Complete meta-analysis generation for neurosurgical outcomes
npx skills add https://github.com/matheus-rech/meta-agent --skill meta-analysisKopieren Sie diesen Befehl und fügen Sie ihn in Claude Code ein, um den Skill zu installieren
| name | meta-analysis |
| version | 1.0.0 |
| description | Complete meta-analysis generation for neurosurgical outcomes |
| author | NeuroResearch Agent |
| license | MIT |
| triggers | [{"pattern":"meta-analysis"},{"pattern":"forest plot"},{"pattern":"funnel plot"},{"pattern":"heterogeneity"},{"pattern":"pool.*studies"},{"pattern":"synthesize.*data"},{"pattern":"publication bias"},{"pattern":"subgroup analysis"},{"pattern":"sensitivity analysis"},{"pattern":"I-squared|I²"}] |
| requires | ["r-execute","filesystem"] |
| tools | [{"name":"analyze_binary","description":"Meta-analysis of binary outcomes (OR, RR, RD)","script":"scripts/binary_meta.R"},{"name":"analyze_continuous","description":"Meta-analysis of continuous outcomes (MD, SMD)","script":"scripts/continuous_meta.R"},{"name":"analyze_proportion","description":"Meta-analysis of proportions (single-arm studies)","script":"scripts/proportion_meta.R"},{"name":"analyze_survival","description":"Meta-analysis of hazard ratios","script":"scripts/survival_meta.R"},{"name":"generate_forest","description":"Generate publication-quality forest plot","script":"scripts/forest_plot.R"},{"name":"assess_heterogeneity","description":"Comprehensive heterogeneity assessment","script":"scripts/heterogeneity.R"},{"name":"assess_publication_bias","description":"Publication bias assessment","script":"scripts/publication_bias.R"},{"name":"run_subgroup","description":"Subgroup meta-analysis","script":"scripts/subgroup.R"},{"name":"run_sensitivity","description":"Sensitivity analyses (leave-one-out, influence)","script":"scripts/sensitivity.R"}] |
| schemas | {"binary_input":{"required":[{"study":"string"},{"events_int":"integer"},{"n_int":"integer"},{"events_ctrl":"integer"},{"n_ctrl":"integer"}],"optional":[{"year":"integer"},{"subgroup":"string"}]},"continuous_input":{"required":[{"study":"string"},{"n_int":"integer"},{"mean_int":"number"},{"sd_int":"number"},{"n_ctrl":"integer"},{"mean_ctrl":"number"},{"sd_ctrl":"number"}],"optional":[{"year":"integer"},{"subgroup":"string"}]},"proportion_input":{"required":[{"study":"string"},{"events":"integer"},{"total":"integer"}],"optional":[{"year":"integer"},{"subgroup":"string"}]}} |
| outputs | ["figures/forest_*.png","figures/funnel_*.png","results/meta_summary.csv","results/heterogeneity.txt","results/publication_bias.txt"] |
This skill provides comprehensive meta-analysis capabilities for neurosurgical systematic reviews. It supports binary outcomes (mortality, complications), continuous outcomes (scores, measurements), proportions (single-arm success rates), and survival data (hazard ratios).
Activate this skill when the user:
Before analysis, ensure data is in the correct format:
# Binary outcomes (2x2 data)
study,year,events_int,n_int,events_ctrl,n_ctrl,subgroup
Smith,2020,10,50,5,50,early
Jones,2021,15,60,8,55,late
# Continuous outcomes
study,year,n_int,mean_int,sd_int,n_ctrl,mean_ctrl,sd_ctrl
Smith,2020,50,25.3,5.2,50,30.1,6.1
# Proportions (single-arm)
study,year,events,total
Smith,2020,42,50
Jones,2021,38,55
For binary outcomes (OR, RR, RD):
Run a meta-analysis of mortality using OR with random effects.
Input: extractions/pooled_data.csv
For continuous outcomes (MD, SMD):
Run a meta-analysis of ODI scores using MD.
Input: extractions/continuous_data.csv
For proportions:
Calculate the pooled proportion of surgical success.
Input: extractions/case_series_data.csv
75%: Considerable heterogeneity
When I² > 50%, consider:
For each meta-analysis, generate:
library(meta)
library(metafor)
data <- read.csv("{{INPUT}}")
ma <- metabin(
event.e = events_int,
n.e = n_int,
event.c = events_ctrl,
n.c = n_ctrl,
studlab = paste(study, year),
data = data,
sm = "{{MEASURE}}", # OR, RR, RD
method = "MH", # MH, Inverse, GLMM
random = TRUE,
fixed = FALSE,
prediction = TRUE,
hakn = TRUE # Knapp-Hartung adjustment
)
# Forest plot
png("{{OUTPUT}}/forest_plot.png", width = 1200, height = 800, res = 150)
forest(ma,
sortvar = TE,
prediction = TRUE,
print.tau2 = TRUE,
print.I2 = TRUE,
leftcols = c("studlab", "event.e", "n.e", "event.c", "n.c"),
rightcols = c("effect", "ci", "w.random"))
dev.off()
# Comprehensive heterogeneity assessment
cat("=== HETEROGENEITY ASSESSMENT ===\n\n")
# Statistics
cat(sprintf("I² = %.1f%% [%.1f%%, %.1f%%]\n",
ma$I2*100, ma$lower.I2*100, ma$upper.I2*100))
cat(sprintf("τ² = %.4f (τ = %.4f)\n", ma$tau2, sqrt(ma$tau2)))
cat(sprintf("H² = %.2f\n", ma$H^2))
cat(sprintf("Q = %.2f, df = %d, p = %.4f\n", ma$Q, ma$df.Q, ma$pval.Q))
cat(sprintf("\nPrediction interval: [%.2f, %.2f]\n",
exp(ma$lower.predict), exp(ma$upper.predict)))
# Baujat plot (outlier detection)
png("{{OUTPUT}}/baujat_plot.png", width = 800, height = 600, res = 150)
baujat(ma)
dev.off()
# GOSH plot (if < 15 studies)
if (ma$k <= 15) {
gosh_result <- gosh(ma)
png("{{OUTPUT}}/gosh_plot.png", width = 800, height = 600, res = 150)
plot(gosh_result)
dev.off()
}
# Only reliable if k >= 10
if (ma$k >= 10) {
# Funnel plot
png("{{OUTPUT}}/funnel_plot.png", width = 800, height = 600, res = 150)
funnel(ma, studlab = TRUE, cex.studlab = 0.7)
dev.off()
# Contour-enhanced funnel plot
png("{{OUTPUT}}/funnel_contour.png", width = 800, height = 600, res = 150)
funnel(ma, contour = c(0.9, 0.95, 0.99),
col.contour = c("darkgray", "gray", "lightgray"))
legend("topright", c("p < 0.10", "p < 0.05", "p < 0.01"),
fill = c("darkgray", "gray", "lightgray"), bty = "n")
dev.off()
# Statistical tests
cat("\n=== PUBLICATION BIAS TESTS ===\n\n")
# Egger's test (continuous outcomes)
egger <- metabias(ma, method = "Egger")
cat(sprintf("Egger's test: t = %.2f, p = %.4f\n",
egger$statistic, egger$p.value))
# Peters' test (binary outcomes - preferred)
peters <- metabias(ma, method = "Peters")
cat(sprintf("Peters' test: t = %.2f, p = %.4f\n",
peters$statistic, peters$p.value))
# Trim-and-fill
tf <- trimfill(ma)
cat(sprintf("\nTrim-and-fill: %d studies imputed\n", tf$k0))
cat(sprintf("Adjusted estimate: %.2f [%.2f, %.2f]\n",
exp(tf$TE.random), exp(tf$lower.random), exp(tf$upper.random)))
# Funnel with imputed studies
png("{{OUTPUT}}/funnel_trimfill.png", width = 800, height = 600, res = 150)
funnel(tf, studlab = TRUE)
dev.off()
} else {
cat("\nNote: k < 10, publication bias tests unreliable.\n")
cat("Visual inspection of funnel plot only.\n")
}
# Leave-one-out analysis
l1o <- metainf(ma, pooled = "random")
png("{{OUTPUT}}/leave_one_out.png", width = 1000, height = 600, res = 150)
forest(l1o)
dev.off()
# Influence diagnostics
inf <- influence(ma)
png("{{OUTPUT}}/influence.png", width = 1000, height = 800, res = 150)
plot(inf)
dev.off()
# Cumulative meta-analysis (by year)
cum <- metacum(ma, sortvar = data$year)
png("{{OUTPUT}}/cumulative.png", width = 1000, height = 600, res = 150)
forest(cum)
dev.off()
# Report influential studies
cat("\n=== INFLUENTIAL STUDIES ===\n\n")
influential <- which(inf$is.infl)
if (length(influential) > 0) {
cat("The following studies are potentially influential:\n")
for (i in influential) {
cat(sprintf(" - %s\n", ma$studlab[i]))
}
} else {
cat("No studies identified as influential.\n")
}
# Update with subgroup
ma_sub <- update(ma, subgroup = data${{SUBGROUP}})
# Forest plot with subgroups
png("{{OUTPUT}}/forest_subgroup.png", width = 1400, height = 1000, res = 150)
forest(ma_sub,
sortvar = TE,
prediction = TRUE,
subgroup = TRUE,
print.subgroup.labels = TRUE,
subgroup.hetstat = TRUE)
dev.off()
# Test for subgroup differences
cat("\n=== SUBGROUP ANALYSIS ===\n\n")
cat(sprintf("Test for subgroup differences:\n"))
cat(sprintf(" Q = %.2f, df = %d, p = %.4f\n",
ma_sub$Q.b.random, ma_sub$df.Q.b, ma_sub$pval.Q.b.random))
# Within-subgroup heterogeneity
for (sg in unique(data${{SUBGROUP}})) {
ma_sg <- subset(ma_sub, subgroup == sg)
cat(sprintf("\n%s (k = %d):\n", sg, ma_sg$k))
cat(sprintf(" Effect: %.2f [%.2f, %.2f]\n",
exp(ma_sg$TE.random), exp(ma_sg$lower.random), exp(ma_sg$upper.random)))
cat(sprintf(" I² = %.1f%%\n", ma_sg$I2 * 100))
}
If SDs are not reported, estimate from:
Add continuity correction:
metabin(..., incr = 0.5, method.incr = "only0")
Convert using Wan et al. (2014) method:
library(estmeansd)
bc.mean.sd(q1.val = Q1, med.val = median, q3.val = Q3, n = n)
If only HR and CI reported:
log_hr <- log(hr)
se <- (log(ci_upper) - log(ci_lower)) / 3.92
After analysis, assess certainty:
Generate Summary of Findings table with absolute effects.
Use when extracting structured data from medical research PDFs, parsing study characteristics, patient demographics, outcomes, and results. Invoke for systematic review data collection from papers.
Use when writing systematic review manuscript sections following PRISMA 2020 guidelines. Covers abstract, introduction, methods, results, and discussion drafting for medical journals. Invoke for academic writing assistance.
Use when performing meta-analysis, pooling study data, generating forest plots, funnel plots, assessing heterogeneity, or conducting subgroup and sensitivity analyses. Invoke for any statistical synthesis of multiple studies.
Use when searching for neurosurgery literature, developing PubMed search strategies, identifying MeSH terms, or building systematic search queries. Invoke for literature searching in neurosurgery topics.
Use when assessing risk of bias or study quality. Covers RoB 2 for RCTs, Newcastle-Ottawa Scale for cohorts, ROBINS-I for non-randomized interventions, and QUADAS-2 for diagnostic studies. Invoke for quality assessment.
Extract structured data from neurosurgical studies