AgentSkillsCN

eda

针对临床与药代动力学数据集开展探索性数据分析，借助 Quarto 报告模板，系统梳理人口学特征、药代动力学概要、协变量分布及数据质量评估等内容。

SKILL.md
--- frontmatter
name: eda
description: Exploratory Data Analysis for clinical and PK datasets. Scaffolds a comprehensive Quarto report with demographics, PK summaries, covariate distributions, and data quality assessment.

Exploratory Data Analysis

Usage

code
/eda
/eda <dataset_path>

What This Does

Scaffolds a comprehensive EDA report:

  1. •Demographics summary
  2. •PK concentration summary by dose, visit, analyte
  3. •Covariate distributions and correlations
  4. •Missing data assessment
  5. •Outlier detection
  6. •Dataset quality flags

Protocol

Step 1: Fetch Documentation

Use Context7 for ggplot2, gt, and data manipulation packages.

Step 2: Identify Dataset

If $ARGUMENTS provides a path, use it. Otherwise ask for:

  • •Dataset location
  • •Dataset type (PK, PD, PKPD, demographics, lab data)
  • •Key column names (ID, TIME, DV, DOSE, covariates)
  • •Any known data issues or exclusions

Step 3: Scaffold Report

Copy templates/eda-report.qmd to reports/.

Step 4: Customize Sections

Demographics:

  • •Summary table: N, Age, Sex, Race, Weight, Height, BMI, renal/hepatic function
  • •Stratified by treatment group
  • •Balance assessment across groups

PK Concentrations:

  • •N observations by dose, visit, timepoint
  • •BLQ summary (count and percentage by timepoint)
  • •Concentration-time spaghetti plots (linear + semi-log)
  • •Dose-normalized concentrations if multiple dose levels

Covariates:

  • •Histograms/density plots for continuous covariates
  • •Bar charts for categorical covariates
  • •Correlation matrix (continuous covariates)
  • •Covariate vs PK parameter scatter plots (if individual estimates available)

Missing Data:

  • •Missing data pattern visualization
  • •Percentage missing by variable
  • •Assessment of missingness mechanism (MCAR/MAR/MNAR)

Outliers:

  • •Statistical outlier detection (e.g., > 3 SD from mean)
  • •Visual outlier flagging on concentration-time plots
  • •Listing of flagged observations for review

Data Quality:

  • •Duplicate record check
  • •Time sequence validation (no negative time gaps)
  • •Dose-concentration consistency
  • •BLQ below actual LLOQ value

Step 5: Generate and Review

  • •Render the Quarto report
  • •Summarize key findings
  • •Flag data issues requiring resolution before modeling

Non-Negotiables

  • •All figures at dpi = 300
  • •Use here::here() for file paths
  • •No hardcoded values in summary tables
  • •Use MM color palette
  • •Document all data exclusions with rationale