• Include age ranges, diagnostic codes, date ranges
• State who made eligibility decisions (automated vs. adjudicated)

4.4 Outcome Definition

• Primary outcome: exact definition, ICD codes, or measurement method
• Secondary outcomes if applicable
• Adjudication process

4.5 Model Development

For deep learning models:

Architecture: [name, version, input dimensions]
Pre-training: [ImageNet / ECG foundation model / from scratch]
Training: [optimizer, learning rate, batch size, epochs, early stopping criterion]
Augmentation: [list augmentations used]
Loss function: [cross-entropy, focal loss, etc.]
Hardware: [GPU model, memory, training time]

For traditional ML:

Features: [list or reference feature table]
Preprocessing: [normalization, imputation, scaling]
Model: [name, hyperparameter search strategy, CV folds]
Feature selection: [method used]

Threshold selection — REQUIRED for binary classification:

State explicitly how the operating threshold is chosen:

• Youden Index (maximizes sensitivity + specificity): recommended default
• Prespecified clinical threshold (e.g., sensitivity ≥ 90%): when clinical constraints drive the operating point
• Cost-sensitive threshold: when FP/FN costs differ clinically

Example: "The operating threshold was selected using the Youden Index (J = sensitivity + specificity − 1) on the internal validation set and applied without modification to all external validation cohorts."

4.6 Statistical Analysis

• Primary metric: AUROC with 95% CI (DeLong method)
• Secondary metrics: AUPRC, Sensitivity, Specificity, PPV, NPV at reported threshold
• Confidence intervals: bootstrap (1000 iterations) or DeLong
• Subgroup analyses: prespecified groups (age, sex, race, disease severity)
• Calibration: Brier score, calibration curves
• Comparison to baseline: likelihood ratio test or net reclassification improvement (NRI)
• Software: Python [version] / R [version] / SAS [version]

---

5. Results

Paragraph order must mirror Methods subsections exactly.

5.1 Cohort Description

• Total included, excluded (with reasons)
• Demographic breakdown (Table 1)
• Outcome prevalence in each split

5.2 Model Performance (Table 2)

Required metrics for classification models:

• AUROC (95% CI)
• AUPRC (95% CI)
• Sensitivity, Specificity, PPV, NPV — all at the reported threshold
• State the threshold explicitly in the table footnote

Table 2 footnote format: "Sensitivity, Specificity, PPV, and NPV calculated at the threshold maximizing the Youden Index (threshold = 0.XX on internal validation set)."

For regression models:

• Mean Absolute Error (MAE)
• Root Mean Square Error (RMSE)
• R² / concordance correlation coefficient

5.3 External Validation

• Repeat Table 2 metrics for each external cohort
• Note if threshold was re-derived or fixed from development set

5.4 Subgroup Analysis (Table 3)

• Performance stratified by key subgroups
• Forest plot recommended for visual summary

5.5 Sensitivity / Supplementary Analyses

• If applicable: performance under data augmentation, different thresholds, model ablation

Results writing rules:

• Past tense throughout
• No "why" statements — no interpretation
• Every table/figure cited in order
• Point estimate + 95% CI for every metric

---

6. Discussion

Structure: Answer 6 questions in order

Paragraph 1 — Main Finding (MANDATORY opening)

State the primary result and its clinical meaning in 2–3 sentences.

"We developed and externally validated a deep learning model for [task] achieving AUROC 0.91 (95% CI 0.88–0.94), demonstrating..."

Paragraph 2 — What Is Different from Prior Work

• Compare your key metrics to the best-cited prior results
• Explain why your approach differs: more data, prospective design, external validation, fairness

Paragraph 3 — What Is Similar to Prior Work

• Acknowledges concordant findings to reinforce validity
• Builds credibility rather than dismissing prior literature

Paragraph 4 — How Results Sections Correlate

• Synthesizes subgroup, calibration, and primary findings
• Draws internal consistency conclusions

Paragraph 5 — Clinical Implications for Practitioners

• Workflow integration: where does this model fit?
• Which patients benefit most?
• What human oversight is still needed?

Paragraph 6 — Research Implications

• What should the field study next?
• What methodological gaps remain?

Paragraph 7 — Limitations (MANDATORY)

Standard limitations to address:

• Retrospective design (if applicable)
• Single-center development (if applicable)
• Missing subgroup data
• Lack of prospective outcome data
• Threshold generalizability

Paragraph 8 — Conclusion (MANDATORY final paragraph)

• 2–4 sentences
• Restate main finding + clinical implication
• No new data or citations

---

7. Tables Reference

Table 1. Baseline Characteristics

Variable	Development (n=X)	External Validation (n=X)
Age, years, mean ± SD
Female sex, n (%)
[Primary outcome], n (%)

Abbreviations: SD = standard deviation.

Table 2. Model Performance

Metric	Internal Validation	External Validation
AUROC (95% CI)	0.91 (0.88–0.94)	0.89 (0.85–0.93)
AUPRC (95% CI)
Sensitivity
Specificity
PPV
NPV

Abbreviations: AUROC = Area under the receiver operating characteristic curve; AUPRC = Area under the precision-recall curve; PPV = Positive predictive value; NPV = Negative predictive value. Threshold: Sensitivity, Specificity, PPV, and NPV calculated at the Youden Index threshold (threshold = 0.XX on internal validation set).

Table 3. Subgroup Analysis

Report AUROC by: Age group, Sex, Disease severity (at minimum).

---

8. Figures Reference

Figure	Content	Required?
Figure 1	Patient/data flowchart	Always
Figure 2	ROC curve (with CI band)	Classification models
Figure 3	Calibration curve	Always
Figure 4	Subgroup forest plot	Recommended
Figure 5	Decision curve analysis	Recommended

---

9. Submission Checklist

• Title ≤ 15 words, includes study design
• Abstract written last, includes threshold statement for classification
• Ethics statement is the first Methods subsection
• LLM use disclosed if applicable
• Figure 1 is a flowchart with exact exclusion numbers
• Table 2 footnote explicitly states threshold and how it was selected
• All metrics reported with 95% CI
• No F1, recall, or accuracy as primary metrics
• Results contain no interpretation
• Discussion addresses all 6 questions
• Limitations section present
• Final paragraph is Conclusion only
• Code/model available on GitHub or HuggingFace
• TRIPOD+AI checklist completed (attach as supplement)
• No em dashes anywhere in manuscript