Clinical Paper Writing

When to use

Drafting a new manuscript for a medical/biomedical journal.
Restructuring sections to comply with reporting guidelines.
Self-review before submission.
Adapting a thesis chapter, technical report, or conference abstract into a publishable paper.
NOT a substitute for the actual scientific work — this guides structure and reporting, not the methodology itself.

Process

Phase 1 — Pre-writing decisions

Identify study type and select the matching reporting guideline:

| Study type | Guideline | Key items reviewers will check | |---|---|---| | Prediction model (DL/ML/radiomics) | TRIPOD+AI (2024) | Sample size justification, train/val/test split at patient level, calibration, fairness across subgroups | | Medical AI intervention/tool | CLAIM (2020, updated 2024) | Architecture, training data provenance, ground truth definition, error analysis | | Diagnostic accuracy | STARD 2015 | Reference standard, blinding, flow diagram of participants | | Observational cohort | STROBE | Setting, eligibility, variables, confounding, sensitivity analyses | | RCT with AI | CONSORT-AI | AI version, integration with workflow, human-AI interaction | | Systematic review | PRISMA 2020 | Flow diagram, risk of bias assessment | | Imaging biomarker | IBSI compliance + REMARK if prognostic | Feature reproducibility, validation cohort |

Confirm target journal: scope match, IF, average article length, structure (some merge Results+Discussion), format (Vancouver vs APA citations), supplementary policy.
Build outline mapping IMRaD sections to the reporting guideline checklist. Every checklist item must have a destination section.

Phase 2 — Section-by-section drafting

Title

Specific, declarative, includes study type ("...: a retrospective multicenter study", "...: a TRIPOD+AI compliant prediction model").
Avoid acronyms unless universally known.
12-15 words is a good target.

Abstract (250-300 words, structured)

Required sub-sections (adapt to journal's exact format):

Background: 1-2 sentences. Clinical problem + gap.
Methods: Design, setting, participants (n), key methods (one phrase: "deep learning model trained on..."), validation strategy, primary outcome.
Results: Cohort characteristics (key numbers), primary result with 95% CI, key secondary.
Conclusions: Clinical implication, NOT overstated. Mention what the study does NOT yet demonstrate.

Introduction (3-5 paragraphs)

Para 1: Clinical context, burden of disease, why it matters.
Para 2-3: State of the art. What's been tried, with citations. Limitations of existing approaches.
Para 4: Explicit gap statement. "Despite [previous work], no study has [specific gap]."
Para 5: Aim, hypothesis, contribution. End with one sentence preview of methods, NOT results.

Anti-pattern: introducing results in the introduction. Results live in Results.

Methods (most detailed; reproducibility is the standard)

Order these sub-sections explicitly:

Study design and setting — retrospective/prospective, single/multi-center, dates, ethics approval (CEIm code).
Participants/dataset — inclusion/exclusion criteria, recruitment, n at each stage. Provide a CONSORT-style flow diagram if any exclusions occurred.
Predictors/features — definition, measurement protocol, reproducibility (intra/inter-observer if applicable, ICC values).
Outcome — definition, ascertainment, blinding of outcome assessors.
Sample size — justification (power analysis for inferential, EPV ≥ 10-20 for prediction models).
Model development (if ML/DL):
- Architecture (cite or describe).
- Hyperparameters and how they were chosen.
- Training protocol (epochs, optimizer, lr, augmentation).
- Hardware (GPU model, framework + version).
Validation strategy — split (patient-level, NEVER image-level for medical imaging), internal vs external, cross-validation if used.
Statistical analysis — software + version, significance level, missing data handling, primary metric with CI estimation method (bootstrap with B replicates, percentile or BCa).

Anti-pattern: vague phrases like "standard preprocessing was applied". Be specific.

Results

Lead with descriptive: cohort characteristics in Table 1, stratified by relevant grouping (e.g., outcome group).
Primary outcome: point estimate + 95% CI. State the comparator explicitly.
Secondary outcomes: in order of importance, not order of significance.
Subgroup/sensitivity analyses: state pre-specification.
Discrimination AND calibration for prediction models (AUC alone is insufficient).
Figures: each should make a single point. Caption must be self-contained.

Anti-pattern: chasing p-values. Effect sizes with CIs are what matter.

Discussion (3-5 paragraphs, ~1000 words)

Para 1: Summary of main findings (in plain language, NO new results).
Para 2-3: Comparison with literature. Where does this fit? Why are results similar/different from prior work?
Para 4: Clinical implications. Calibrated, not overhyped. State the population to which results generalize.
Para 5: Limitations. Be honest, not strategic. "Single-center", "retrospective", "no external validation", "sample size limited for subgroup analysis".
Final paragraph: Conclusion — one sentence on contribution, one on future work.

Anti-pattern: "future work" used as a way to dismiss limitations. Limitations and future work are different.

Phase 3 — Self-review checklist (mandatory before declaring done)

[ ] Every claim in Abstract has support in Results.
[ ] Methods reproducible by another team without contacting authors.
[ ] No data leakage: patient-level split, no contamination between train/val/test.
[ ] CIs reported for all key metrics, not just point estimates.
[ ] Calibration reported (not just discrimination) for prediction models.
[ ] Reporting guideline checklist completed and submitted as supplementary.
[ ] Limitations explicit and proportionate to claims.
[ ] Author contributions, funding, COI, data/code availability statements present.
[ ] No "AI will replace radiologists" overclaiming.
[ ] Figures/tables referenced in order in text.
[ ] Citations: format matches journal, all DOIs present, no missing references.

Anti-patterns

| Pattern | Why it fails | Correction | |---|---|---| | "State-of-the-art" claim without CI comparison vs published baselines | Reviewers reject vague superlatives | Report metric ± 95% CI vs cited baseline numbers | | Image-level train/val/test split for patient-level outcome | Same patient in train+test inflates performance | Patient-level split, document explicitly with flow diagram | | Single-center validation claimed as "generalizable" | External validity unsupported | Acknowledge as limitation, propose multi-center external validation | | AUC as sole metric for imbalanced data | Misleads on minority class performance | Add F1, precision-recall AUC, calibration, decision curve analysis | | "Our model achieves [metric]" with no comparator | Performance has no context | Always compare to a baseline (clinical score, prior model, radiologist) | | Reporting only the best of multiple model variants | Multiplicity inflation | Pre-specify primary model; report all variants in supplementary | | "Future work" used to dismiss limitations | Reviewers see through it | Limitations = what this study cannot address; future work = next step | | Discussion that re-states results | Wastes word count | Discussion interprets, contextualizes, places in literature |

Verification gates

Before declaring a section complete, the agent must produce evidence:

Methods: Cite the reporting guideline checklist items covered in this section.
Results: Every number traces back to a specific analysis script/notebook (mention which).
Discussion: Every literature comparison has a citation; every clinical implication has a stated population scope.

Default style

English: academic, formal but not pompous. Sentences under 30 words where possible.
Tense: past tense for Methods/Results ("we trained"), present tense for established knowledge ("CT is the standard imaging modality").
Voice: active where it improves clarity ("we collected"), passive where the agent is irrelevant ("samples were processed").
Hedging: "may", "suggests", "is consistent with" for inference. Never for facts.
Citations: Vancouver numeric for most clinical journals (verify journal guide). Avoid citing review articles when primary sources exist.
Numbers: 95% CI in parentheses, not separate sentence. Decimal places matched to measurement precision (don't report Dice = 0.84231).

Resources to reference if needed

TRIPOD+AI 2024 checklist: https://www.tripod-statement.org/
CLAIM 2024: https://pubs.rsna.org/doi/10.1148/ryai.240300
STARD 2015: https://www.equator-network.org/reporting-guidelines/stard/
STROBE: https://www.strobe-statement.org/
IBSI for radiomics features: https://theibsi.github.io/