medsci-skills 4.1.0

package/skills/write-paper/references/journal_profiles/npj_Digital_Medicine.md

@@ -0,0 +1,93 @@
+# npj Digital Medicine
+
+## Journal Identity
+
+- **Full name**: npj Digital Medicine
+- **Abbreviation**: npj Digit. Med.
+- **Publisher**: Nature Portfolio (Springer Nature)
+- **ISSN**: 2398-6352 (online)
+- **Frequency**: Continuous (online only)
+- **Impact Factor**: High-impact digital health journal
+- **Open Access**: Yes (CC BY 4.0 or CC BY-NC-ND 4.0)
+- **APC**: ~EUR 2,490 (check current rates)
+
+## Manuscript Types and Word Limits
+
+| Type | Abstract | Manuscript Body | References | Notes |
+|------|----------|----------------|------------|-------|
+| Article | Structured | No strict limit (concise recommended) | No cap | Primary research |
+| Review | Structured | No strict limit | No cap | |
+| Brief Communication | Unstructured | Shorter than Article | No cap | |
+| Perspective | Unstructured | ~3,000 | No cap | |
+| Comment | Unstructured | ~1,500 | No cap | |
+| Matters Arising | Brief summary | ≤ 1,200 | 15 | No APC; on npj Digit. Med. papers only |
+
+## Abstract Format
+
+Structured with three headings:
+1. **Background**
+2. **Methods**
+3. **Results**
+
+Typical length 250-350 words (no explicit cap stated).
+
+## Keywords
+
+Not explicitly required in guidelines.
+
+## Required Sections
+
+Nature-style order (Methods after Discussion):
+1. **Introduction**
+2. **Results**
+3. **Discussion**
+4. **Methods** (placed after Discussion)
+5. **Data Availability Statement**
+6. **Code Availability Statement** (if custom code used)
+7. **Acknowledgements** (including funding)
+8. **Author Contributions** (CRediT recommended)
+9. **Competing Interests** (mandatory, even if none)
+10. **References**
+11. **Figure Legends** (after References)
+12. **Tables** (at end)
+
+## Citation Style
+
+- Nature referencing style (numbered, square brackets)
+- Numbered sequentially in order of appearance
+- More than 5 authors: first author only, then "et al."
+- Journal names: italicized, abbreviated with full stops
+- Volume numbers: **bold**
+- URLs: cited parenthetically in text, not in reference list
+- Example: `Schott, D. H., Collins, R. N. & Bretscher, A. Secretory vesicle transport velocity... *J. Cell Biol.* **156**, 35-39 (2002).`
+
+## Reporting Guidelines
+
+| Study Type | Required Checklist |
+|------------|-------------------|
+| Randomized trials | CONSORT |
+| Systematic reviews | PRISMA |
+| Observational studies | STROBE |
+| Diagnostic/prognostic | STARD, TRIPOD |
+| AI diagnostic accuracy | STARD-AI |
+| AI intervention trial | CONSORT-AI |
+| AI prediction model | TRIPOD+AI |
+| Qualitative research | COREQ or SRQR |
+| Case reports | CARE |
+
+Nature Portfolio Reporting Summary required with revised manuscript.
+
+## Special Notes
+
+- **Initial submission formatting is relaxed**; strict formatting required only at acceptance.
+- **Methods section after Discussion** (Nature family convention, not standard IMRAD).
+- **No Supplementary Methods allowed**: all methods must be in main manuscript.
+- **Single-blind peer review**; typical first decision ~4-6 weeks; desk reject rate ~50%+.
+- **ORCID required** for corresponding author before final submission.
+- **Data Availability and Code Availability** statements mandatory.
+- **LLM/AI policy**: LLMs cannot be authors; document use in Methods section.
+- **Figures**: 300 dpi minimum; panel labels lowercase bold (a, b, c); colorblind-accessible palettes required (avoid red/green).
+- **Supplementary Information**: single merged PDF; large tables as separate "Supplementary Data" files.
+- **Preprints** encouraged and do not compromise novelty.
+- **Manuscript transfer service**: rejected papers can transfer to other Nature Portfolio journals with referee reports.
+- **Statistics section mandatory**: must include test name, n, alpha level, one-tailed vs two-tailed, exact P values.

package/skills/write-paper/references/paper_types/ai_validation.md

@@ -0,0 +1,270 @@
+# AI Validation Study — Template for Radiology
+
+## Overview
+
+Template for studies that develop, validate, or compare AI/ML models for medical imaging tasks. Must follow at least one of STARD-AI, TRIPOD+AI, or CLAIM depending on the study type. This template extends the standard IMRAD structure with AI-specific requirements.
+
+---
+
+## Applicable Reporting Guidelines
+
+Select based on study type (multiple may apply):
+
+| Study Type | Primary Guideline | Also Consider |
+|------------|------------------|---------------|
+| Diagnostic accuracy of AI | STARD-AI | CLAIM |
+| AI prediction model (development) | TRIPOD+AI (Type 1a/1b) | CLAIM |
+| AI prediction model (validation) | TRIPOD+AI (Type 2a/2b/3) | CLAIM |
+| Any AI study in medical imaging | CLAIM 2024 | STARD-AI or TRIPOD+AI |
+| AI-assisted reader study | STARD-AI + reader study guidelines | CLAIM |
+| LLM/MLLM evaluation | CLAIM + custom LLM reporting | -- |
+
+---
+
+## Title
+
+- Include the AI/ML method, clinical task, and study type.
+- Good: "Validation of a Deep Learning Model for Automated Detection of Intracranial Hemorrhage on Noncontrast CT: A Multicenter Study"
+- Good: "Large Language Model Performance in Radiology Board-Style Questions: A Multi-Agent Validation Study"
+- Avoid: vague terms like "novel," "state-of-the-art," or "revolutionary."
+
+---
+
+## Abstract (Structured, 250 words)
+
+Follow the target journal's format. Include:
+- **Purpose**: Clinical problem + study objective.
+- **Materials and Methods**: Dataset size and source, model type, reference standard, primary metric.
+- **Results**: Primary performance metric with 95% CI, key comparison result.
+- **Conclusion**: Clinical implication, not just "the model performed well."
+
+---
+
+## Introduction (3-4 paragraphs)
+
+### Paragraph 1: Clinical Problem
+- The specific clinical task and its importance.
+- Current diagnostic performance, error rates, or workflow bottlenecks.
+- Why AI could help (workload, accuracy, speed, access).
+
+### Paragraph 2: Current AI Landscape
+- Prior AI approaches for this task (brief, cite key studies).
+- Gaps: limited validation, single-center, no reader comparison, no generalizability data.
+- For LLM studies: current evidence on LLM/MLLM performance in medical domains.
+- For NLP/LLM extraction studies: name 2-3 closest prior extraction or report-mining papers/systems
+  and state the substantive delta (task, dataset, input constraints, method, validation, or workflow),
+  not merely that the new system is "novel."
+
+### Paragraph 3: Study Objective
+- Precise aim with study design mentioned.
+- Hypothesis if applicable.
+- Template: "The purpose of this study was to {develop and validate / externally validate / compare} {model/system} for {task} using {dataset description}."
+
+---
+
+## Materials and Methods
+
+### 4.1 Study Design
+- Retrospective vs prospective.
+- Single-center vs multicenter.
+- Development vs validation vs both.
+- Dates of data collection.
+- If development+validation: describe the temporal or geographic separation.
+
+### 4.2 Dataset
+
+#### Source and Collection
+- Institution(s), imaging modality, date range.
+- How cases were identified (consecutive, sampled, queried).
+- Inclusion and exclusion criteria (be specific).
+
+#### Dataset Splitting (for development studies)
+- Training / validation / test split ratios and method.
+- **Critical**: confirm patient-level splitting (not image or exam level).
+- Temporal split preferred over random split (state which was used).
+- If external test set: describe source and independence.
+
+#### Dataset Characteristics
+- Total cases, positive/negative distribution.
+- Demographics: age, sex, relevant clinical variables.
+- Image characteristics: scanner models, protocols, acquisition parameters.
+- Case complexity distribution if relevant.
+
+### 4.3 Reference Standard
+- How ground truth was established (pathology, follow-up, expert consensus, clinical diagnosis).
+- Number of annotators, their experience level.
+- Annotation process: independent reading, consensus, adjudication.
+- Blinding: annotators blinded to model output and clinical information?
+- Inter-annotator agreement reported (kappa, ICC).
+- Timing between index test and reference standard.
+
+### 4.4 AI Model
+
+#### For Traditional DL Models
+- Architecture (e.g., ResNet-50, U-Net, YOLO v8) with citation.
+- Input: image size, preprocessing steps (normalization, augmentation).
+- Training: optimizer, learning rate (schedule), batch size, epochs, early stopping criteria.
+- Loss function.
+- Hardware: GPU model, training time.
+- Framework and version (PyTorch 2.x, TensorFlow 2.x).
+
+#### For LLM/MLLM Studies
+- Model name and version (e.g., GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro).
+- API access date range (models change over time).
+- Temperature and other generation parameters.
+- Prompt design: system prompt, few-shot examples, chain-of-thought.
+- Full prompts in supplement (or verbatim in methods if short).
+- If multi-agent: agent roles, interaction protocol, consensus mechanism.
+- Input format: text-only, image+text, structured data.
+- For report-text extraction: list each supplied text field separately (e.g., findings, impression,
+  clinical history, indication, prior diagnosis, referral text). State whether fields that may contain the
+  target label were masked, excluded, or evaluated in a sensitivity analysis.
+- For fine-tuning, LoRA, prompt-engineering, or multi-agent claims: include a same-backbone zero-shot or
+  few-shot comparator using the identical input fields, output schema, and test split.
+
+#### For Pipeline/System Studies
+- End-to-end system description.
+- Each stage: input, processing, output.
+- Decision points and thresholds.
+- Human-in-the-loop components.
+
+### 4.5 Comparators (if applicable)
+- Reader study design: number of readers, experience levels, subspecialty.
+- Reading conditions: blinding, clinical information provided, time constraints.
+- AI-assisted vs unassisted reading protocol.
+- Washout period between reading sessions.
+
+### 4.6 Statistical Analysis
+- **Primary metric**: AUC, sensitivity, specificity, accuracy, F1 -- state which and why.
+- **Confidence intervals**: 95% CIs for all primary metrics (bootstrap or exact binomial).
+- **Comparison tests**: DeLong test for AUC comparison, McNemar for sensitivity/specificity.
+- **Adaptation ablation**: when claiming benefit from fine-tuning, LoRA, prompt design, or a wrapper,
+  pre-specify the paired comparison against the same-backbone zero-shot/few-shot baseline and report the
+  uncertainty for the delta, not only the adapted model's standalone performance.
+- **Input-contamination sensitivity**: for NLP/LLM extraction, report the primary analysis with the
+  pre-specified input fields and, when applicable, a sensitivity analysis excluding fields likely to
+  contain the answer.
+- **Calibration**: calibration plot, Brier score, calibration slope/intercept (for prediction models).
+- **Subgroup analyses**: pre-specified subgroups with rationale.
+- **Multiple comparisons**: correction method if applicable.
+- **Sample size**: justification or power analysis if prospective.
+- **Software**: version numbers for all statistical packages.
+
+### 4.7 Ethics
+- IRB approval with protocol number.
+- Informed consent status.
+- Data de-identification method.
+
+---
+
+## Results
+
+> **Results = facts and numbers only.** No interpretation, no "why," no prior study
+> comparisons. See SKILL.md Phase 4 anti-interpretation guardrails.
+
+### 5.1 Dataset Characteristics
+- Flow diagram (Figure 1): screening, exclusions, final cohorts.
+- Table 1: demographics and baseline characteristics by split (train/val/test) or group.
+- Class distribution: positive/negative cases per split.
+
+### 5.2 Model Performance (Primary Endpoint)
+- Table 2: Primary performance metrics with 95% CIs.
+- ROC curve (Figure 2) with AUC and CI.
+- At the chosen operating point: sensitivity, specificity, PPV, NPV.
+- How the operating point was selected (Youden, clinical threshold, fixed sensitivity).
+
+### 5.3 Reader Comparison (if applicable)
+- Table 3: Reader performance vs AI performance.
+- Statistical comparison (DeLong, McNemar).
+- AI-assisted vs unassisted reader performance.
+- Reading time comparison if measured.
+
+### 5.4 Subgroup Analysis
+- Performance by key subgroups (age, sex, scanner, pathology subtype, difficulty).
+- Present as table or forest plot.
+- State the numbers; do NOT interpret why subgroups differ (save for Discussion).
+
+### 5.5 Error Analysis
+- Types of errors (false positives and false negatives).
+- Representative examples (Figure 3+): correctly classified, false positive, false negative.
+- Describe error patterns factually; do NOT speculate on causes (save for Discussion).
+
+### 5.6 Calibration (for prediction models)
+- Calibration plot (Figure).
+- Brier score, calibration slope, intercept.
+- Hosmer-Lemeshow or similar test.
+
+---
+
+## Discussion
+
+> **Before writing:** Collect anchor papers and user input via SKILL.md Phase 5a
+> interactive planning gate.
+
+### Paragraph 1: Summary
+- Restate key findings without repeating exact numbers.
+- Frame in context of the study aim.
+
+### Paragraphs 2-3: Performance in Context (anchor paper driven)
+- Organize around anchor papers provided by the user.
+- For each anchor paper: state their result → compare with ours → explain discrepancy.
+- Compare to prior models on the same task (cite specific AUC/accuracy from prior studies).
+- Compare to human reader performance (if applicable).
+- Explain performance differences: data, methodology, population.
+
+### Paragraph 4: Clinical Deployment Implications
+- Where in the workflow would this fit? (triage, second reader, standalone, quality assurance)
+- What is the intended use case?
+- What decision, workflow step, or downstream action changes if the model is correct? State this before
+  making a clinical-utility claim.
+- Implementation considerations: speed, hardware, integration.
+- Regulatory pathway considerations (if relevant).
+
+### Paragraph 5: Limitations
+Typical AI study limitations to address honestly:
+1. Retrospective design and selection bias.
+2. Single-center / limited generalizability.
+3. Reference standard limitations (imperfect ground truth).
+4. Dataset size and class imbalance.
+5. Temporal/geographic bias (training and test from same distribution).
+6. For LLMs: model version changes, non-reproducibility, API dependency.
+7. No prospective clinical validation.
+8. No assessment of clinical outcome impact.
+For each: (a) what it is, (b) how mitigated, (c) direction of residual bias.
+
+### Conclusion
+- One to two sentences: main finding and primary clinical implication.
+- Must be a citable statement — memorable and specific.
+
+---
+
+## Special Required Sections
+
+### Data Availability Statement
+- Template: "The datasets generated and/or analyzed during the current study are {available in the [repository name] repository, [URL] / not publicly available due to [reason] but are available from the corresponding author on reasonable request}."
+
+### Code Availability
+- Template: "The code for {model training / analysis / evaluation} is available at {URL}."
+- If not available: "Code is available from the corresponding author on reasonable request."
+
+### AI Disclosure
+- In Methods: "Large language model {name, version} was used for {specific task}. The model was accessed via {API/interface} between {dates}. Generation parameters were set to {temperature, etc.}."
+- In Acknowledgments: "{Tool name} was used for {writing assistance / code generation / data analysis}. All AI-generated content was reviewed and verified by the authors."
+
+---
+
+## Common Pitfalls to Avoid
+
+1. **Data leakage**: Preprocessing (normalization statistics) computed on full dataset before splitting.
+2. **Image-level splitting**: Splitting by image rather than by patient allows the same patient in train and test.
+3. **Missing CIs**: Reporting AUC = 0.95 without confidence interval is incomplete.
+4. **No calibration**: AUC measures discrimination only; calibration is essential for prediction models.
+5. **Overclaiming**: "Our model outperformed radiologists" when CIs overlap or test set is small.
+6. **Missing error analysis**: Knowing where the model fails is as important as overall metrics.
+7. **Vague prompts**: For LLM studies, not providing the exact prompts used.
+8. **No intended use**: Not specifying where in the clinical workflow the model would be deployed.
+9. **Cherry-picked operating point**: Choosing the threshold that looks best on the test set.
+10. **Ignoring prevalence**: Reporting PPV/NPV without discussing the study prevalence vs real-world prevalence.
+11. **Input-text contamination**: clinical history, indication, impression, or referral text contains the target diagnosis or label.
+12. **Missing same-backbone baseline**: fine-tuning or prompt-engineering benefit is claimed without a zero-shot/few-shot comparator on the same task.
+13. **Presence-only novelty**: the Introduction says the task is novel but does not position the work against the closest 2-3 prior papers or systems.

package/skills/write-paper/references/paper_types/animal_study.md

@@ -0,0 +1,194 @@
+# Paper Type: Animal Study
+
+## Overview
+
+- **Reporting guideline:** ARRIVE 2.0 (2020) — mandatory
+- **Ethical approval:** IACUC (Institutional Animal Care and Use Committee) approval number required
+- **Typical word count:** 3500–5000 words
+- **Structure:** Abstract → Introduction → Materials and Methods → Results → Discussion → Conclusions
+- **Key requirement:** 3Rs (Replacement, Reduction, Refinement) statement in Methods
+
+---
+
+## ARRIVE 2.0 — Essential 10 Items
+
+These 10 items are the minimum required for publication. **All must be present.**
+
+| # | Item | Where in Paper |
+|---|------|----------------|
+| 1 | Study design | Methods: Study Design |
+| 2 | Sample size | Methods: Sample Size |
+| 3 | Inclusion/exclusion criteria | Methods: Animals |
+| 4 | Randomization | Methods: Randomization |
+| 5 | Blinding | Methods: Blinding |
+| 6 | Outcome measures | Methods: Outcome Measures |
+| 7 | Statistical methods | Methods: Statistical Analysis |
+| 8 | Experimental animals | Methods: Animals |
+| 9 | Experimental procedures | Methods: Procedures |
+| 10 | Results | Results section |
+
+---
+
+## Structured Abstract (250 words)
+
+```
+Background: [Scientific rationale; why this animal model was used]
+Objective: [Specific aim and hypothesis]
+Methods: [Species, strain, sex, N per group, intervention, primary outcome,
+          statistical test]
+Results: [Primary outcome with effect size, 95% CI, and exact p-value]
+Conclusions: [Main finding and translational implication]
+```
+
+---
+
+## Introduction (400–600 words)
+
+1. **Clinical/biological problem:** Why does this research question matter?
+2. **Gap in knowledge:** What is unknown and why animal work is needed to address it.
+3. **Rationale for animal model:** Why was this species/model chosen? What aspects of the human condition does it recapitulate? Cite validation literature for the model.
+4. **Hypothesis and aims:** State explicitly. "We hypothesized that [X]. To test this, we [aim 1], [aim 2], [aim 3]."
+5. **3Rs statement:** "All procedures were performed in accordance with the 3Rs principles (Replacement, Reduction, Refinement)."
+
+---
+
+## Materials and Methods (1200–1800 words)
+
+### 2.1 Ethical Approval
+
+"All animal experiments were approved by the Institutional Animal Care and Use Committee of [Institution] (protocol number [XXXXX]) and conducted in accordance with the [National Institutes of Health Guide for the Care and Use of Laboratory Animals / Korean Animal Protection Act / relevant local regulation]."
+
+### 2.2 Animals (ARRIVE Essential Item 8)
+
+Report all of the following:
+- Species, strain, substrain (e.g., Sprague-Dawley rat, Charles River, Wilmington, MA)
+- Source (commercial vendor or in-house breeding)
+- Sex (justify if single sex: biological rationale required)
+- Age and weight at the start of the experiment (mean ± SD or range)
+- Health status (specific pathogen-free, germ-free, etc.)
+- Any genetic modification (full genotype for transgenic models)
+
+### 2.3 Housing and Husbandry (ARRIVE Recommended)
+
+- Housing conditions: cage type, group or individually housed, N per cage
+- Environmental conditions: temperature (°C), humidity (%), light cycle (e.g., 12h light/12h dark, lights on at 07:00)
+- Access to food and water (ad libitum or restricted; specify diet)
+- Acclimatization period before experiment
+- Any enrichment provided
+
+### 2.4 Study Design (ARRIVE Essential Item 1)
+
+State clearly:
+- Design type: parallel group, crossover, factorial
+- Control groups: sham, vehicle, naïve
+- Number of experimental groups and treatment allocations
+- Timeline (from baseline to final outcome measurement)
+
+Include a figure showing the experimental design timeline if complex.
+
+### 2.5 Sample Size Justification (ARRIVE Essential Item 2)
+
+**Required.** Calculate using G*Power or equivalent. Report:
+- Effect size expected (based on pilot data or literature)
+- α (0.05) and β (0.20, i.e., 80% power)
+- Formula used
+- Calculated N per group
+- Inflation for expected attrition (typically +15–20%)
+
+"Sample size was estimated using G*Power (version 3.1.9.7) based on a two-sample t-test with an expected effect size of Cohen's d = [X] (derived from [pilot data/[Author Year]]). With α = 0.05 and 80% power, [N] animals per group were required. Accounting for an estimated [15]% attrition, [N] animals per group were enrolled."
+
+### 2.6 Inclusion and Exclusion Criteria (ARRIVE Essential Item 3)
+
+State pre-specified criteria applied at the time of animal selection AND during the experiment. Common exclusion reasons: failure to reach surgical endpoint, infection, weight loss >20%, death before planned outcome measurement.
+
+"Animals were excluded if: (a) they failed to recover from anesthesia, (b) technical failure of the intervention occurred, or (c) the primary outcome could not be measured. All exclusions were recorded in the study log."
+
+### 2.7 Randomization (ARRIVE Essential Item 4)
+
+"Animals were randomly allocated to treatment groups using [computer-generated random numbers / block randomization / stratified randomization] with block size [N]. Allocation was performed by [person] who was not involved in [surgery/treatment/outcome assessment]."
+
+If randomization was not performed, justify.
+
+### 2.8 Blinding (ARRIVE Essential Item 5)
+
+"Outcome assessment was performed by [role] who was blinded to group allocation. [Describe specifically what was blinded at each stage: surgery, treatment, outcome assessment, statistical analysis.]"
+
+If full blinding was not possible, explain why and what steps were taken to minimize bias.
+
+### 2.9 Anesthesia and Procedures (ARRIVE Essential Item 9)
+
+- Anesthetic agent, dose, route (e.g., isoflurane 2–3% via inhalation, ketamine 80 mg/kg + xylazine 10 mg/kg i.p.)
+- Monitoring during anesthesia (temperature, heart rate, respiratory rate)
+- Analgesia (pre- and post-operative)
+- Humane endpoints: criteria for euthanasia
+- Specific surgical or experimental procedure: describe in enough detail for replication
+
+### 2.10 Outcome Measures (ARRIVE Essential Item 6)
+
+For each outcome:
+- Primary or secondary
+- Operational definition
+- Measurement method and equipment (manufacturer, location)
+- Timing of measurements
+- Observer reliability (if applicable: ICC, kappa)
+
+### 2.11 Statistical Analysis (ARRIVE Essential Item 7)
+
+- State the primary outcome and the test used for the primary analysis
+- Parametric vs. non-parametric decision (justify based on distribution)
+- Mixed-effects models for repeated measures (preferred over repeated ANOVA)
+- Correction for multiple comparisons (Bonferroni, FDR — state which and why)
+- Handling of missing data
+- Software: R (version X.X.X) or SPSS (version X.X); packages used
+- Significance threshold: α = 0.05 (two-tailed)
+
+---
+
+## Results (900–1200 words)
+
+### 3.1 Animal Attrition
+
+Report in a CONSORT-style flow diagram or table:
+- Animals enrolled
+- Animals excluded (reason)
+- Animals analyzed per group
+
+"Of [N] enrolled animals, [N] were excluded: [N] due to [reason 1], [N] due to [reason 2]. Final analysis included [N] animals per group."
+
+### 3.2 Baseline Characteristics
+
+Table comparing groups at baseline (weight, age, any baseline measurements). Include statistical comparison — groups should be comparable.
+
+### 3.3 Primary Outcome
+
+Report: mean ± SD (or median [IQR]) for each group. Effect size with 95% CI. Exact p-value.
+
+"[Primary outcome] was significantly [higher/lower] in [Group A] compared with [Group B] ([mean ± SD] vs. [mean ± SD]; difference, [X] [units]; 95% CI, [lower–upper]; P = [exact value])."
+
+### 3.4 Secondary Outcomes and Adverse Events
+
+Report all pre-specified secondary outcomes. **Always report adverse events** (infections, weight loss, unexpected deaths) even if none occurred: "No adverse events were observed during the study period."
+
+---
+
+## Discussion (600–800 words)
+
+1. Summary of main findings (with effect size)
+2. Mechanistic interpretation
+3. Comparison with previous animal studies of same model or intervention
+4. Translational relevance: how do findings relate to the human condition?
+5. Limitations: model limitations (does not fully recapitulate human disease), single sex (if applicable), limited follow-up, sample size
+6. Future directions
+
+---
+
+## Common Animal Study Pitfalls
+
+1. **Missing IACUC number** — always required; reviewers check this.
+2. **No sample size justification** — G*Power calculation is mandatory.
+3. **Not reporting sex** — increasing pressure to include both sexes or justify exclusion.
+4. **Confounding housing conditions** — different cages, litters, or technicians are confounders; report and control.
+5. **Outcome assessor not blinded** — blinding of outcome assessment is the most important bias control in animal studies.
+6. **Inconsistent or missing humane endpoints** — specify pre-determined criteria for euthanasia.
+7. **Reporting group means without individual data points** — show individual data (dot plots) for small N.
+8. **Using repeated ANOVA without accounting for sphericity** — use linear mixed-effects models instead.