The Evidence Layer for Scientific AI

Radiomics → Post-pancreatic-surgery Diabetes Meta-analysis (External Validation): Evidence Status

Evidence mismatch (critical): In the provided research payload, I see radiomics models for hepatocellular carcinoma (HCC) and other unrelated biomedical meta-analyses, but no studies about radiomics predicting diabetes after pancreatic surgery, nor any external-validation DTA/meta-analysis in that specific clinical target.

Therefore, any attempt to compute pooled external-validation performance (e.g., pooled AUC/sensitivity/specificity) for the pancreatic-surgery-diabetes question would require studies that are not present in your supplied data.

What I *can* visualize now (from provided radiomics example)

The only provided radiomics study with explicit external testing is an HCC CT radiomics model predicting tumor micronecrosis. I visualize its transport to an external set to demonstrate the exact metric types you’ll later pool for your pancreatic target. HCC radiomics external validation performance

1) AUC: Development vs Internal Test vs External Test

Observed pattern: AUC drops from 0.83 (development) → 0.80 (internal) → 0.76 (external), consistent with the common phenomenon that external transport is harder than internal validation. AUC across splits

2) Sensitivity/Specificity/Accuracy: Development vs External Test

Note the trade-off: external sensitivity (0.559) is lower than development (0.724), while external specificity is higher (0.820 vs 0.781) in this example. Sensitivity/specificity/accuracy across splits

3) Class prevalence by split (why sensitivity/specificity can shift)

Prevalence differs across splits (e.g., external test ~0.595), which can influence sensitivity/specificity depending on thresholds and calibration. Positive prevalence by split

How to perform the requested meta-analysis (rigorous checklist)

Below is a target-agnostic but external-validation DTA meta-analysis-ready workflow. Once you provide the pancreatic-surgery-diabetes radiomics papers (with external cohorts), we can compute pooled performance and quantify heterogeneity.

• Inclusion definition: radiomics predicting post-pancreatic surgery diabetes with an external validation cohort (not just internal random split). (Right now, that evidence isn’t in the payload.)
• Extract: AUC/ROC metrics plus sensitivity & specificity (and the positivity threshold used), ideally with sample sizes and event counts.
• Model transport & calibration: report whether external performance used the same preprocessing, segmentation protocol, and scanner harmonization. External test performance is often worse (as in the example study), so heterogeneity matters. Reproducibility/generalizability limitations
• Meta-analytic layer: choose a DTA framework that respects threshold variability (e.g., bivariate HSROC / Reitsma-style approaches). Also consider meta-regression for scanner/protocol differences.
• Bias checks: assess publication bias and selective reporting (radiomics often has “winner’s curse” risk). The example highlights that code/data were not fully public, which can hamper replication. Data/code availability limitation

Scientific critical note (why I’m stopping short)

The specific requested target (post-pancreatic surgery diabetes) is not represented in the supplied evidence. I’m therefore not computing a pancreatic-surgery pooled estimate—because doing so would require study-level data that I do not have.