medsci-skills 4.1.0

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

@@ -0,0 +1,127 @@
+# Paper Type: Letter to the Editor
+
+## Overview
+
+- **Purpose:** Concise scholarly comment on a recently published article, or a brief original observation
+- **Word count:** 400–600 words (some journals allow up to 750)
+- **Structure:** No abstract, no section headers, 3–5 references maximum
+- **Key rule:** ONE main point. A letter with two main points is a letter that should not be published.
+
+---
+
+## When to Write a Letter
+
+1. **Comment on a published article:** Methodological concern, missing analysis, alternative interpretation, replication in a different population
+2. **Correspondence on your own article:** Response to a letter about your published paper
+3. **Brief original observation:** Too small for a full paper but notable; single case or very small series
+
+---
+
+## Standard Letter Format
+
+```
+[Journal Name]
+[Date]
+
+To the Editor:
+
+[Opening sentence: reference the article being commented on, with citation]
+
+[Body: ONE main point, supported by evidence, analysis, or data]
+[If applicable: additional supporting data, calculation, or comparison]
+
+[Closing: what the main point means — brief, direct]
+
+[Optional: acknowledgment of limitations of your own point]
+
+Sincerely,
+[Name, degree]
+[Institution, City, Country]
+[Email]
+[On behalf of co-authors if any]
+
+References
+[Maximum 5 citations]
+```
+
+---
+
+## Opening Sentence Templates
+
+**Commenting on methodology:**
+"We read with interest the article by [Author] et al. [citation] reporting [brief description]. However, we wish to raise a concern regarding [specific methodological point]."
+
+**Providing additional data:**
+"We read with interest the study by [Author] et al. [citation]. Our experience with [N] patients at [Institution] provides additional perspective on this topic."
+
+**Alternative interpretation:**
+"The study by [Author] et al. [citation] makes an important contribution. We would like to offer an alternative interpretation of the reported findings."
+
+**Replication:**
+"We have recently conducted an analysis of [similar cohort/dataset] that provides independent support for (or challenges) the findings of [Author] et al. [citation]."
+
+---
+
+## Body (200–350 words)
+
+Make your ONE main point clearly. Structure:
+1. State the specific concern or observation (1–2 sentences)
+2. Provide supporting evidence: cite literature, provide a calculation, reference an existing guideline, or briefly describe your own data
+3. Explain the clinical or scientific significance
+
+**If pointing out a methodological issue:**
+- Be specific: cite the exact figure/table/section where the issue appears
+- Quantify the impact if possible ("This could overestimate sensitivity by approximately X percentage points")
+- Propose what analysis should have been done
+
+**If providing additional data:**
+- State N, population, and key result with CI
+- Keep statistics minimal — no full table, just the key number
+
+---
+
+## Closing (50–100 words)
+
+"In summary, [restate main point in one sentence]. We believe [clinical or methodological implication]. [Optional: call for future research or transparency]."
+
+---
+
+## Response Letter Format (When Replying to a Letter About Your Paper)
+
+If someone writes a letter about your article and the journal asks you to respond:
+
+```
+To the Editor:
+
+We thank [Author] et al. for their interest in our study [citation] and for
+raising the issue of [topic].
+
+[Address each point they raised, one by one. Be concise and professional.
+Do not dismiss their concerns; acknowledge what is valid, then explain
+your position.]
+
+[If they are correct: "We agree that [point]. This limitation was
+acknowledged in our Discussion [citation, page X]. We are planning
+[future work] to address this."]
+
+[If you disagree: "We respectfully disagree with [point] for the
+following reasons: [evidence]. [Author Year] demonstrated that
+[supporting evidence]."]
+
+We thank [Author] et al. for contributing to the discussion of
+[topic] in the field.
+
+Sincerely,
+[Name, degree]
+[Institution]
+```
+
+---
+
+## Common Letter Pitfalls
+
+1. **More than one main point** — editors reject letters that try to make multiple arguments. Submit separately or pick the strongest point.
+2. **Too long** — strictly stay within the word limit (check each journal); editors cut letters without revision.
+3. **Personal or aggressive tone** — always professional and respectful, even when pointing out errors.
+4. **Citing too many references** — maximum 5; focus on the 1–2 most directly relevant.
+5. **Not referencing the specific article** — the opening must cite the article you are commenting on with the full journal citation.

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

@@ -0,0 +1,181 @@
+# Paper Type: Systematic Review and Meta-Analysis
+
+## Overview
+
+- **Reporting guideline:** PRISMA 2020 (mandatory)
+- **Quality assessment:** QUADAS-2 (diagnostic), RoB 2 (RCTs), MINORS (non-randomized surgical studies)
+- **Registration:** PROSPERO required before data extraction (cite protocol)
+- **Typical word count:** 4500–6000 words
+- **Structure:** Abstract → Introduction → Methods → Results → Discussion → Conclusions
+
+---
+
+## Required at Phase 0 (Init)
+
+Before writing, confirm:
+1. PROSPERO registration number (or justify non-registration)
+2. Search databases (PubMed, Embase, Cochrane, Web of Science minimum)
+3. Search date and date of last update
+4. Effect measure (OR, RR, HR, MD, SMD, AUC, sensitivity/specificity)
+5. Planned subgroup analyses (pre-specified in protocol)
+
+---
+
+## Structured Abstract (250 words)
+
+```
+Background: [One sentence: why this systematic review is needed]
+Purpose: [PICO: Population, Intervention/Index test, Comparator, Outcome]
+Methods: [Databases searched, dates, study design, quality tool]
+Results: [N studies, N patients, pooled estimate with 95% CI, 95% PI when I² > 50%, I²]
+Conclusion: [Main finding, certainty of evidence]
+```
+
+---
+
+## Introduction (400–600 words)
+
+1. **Background:** Why is this question clinically important? What is the burden of disease or diagnostic challenge?
+2. **Knowledge gap:** What do existing studies show, and why is a synthesis needed? (Reference 2–4 key studies or previous meta-analyses)
+3. **Objective:** State the PICO question explicitly.
+   - "The purpose of this systematic review and meta-analysis was to evaluate the diagnostic accuracy of [index test] for [condition] in [population]."
+4. **PROSPERO registration:** "This review was prospectively registered with PROSPERO (CRD[number])."
+
+---
+
+## Methods (1200–1800 words)
+
+### 2.1 Protocol and Registration
+State PROSPERO number. Confirm adherence to PRISMA 2020.
+
+### 2.2 Search Strategy
+List all databases with dates. Provide full search string for at least one database (or supplementary). Include grey literature (ClinicalTrials.gov, conference abstracts if applicable).
+
+**Example:**
+"MEDLINE (via PubMed) was searched from inception to [date] using the following strategy: ([MeSH term 1] OR [free text term]) AND ([MeSH term 2] OR [free text term]) AND ([MeSH term 3]). The complete search strategies for all databases are provided in Supplementary Table 1."
+
+### 2.3 Eligibility Criteria
+
+**Inclusion criteria table:**
+
+| Criterion | Specification |
+|-----------|---------------|
+| Study design | [Prospective/retrospective, cross-sectional, RCT, etc.] |
+| Population | [Patient characteristics, age, clinical setting] |
+| Index test | [Specific test/intervention] |
+| Reference standard | [Gold standard or comparator] |
+| Outcomes | [Primary and secondary] |
+| Language | English (and specify others if included) |
+| Time period | [If applicable] |
+
+**Exclusion criteria:** [List explicitly, not just "did not meet inclusion criteria"]
+
+### 2.4 Study Selection
+"Two reviewers (initials) independently screened titles/abstracts, then full texts. Inter-reviewer agreement was [% agreement / Cohen's kappa = X.XX] at the title/abstract stage and [% agreement / Cohen's kappa = X.XX] at the full-text stage. [N] discrepancies were resolved by consensus or a third reviewer (initials). The screening process is summarized in the PRISMA 2020 flow diagram (Figure 1)."
+
+### 2.5 Data Extraction
+"Two reviewers independently extracted data using a standardized form. The following variables were extracted: [list 8–15 variables]. Disagreements were resolved by discussion."
+
+**Data extraction table template (Supplementary Table):**
+
+| Study | Year | Country | Design | N | Age (mean±SD) | [Variable 1] | [Variable 2] | Outcome measure | Follow-up |
+|-------|------|---------|--------|---|--------------|-------------|-------------|-----------------|-----------|
+
+### 2.6 Quality Assessment
+Specify which tool and how it was applied:
+- **QUADAS-2** (4 domains: patient selection, index test, reference standard, flow and timing) → 2 reviewers, signaling questions → low/high/unclear risk
+- **RoB 2** (5 domains for RCTs) → randomization, deviations, missing data, outcomes, reporting
+- **MINORS** (12 items for non-randomized surgical studies) → score 0–24
+
+### 2.7 Statistical Analysis
+
+**Pooling:**
+"We pooled effect estimates using a random-effects model with the DerSimonian-Laird estimator. We chose a random-effects model a priori given the expected clinical heterogeneity across studies."
+
+**Heterogeneity:**
+"Statistical heterogeneity was assessed using the I² statistic (threshold: I² > 50% indicates substantial heterogeneity) and Cochran's Q test (P < 0.10). The τ² (between-study variance) and 95% prediction interval were reported."
+
+**Subgroup analyses (pre-specified):**
+List subgroup variables. Example: scanner type, field strength, patient age group, reference standard type.
+
+**Sensitivity analyses:**
+- Leave-one-out analysis
+- Restricting to prospective studies only
+- Restricting to studies with low QUADAS-2 risk of bias
+
+**Publication bias:**
+"Publication bias was assessed using Egger's test (for continuous outcomes) or Deeks' funnel plot asymmetry test (for diagnostic accuracy) when ≥10 studies were available. Trim-and-fill analysis was performed if asymmetry was detected."
+
+**Software:** R (version X.X.X), packages: `meta`, `metafor`. Seed: set.seed(42).
+
+---
+
+## Results (1200–1500 words)
+
+### 3.1 Study Selection
+"Database searches identified [N] records. After removing [N] duplicates, [N] records were screened, of which [N] were excluded at the title/abstract stage. A total of [N] full-text articles were assessed for eligibility; [N] were excluded [with reasons]. Finally, [N] studies (N = [total patients]) were included in the systematic review, of which [N] were included in the meta-analysis. The PRISMA 2020 flow diagram is shown in Figure 1."
+
+### 3.2 Study Characteristics
+"Included studies were published between [year] and [year]. Sample sizes ranged from [N] to [N] participants (median [N], IQR [N–N]). [N] studies were prospective and [N] were retrospective. Studies were conducted in [list countries]."
+
+Reference the characteristics table.
+
+### 3.3 Quality Assessment
+"Quality assessment results are summarized in Figure 2 (QUADAS-2 summary plot). [N] studies were rated at low risk of bias across all four domains. The most common methodological concern was [domain]: [N/N] studies had high or unclear risk."
+
+### 3.4 Quantitative Synthesis
+"[N] studies (N = [patients]) contributed to the primary meta-analysis. The pooled [effect measure] was [value] (95% CI, [lower]–[upper]; I² = [value]%, τ² = [value], P for heterogeneity = [value]). The 95% prediction interval ranged from [value] to [value], indicating that in a new study setting, the true effect could be as low as [value]."
+
+Report each subgroup analysis with test for subgroup interaction (P for interaction).
+
+### 3.5 Publication Bias
+If ≥10 studies: "Egger's test showed [evidence/no evidence] of funnel plot asymmetry (P = [value]). Trim-and-fill analysis imputed [N] studies, resulting in an adjusted pooled estimate of [value] (95% CI, [lower]–[upper])."
+
+---
+
+## Discussion (800–1000 words)
+
+1. **Summary of evidence:** Restate main finding in one sentence with the effect size and units.
+2. **Context with prior meta-analyses:** Compare your estimate to the 2–3 most recent related meta-analyses.
+3. **Heterogeneity explanation:** Clinical (population, technique) vs. statistical. Subgroup findings.
+4. **Quality of evidence:** Comment on overall QUADAS-2/RoB2/MINORS findings. GRADE (if applicable): downgrade for risk of bias, inconsistency, indirectness, imprecision, publication bias.
+5. **Limitations:** (a) cannot rule out publication bias; (b) retrospective study designs; (c) clinical heterogeneity; (d) limited number of studies for subgroup analyses.
+6. **Clinical implications:** What should clinicians take from this?
+
+---
+
+## Conclusions (100–150 words)
+
+"In this systematic review and meta-analysis of [N] studies including [N] patients, [index test] demonstrated a pooled [metric] of [value] (95% CI, [lower]–[upper]) for [condition/outcome]. [One sentence on heterogeneity or quality.] These findings suggest that [clinical recommendation or need for further research]."
+
+---
+
+## Required Figures and Tables
+
+| Item | Content |
+|------|---------|
+| Figure 1 | PRISMA 2020 flow diagram |
+| Figure 2 | Forest plot (primary outcome) |
+| Figure 3 | Funnel plot (if ≥10 studies) |
+| Figure 4 | QUADAS-2 / RoB 2 summary plot |
+| Table 1 | Study characteristics |
+| Table 2 | Quality assessment results |
+| Suppl. Table 1 | Full search strategy (all databases) |
+| Table 3 (or next) | GRADE Summary of Findings table |
+| Table 4 (or next) | Per-study risk of bias table (NOS domain scores for comparative; JBI item scores for single-arm) |
+| Suppl. Table 2 | Excluded studies with reasons |
+
+---
+
+## Common Meta-Analysis Pitfalls
+
+1. **Not registering on PROSPERO** before data extraction — reviewers will ask.
+2. **Fixed-effects model without justification** — use random-effects unless strong homogeneity and similar populations.
+3. **Reporting I² without prediction interval** — I² tells proportion, not magnitude; prediction interval shows the range of true effects.
+4. **Subgroup analyses not pre-specified** — post-hoc subgroups must be labeled as exploratory.
+5. **Including studies where the same dataset was used multiple times** — check for overlapping populations.
+6. **Misinterpreting non-significant Egger's test** — power is often low; state the limitation.
+7. **Confusing statistical significance with clinical significance** — a statistically significant pooled OR = 1.05 may be clinically trivial.
+8. **Missing duplicate publications** — check corresponding author + institution + dates.
+9. **Not reporting absolute numbers** — always provide N patients per group, not just proportions.
+10. **Ignoring GRADE** — many high-impact journals now expect GRADE certainty of evidence ratings.

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

@@ -0,0 +1,297 @@
+# NHIS Cohort Study — Paper Type Template
+
+## Overview
+
+Template for population-based cohort studies using national health insurance claims databases (e.g., Korean NHIS, HIRA, KDCA, or equivalent national registries). Follows the IMRAD structure with specific patterns for emulated target trial frameworks, propensity score balancing, and claims-based outcome definitions. Applicable to vaccination effectiveness, disease-risk, and longitudinal exposure-outcome studies using administrative health data.
+
+---
+
+## Title
+
+- Include study design: "A Nationwide Population-Based Cohort Study" or "An Emulated Target Trial."
+- Specify the population or exposure: "Among Adults Aged 50 Years and Older."
+- Under 20 words preferred.
+- Good: "Association Between Statin Use and Dementia Risk: A Nationwide Population-Based Cohort Study"
+- Bad: "Investigating Statins and Brain Health Using Big Data"
+
+---
+
+## Abstract (Structured, 250 words max)
+
+### Background / Purpose
+- One to two sentences. State the clinical question and gap.
+- Template: "Evidence on the association between {exposure} and {outcome} at the population level remains limited."
+
+### Methods
+- Data source (nationwide cohort, N, date range).
+- Emulated target trial or standard cohort design.
+- Propensity score method (overlap weighting, SIPTW, or 1:1 matching).
+- Cox proportional hazards model for aHRs with 95% CIs.
+- Three to four sentences maximum.
+
+### Results
+- Final analytic cohort size after exclusions.
+- Primary aHR with 95% CI.
+- Key subgroup or time-dependent findings.
+- Three to four sentences maximum.
+
+### Conclusion
+- One to two sentences. Restate the main finding and population-level implication.
+- Be specific about effect magnitude and clinical relevance.
+
+---
+
+## Introduction (3-4 paragraphs, ~400-500 words)
+
+### Paragraph 1: Disease Burden and Clinical Context
+- Establish the clinical importance with global or national prevalence data.
+- Cite GBD, WHO, or national registry statistics.
+- Describe current clinical practice or public health approach.
+
+### Paragraph 2: Prior Evidence and Knowledge Gap
+- Summarize key prior studies (single-country cohorts, small sample sizes, limited follow-up).
+- Flag specific limitations: "were limited to single-country cohorts," "did not adjust for key confounders," "lacked population-level data."
+- State what remains unknown.
+
+### Paragraph 3: Study Objective
+- State the specific aim.
+- If using emulated target trial: "Using an emulated target trial framework, we evaluated the association between {exposure} and {outcome} in a nationwide cohort of {N} individuals."
+- Optional: preview key methodological strengths (large cohort, long follow-up, claims-linked data).
+
+---
+
+## Materials and Methods (~1000-1400 words)
+
+### 3.1 Data Source
+
+Describe the national claims database linkage. Standard components:
+
+1. **Database identification**: Name the data sources (e.g., national health insurance claims, health examination records, disease registry).
+2. **Population coverage**: State the coverage rate (e.g., "covers approximately 98% of the national population").
+3. **Data domains**: Specify linked domains — outpatient and inpatient claims, pharmaceutical records, health examination results, death certificates, and any disease-specific registries.
+4. **Anonymization**: State how personally identifiable information was handled.
+
+Template:
+> This {study type} utilized a large-scale, population-based, nationwide cohort comprising {N} individuals aged {X} years and older from {START DATE} to {END DATE}. The national universal health insurance system enables the integration of comprehensive health data across multiple domains. We integrated outpatient and inpatient data, pharmaceutical records, and death certificates from the national claims database, along with health examination data from the national health insurance service.
+
+### 3.2 Ethics Approval
+
+- State the approving body and protocol number.
+- For claims-based studies: informed consent is typically waived due to use of anonymized records.
+- Template: "This study was approved by {IRB/authority} (approval no. {NUMBER}). The requirement for informed consent was waived due to the use of anonymized health records."
+
+### 3.3 Study Design Framework
+
+If using the emulated target trial approach:
+- State explicitly: "This study was designed and analysed following the emulated target trial framework, which mimics a randomized controlled trial."
+- Reference a supplementary table mapping the target trial components (eligibility, treatment strategy, assignment, follow-up, outcome, causal contrast, analysis plan).
+
+### 3.4 Study Population (Inclusion/Exclusion)
+
+Standard three-tier exclusion cascade:
+1. Death before index date.
+2. Prevalent outcome (any documented diagnosis of the study outcome prior to the index date).
+3. Missing demographic or health examination data (complete-case approach).
+
+For disease-specific studies, add:
+4. Disease-specific exclusion (e.g., diagnosis after the index date for subcohort studies).
+
+Template:
+> Participants were excluded if they (1) died before their assigned index date (n = {X}), (2) had any documented {outcome} diagnosis prior to the index date (n = {X}), or (3) lacked complete demographic information or health examination data (n = {X}). Following exclusions, {N} individuals were included in the final analytic cohort (Fig. 1).
+
+Always include a CONSORT-style flow diagram (Fig. 1).
+
+### 3.5 Exposure Definition
+
+- Define the exposure with specific dates, codes, or criteria.
+- For time-dependent exposures: address immortal time bias.
+- Template for index date assignment: "To address immortal time bias, index dates for unexposed individuals were randomly assigned to match the distribution of index dates in the exposed group."
+- Specify follow-up: "Participants were followed from the index date until the first occurrence of an outcome, death, or the end of follow-up ({END DATE}), whichever occurred first."
+
+### 3.6 Outcome Definition
+
+- Define using ICD-10 codes (defer code lists to supplementary table).
+- For composite outcomes: define each component and the time windows.
+- Template: "The primary outcomes were: (1) {OUTCOME 1}, (2) {OUTCOME 2}. Outcomes were identified using ICD-10 codes (Table S{X}). {COMPLICATION} was defined as a claim occurring within {X days} to {Y year} after the incident event."
+- For stricter definitions: "An outcome was considered valid if at least two ICD-10 diagnostic claims within 1 year and related medication prescriptions were recorded."
+
+### 3.7 Covariate Adjustment
+
+Standard covariate set for national health insurance cohort studies:
+
+| Category | Variables | Typical Categorization |
+|----------|-----------|----------------------|
+| Demographics | Age | 5-year bins or clinical cutoffs |
+| | Sex | Male, Female |
+| | Region of residence | Urban, Rural |
+| | Household income | Low, Middle, High (percentile-based) |
+| Clinical (exam) | BMI | Country-specific cutoffs (e.g., Asia-Pacific: <23.0, 23.0-24.9, >=25.0 kg/m2) |
+| | Blood pressure | SBP/DBP threshold-based (e.g., SBP >=140 or DBP >=90 mmHg) |
+| | Fasting blood glucose | Threshold-based (e.g., <100, >=100 mg/dL) |
+| | GFR | Categories (e.g., <60, 60-89, >=90 mL/min/1.73 m2) |
+| Comorbidity (claims) | Charlson comorbidity index | 0, 1, >=2 |
+| | Medication history | Hypertension, Hyperlipidemia, Coronary artery disease |
+| Behavioral (survey) | Smoking | Never, Former, Current |
+| | Alcohol consumption | Frequency categories |
+| | Physical activity | Sufficient/Insufficient (MET-based) |
+
+Notes:
+- BMI cutoffs should follow country-specific guidelines with explicit justification.
+- Covariates are collected from the most recent data available prior to the individual index date.
+- State the missing data approach (complete-case or multiple imputation with chained equations).
+
+### 3.8 Statistical Analysis
+
+#### Propensity Score Balancing
+
+Three validated methods (choose one):
+
+1. **Overlap weighting**: Maintains full sample size; weights each individual by the probability of being assigned to the opposite group.
+2. **Stabilized IPTW**: Maintains full sample size; stabilized weights reduce variance.
+3. **1:1 Propensity score matching**: Greedy nearest-neighbor within a specified caliper (e.g., 0.001 SD), without replacement.
+
+Template:
+> Propensity scores were estimated using multivariable logistic regression incorporating {list all covariates from 3.7}. Covariate balance was evaluated using standardized mean differences, with values <0.1 indicating adequate balance.
+
+#### Effect Estimation
+
+- Cox proportional hazards models for adjusted hazard ratios (aHRs) with 95% CIs.
+- Optional: restricted mean survival time (RMST) for absolute clinical benefit.
+- Censoring: outcome, death, or administrative end date (competing-risks framework).
+
+#### Stratification and Time-Dependent Analysis
+
+- Stratify by all major covariates (age, sex, income, region, BMI, smoking, alcohol, physical activity, comorbidities).
+- Time-dependent effectiveness in interval bins (e.g., <1, 1-2, 2-4, 4-6, 6-8, >=8 years).
+- Note: "Stratified analyses were adjusted for all covariates except the stratification variable."
+
+#### Sensitivity Analyses
+
+Common approaches for claims-based cohort studies:
+- Stricter outcome definition (e.g., 2+ claims + medication prescription).
+- Negative control outcomes (outcomes with no expected association).
+- Mediation analysis (if plausible intermediate pathway exists).
+- Alternative PS method (e.g., matching vs. weighting).
+- Subgroup by exposure timing or severity.
+
+#### Software and Significance
+
+- State software and version (e.g., SAS 9.4, R, Python).
+- "A two-sided p value of less than 0.05 was considered statistically significant."
+- For multiple primary outcomes: consider Bonferroni correction.
+
+---
+
+## Results (~800-1200 words)
+
+### 4.1 Study Population
+- Total screened, excluded (with per-criterion counts), and final cohort.
+- Reference the flow diagram (Fig. 1).
+- Baseline characteristics table (Table 1) showing pre- and post-weighting/matching balance.
+- Report SMDs to confirm adequate balance (<0.1).
+
+### 4.2 Primary Endpoint
+- aHR with 95% CI for the primary outcome.
+- Event counts and incidence rates in exposed vs. unexposed groups.
+- Template: "In the {weighted/matched} cohort, {exposure} was associated with a {increased/decreased} risk of {outcome} (aHR, {value}; 95% CI, {lower}-{upper})."
+
+### 4.3 Secondary Endpoints and Subgroup Analyses
+- Present secondary outcomes with aHRs and 95% CIs.
+- Stratification results: forest plot (Fig. 2) or table.
+- Time-dependent effectiveness: present by time interval.
+
+### 4.4 Sensitivity Analyses
+- Confirm or qualify main findings with sensitivity results.
+- Reference supplementary tables and figures.
+
+### Rules
+- Same rules as original article: no interpretation in Results.
+- Every table and figure must be referenced.
+- Numbers must match tables exactly.
+- Use "was associated with," not causal language.
+
+---
+
+## Discussion (4-5 paragraphs, ~800-1200 words)
+
+### Paragraph 1: Summary of Key Findings
+- Restate principal findings in context of the study aim.
+- Paraphrase, do not repeat exact numbers.
+
+### Paragraphs 2-3: Comparison with Prior Literature
+- Anchor paper-driven comparison.
+- Address concordant and discordant findings.
+- Discuss potential reasons for discrepancies (population, exposure definition, follow-up duration, confounding adjustment).
+
+### Paragraph 4: Clinical Implications
+- Population-level public health implications.
+- Policy relevance (screening, vaccination, treatment guidelines).
+
+### Paragraph 5: Limitations
+
+Standard limitation categories for claims-based cohort studies (ordered by severity):
+
+| # | Category | Template |
+|---|----------|----------|
+| 1 | Generalizability | "The study population comprised predominantly {nationality/ethnicity} adults, which may limit generalizability to other populations." |
+| 2 | Residual confounding | "Despite {PS method}, residual confounding from unmeasured variables cannot be entirely excluded." |
+| 3 | ICD-based misclassification | "Both exposure and outcomes were identified through administrative records, which may be subject to misclassification." |
+| 4 | Unmeasured confounders | "Data on {specific confounders} were unavailable." |
+| 5 | Rare outcome precision | "Effectiveness estimates for rare outcomes should be interpreted cautiously due to low event counts and wide confidence intervals." |
+| 6 | Missing comparator | "{Alternative treatment} was not widely available during the study period, precluding direct comparison." |
+| 7 | Administrative coding accuracy | "ICD-based outcome definitions may not capture all clinically relevant cases." |
+
+Follow each limitation with a mitigation sentence. End with a strengths paragraph.
+
+### Conclusion
+- One to two sentences. Citable statement with effect magnitude and implication.
+
+---
+
+## Tables
+
+- **Table 1**: Baseline characteristics, pre- and post-weighting/matching, with SMDs.
+- **Table 2**: Primary and secondary outcome aHRs with 95% CIs (overall and by subgroup).
+- **Table S1+**: ICD-10 code definitions (supplementary).
+- **Table S{X}**: Emulated target trial specification (if applicable).
+
+## Figures
+
+- **Figure 1**: Patient flow diagram (exclusion cascade with counts).
+- **Figure 2**: Forest plot of stratified analyses.
+- **Figure 3**: Time-dependent effectiveness plot (aHR by follow-up interval).
+- Optional: Kaplan-Meier curves, RMST difference plots.
+
+---
+
+## ML Prediction Model Variant
+
+When the NHIS cohort study is a machine learning prediction model (rather than Cox regression):
+
+### Key Differences
+- **Exposure → Prediction target**: Define the outcome prediction window (e.g., 5-year incidence).
+- **PS balancing → Train/test split**: Stratified 80:20 split of discovery cohort.
+- **Class imbalance**: Address with SMOTE or other oversampling techniques.
+- **Model selection**: List candidate models with hyperparameters; ensemble top performers.
+- **Evaluation**: Sensitivity, specificity, balanced accuracy, AUROC. Primary metric justified by class imbalance.
+- **Feature importance**: MDI or SHAP values; ablation study for validation.
+- **External validation**: Independent cohort with same preprocessing pipeline.
+- **Software**: Python (TensorFlow/Scikit-learn) for ML; SAS for descriptive/inferential statistics.
+
+---
+
+## Checklist Before Submission
+
+- [ ] Data source described with coverage rate, linked domains, and anonymization
+- [ ] Ethics approval with specific protocol number
+- [ ] Emulated target trial table in supplementary (if applicable)
+- [ ] Exclusion cascade with per-criterion counts and flow diagram
+- [ ] Immortal time bias addressed (random index date assignment or equivalent)
+- [ ] All covariates listed with categorization and country-specific cutoff justification
+- [ ] PS method specified with balance assessment (SMD < 0.1)
+- [ ] ICD-10 codes deferred to supplementary table
+- [ ] Time-dependent analysis with interval bins
+- [ ] At least 2 sensitivity analyses performed
+- [ ] Limitation paragraph covers generalizability, residual confounding, and misclassification
+- [ ] Numbers consistent between text, tables, and figures
+- [ ] Reporting guideline (STROBE or RECORD) items addressed