Graphs Built From Reported Results

 Quick Analysis Plan Copied

 Long Analysis Plan

What the evidence supports plotting now vs what is missing

The supplied records contain no per-gene AUC values (no gene-level ROC statistics for LDOC1, FUT8, EGFR, or any other gene), so per-gene ROC curves cannot be rendered from these records alone. What IS reported: (1) the NHR nomogram AUCs with confidence intervals, and (2) the GSRC survival nomogram AUCs at 3/6/12 months, both from cohort splits (994/427 training-validation and 703/300 respectively) GDM NHR nomogram AUC study Gastric signet ring carcinoma survival nomogram. These are graphed below (reported values only, not simulated ROC curves).

Note: blue = training cohort, light blue = validation. Error bars (CI 0.72–0.80 and 0.66–0.78) exist only for the GDM model; GSRC AUCs were reported without CIs, an asymmetry reflecting differences in reporting completeness.

Analysis plan: generate full ROC curves + per-gene AUCs from raw data

Once you supply the raw per-gene expression/score data and nomogram risk scores with outcome labels, run this workflow:

Python workflow (sklearn): 1) Load a tidy table: rows = samples, columns = per-gene values (e.g., LDOC1, FUT8, EGFR IHC/expression), nomogram total point score, and binary outcome (GDM yes/no, survival at horizon, or gefitinib response). 2) For each predictor, call sklearn.metrics.roc_curve(y, predictor) and roc_auc_score; bootstrap 1000 resamples for 95% CIs. 3) Overlay the nomogram ROC on the same axes as the best single gene to test added discriminative value (DeLong test via roc_auc comparisons or statsmodels). 4) Plot in Plotly with diagonal reference line (AUC 0.5), per-gene curves in muted colors, nomogram curve emphasized in the accent color. Validation-set curves should be a separate panel to avoid optimistic bias, since single-center, internally validated nomograms can overestimate performance GDM study limitations GSRC study limitations.

Blind spots to check: the GSRC nomogram AUCs decline monotonically with prediction horizon (0.845 → 0.693), typical for survival models — compare AUCs only at matched horizons. Neither nomogram record reports per-gene contributions, and the EGFR SISH data (10 mutation-positive patients listed) are far too small (n=96 cohort, 31 mutations) to support a stable ROC curve for EGFR copy number as a screening test EGFR SISH screening study.

What would change this conclusion

If you upload per-gene expression matrices with outcomes (e.g., the LDOC1 cohort of 200 NSCLC patients or the GDM lab panels), per-gene AUCs become directly computable and the plot above would be replaced by true ROC curves. Falsifying the nomogram utility shown here would require external-cohort validation showing AUC degradation below ~0.65.