CCGA substudy 1 training set: cancer signal sensitivity at 98% specificity under 10-fold cross-validation

Stable ID: ctdnajam-20261010-protocol-ccga1-training-cv-sens-98spec

areas

dna-genomes

contexts

clinical_research

protocol

10-fold cross-validation on the 1,414-participant training set; held-out fold scores are pooled and the threshold set post hoc to give 98% specificity in the training non-cancer participants. The pan-feature classifier has no training-set value.

version

Jamshidi et al. 2022, Table 3 (training columns)

metric

sensitivity-at-98-percent-specificity

metric direction

higher

unit

percent

metric definition

Share of cancer participants called positive at the score threshold that gives 98% specificity among the non-cancer participants of the same set. 95% CIs are Clopper-Pearson.

limitations

Developer study: GRAIL designed the assays and classifiers, ran every arm, and reports the comparison.; The 98% specificity threshold was set post hoc within each set, including the validation set, so it is not a threshold locked before validation.; Case-control enrolment of clinically diagnosed cancers and matched non-cancer participants, not an intended-use screening population.; Prototype assays; the targeted methylation test developed afterwards is a different assay and is not scored here.; The fragment endpoints, fragment lengths, allelic imbalance and pan-feature classifiers were developed after validation blinding was lifted; the other classifiers were blinded.; Table 3 scores 833 training and 464 validation cancers against 854 and 485 analysable cancers in Table 1. STAR Methods restrict detection to solid cancers, and Table 1 lists 11 plasma cell neoplasms and 10 leukaemias in training and 8 and 13 in validation, which is 21 in each set; lymphomas stay in.; Specificity was 97.9% (548/560) in training and 97.8% (354/362) in validation for every cfDNA classifier, slightly below the 98% target.; Cross-validated within the set used to design the classifiers; the fragment endpoint threshold tier was itself chosen to maximise sensitivity at 98% specificity in these folds.; For the validation set, the WG methylation, SNV, SNV-WBC, SCNA, SCNA-WBC and clinical data classifiers were analysed double-blinded (STAR Methods); the fragment endpoint, fragment length, allelic imbalance and pan-feature classifiers were developed after blinding was lifted.

source locator

Table 3; STAR Methods, classifier descriptions

• uses data: CCGA substudy 1 training set (1,414 analysable participants)
• assessment: allelic imbalance on CCGA substudy 1 training set: cancer signal sensitivity at 98% specificity under 10-fold cross-validation
• assessment: clinical data on CCGA substudy 1 training set: cancer signal sensitivity at 98% specificity under 10-fold cross-validation
• assessment: fragment endpoints on CCGA substudy 1 training set: cancer signal sensitivity at 98% specificity under 10-fold cross-validation
• assessment: fragment lengths on CCGA substudy 1 training set: cancer signal sensitivity at 98% specificity under 10-fold cross-validation
• assessment: SCNA on CCGA substudy 1 training set: cancer signal sensitivity at 98% specificity under 10-fold cross-validation
• assessment: SCNA-WBC on CCGA substudy 1 training set: cancer signal sensitivity at 98% specificity under 10-fold cross-validation
• assessment: SNV on CCGA substudy 1 training set: cancer signal sensitivity at 98% specificity under 10-fold cross-validation
• assessment: SNV-WBC on CCGA substudy 1 training set: cancer signal sensitivity at 98% specificity under 10-fold cross-validation
• assessment: WG methylation on CCGA substudy 1 training set: cancer signal sensitivity at 98% specificity under 10-fold cross-validation
• assessed by: Select a plasma ctDNA fragmentomics detection workflow
• assessed by: Select a plasma ctDNA methylation detection workflow