Peer review round 5: LivingGLP1 protocol 0.5.0 (reviewer grok-4.7)

Round 4 comments

G4-1. Resolved. Confidence in cumulative evidence builds the hazard scale contribution network from the trial level log hazard ratios that entered the primary model, adds 0.5 to every cell of a zero event or all event trial for contributions only, and keeps every primary trial. Restricted sensitivity analyses, including the odds ratio analysis, are rated on their own. The qualitative fallback is sound: it applies only when a finite contribution matrix still cannot be computed, and within study bias and indirectness then use the full evidence set, with reasons recorded. Extending 0.5 to all event arms is right, because a zero non-event cell leaves the same inverse variance weight undefined.

G4-2. Partly resolved. The outcomes.csv mapping rules for O22D, O22E, and O50 match the proposed text, and O22E now includes P08A and P08B. Composite definitions still names O22C twice in one sentence: first as "doubling of serum creatinine, about a 57 percent decline," then as "kidney failure, sustained doubling of serum creatinine, or kidney death." The row is the second definition. The table is authoritative, so extraction would follow the row, but the prose contradicts it. Delete the first O22C parenthesis.

G4-3. Resolved. The network key and the connectivity rule both use population, population stratum, outcome, definition stratum, window, and estimand. The definition stratum is the component set and threshold when an identifier allows more than one.

G4-4. Resolved. A reported hazard ratio enters a normal likelihood on the log hazard ratio. A cumulative incidence at a common time enters as log(minus log(1 minus p)) with a delta method standard error and offset log t. Counts use the complementary log log model. Denominators are the number randomized. The safety population is used only when the randomized number is missing, and those trials are flagged. The 95 percent known status gate is an added restriction from the harms revision. The denominator rule still stands. The new cumulative incidence input leaves a contribution gap, listed below.

G4-5. Resolved. IGLARLIXI_SWITCH_<strategy>_SC_TIT_QD has is_template = yes and stays out of every network until a strategy identifier, the permitted pens, the switching threshold, and the titration algorithm are recorded. The prose uses one node per set of permitted pens, switching threshold, and titration algorithm, combined with the two fixed pens only in a sensitivity analysis.

G4-6. Resolved. The sensitivity list includes iGlarLixi nodes combined; recorded harms excluding trials with more than 20 percent incomplete follow-up or an arm difference above 10 percentage points; the 100 percent known status restriction; and exclusion of the O04 mapping.

G4-7. Resolved. When only one reference arm covers the analyzed horizon, that arm's risk is used and no between arm heterogeneity is estimated.

New issues

Major

1. Competing risk synthesis is not yet a network (protocol.qmd, Competing risks; Networks and nodes; Absolute effects). Cause-specific and subdistribution hazard ratios are different parameters, and the paragraph says they are synthesized separately. The network key has no hazard type. Definition stratum is the component set and threshold. Estimand is the ICH E9 strategy. Results that share those fields fall in one network. Absolute effects call for an Aalen-Johansen cumulative incidence, and they allow a cause-specific conversion only when the competing event hazard is available, without a formula. Binary harms that death can preclude and that are outside the named list (O32A, O32B, O35B, O38, O39, O121) still use the randomized denominator, so a death without the event counts as event free. "Negligible mortality" has no cutoff.

Replacement: "Add competing risk hazard type to the network key and the connectivity rule, with values cause-specific, subdistribution, or not applicable. Those two hazard types never share a network. For a subdistribution hazard ratio, the absolute risk at the analyzed horizon is 1 minus (1 minus F_ref) raised to HR, where F_ref is the reference Aalen-Johansen cumulative incidence. For a cause-specific hazard ratio, convert only when the reference cause-specific hazard λ1 and the competing hazard λ2 are both available, as F(t) = λ1/(λ1+λ2) times (1 minus exp(minus (λ1+λ2)t)); otherwise report relative effects only. Use the same split for every outcome death can preclude, including cancer, suicide death, and acute kidney injury. Exempt an adverse event analysis only when all cause mortality is below 1 percent in every arm." Advisors may lock a different mortality cutoff with the other thresholds before extraction. The split itself belongs in the registered methods.

2. The 95 percent count rule does not choose a time or a meaning of known status (protocol.qmd, Event outcomes and harms, Hazard ratio analysis). Counts enter when event status is known for at least 95 percent of randomized participants in every arm "at a common within trial time." Which time, and whether administrative censoring counts as known, is unstated. The worked example (risks 0.50 and 0.75, hazard ratio 2.0, giving 1.93, 1.88, and 1.79 after 5, 10, and 20 percent lost at the start) is the bias from leaving unknown participants in the randomized denominator and coding them as event free. Independent censoring that removes them from the risk set leaves the hazard ratio at 2.0.

Replacement: "For each trial, t is the longest follow-up at which at least 95 percent of randomized participants in every arm have known status. Known status means the participant had the event by t or was observed event free through t. Administrative end of follow-up counts as observed. Loss to follow-up before t is unknown. Counts of participants with at least one event by t enter the binomial complementary log log likelihood with offset log t in years. Unknown participants remain in the denominator and count as event free. That coding pulls the contrast toward the null and produces the 1.93, 1.88, and 1.79 figures above; below risks of 5 percent the bias is under 2 percent. The contrast estimates a log hazard ratio under proportional hazards within the trial. The log time offset equates arms with different follow-up only under a constant hazard; flag those arms."

Minor

1. Duplicate O22C label (protocol.qmd, Composite definitions). Use: "O22 (50 percent eGFR decline), O22B (40 percent), O22C (kidney failure, sustained doubling of serum creatinine, or kidney death), O22D (any composite including cardiovascular death), and O22E (composites including albuminuria or smaller declines without cardiovascular death) are never pooled with each other. A composite that includes cardiovascular death is O22D even when it also includes albuminuria or a smaller eGFR decline."

2. Contribution weights skip cumulative incidence (protocol.qmd, Confidence in cumulative evidence). After a reported log hazard ratio, the rule rebuilds every other trial from arm counts. A trial that entered through a cumulative incidence then gets weights from different data. Use: "Where the primary model used a cumulative incidence, the contribution network uses that same log(minus log(1 minus p)) contrast and its delta method variance. Where it used arm counts, derive the complementary log log estimate from those counts, adding 0.5 to every cell of a trial with a zero event arm, an all event arm, or a cumulative incidence of 0 or 1. Where it used a reported log hazard ratio, use that log hazard ratio."

3. Compatible follow-up has no bounds (protocol.qmd, Frequentist network meta-analysis). The only numeric definition, 20 to 104 weeks, was removed from Event outcomes and harms. The frequentist odds ratio sensitivity still says "compatible follow-up." Restore "restricted to trials whose follow-up is 20 to 104 weeks."

4. Recorded harms absolute differences are unnamed (protocol.qmd, Event outcomes and harms; Confidence in cumulative evidence). That analysis reports a complementary log log contrast with no hazard ratio interpretation, and "recorded risk differences over the median observation period," never at a fixed horizon. Imprecision for non-rare harms uses the absolute scale at the horizon. The median, the baseline, and the scale name are unspecified. Use: "The recorded harms coefficient is the log relative rate of recorded first events. It is not a hazard ratio. Its absolute difference is the difference in predicted recorded risks at the median of the contributing trials' observation periods, using the reference arm proportion rescaled to that median under a constant hazard. Imprecision uses this difference against the outcome's absolute threshold. Rare outcomes use the numerical range 1/1.25 to 1.25 on this recorded event rate ratio, and the rating states the scale."

Checks that held

There are still 77 outcomes. Every windowed outcome and population pair has one primary window and no numeric overlap. O36A, O36B, and O120 have no week window, and W00 omits O36A and O36B. Turner mappings cover every log odds, log hazard, complementary log log, and log rate outcome. Every mean difference and standardized mean difference has a reference standard deviation, including O24B and O24C at s_y = 1. Nonrandomized flags match harms_nrs.csv. Node classes resolve, and the switching row is a template. Of the authoritative tables, only outcomes.csv, nodes.csv, and amendments.csv (A004) changed from 0.4.0. The concepts file and R/search_strategy.R still render the four sources. The 2026-10-12 harvest remains a pipeline test of concepts 1.0.0 on three sources.

Pre-registration conditions

Team names, affiliations, and postal address; funding and competing interests; PRESS 2015 by a human information specialist; confirmation of Embase access, with omission stated as a limitation if access is absent; verification of Chinese terms by a Chinese-reading team member; locking of provisional thresholds by clinical advisors and patient partners before any outcome data are extracted; an OpenAlex API key; and a logged four source harvest of the registered concepts version, with reconciled counts, before the baseline search. Absence of these alone does not withhold acceptance.

Judgment

Six of the seven round 4 comments are resolved. G4-2 needs one parenthesis deleted. The harms revision is the right structure: a hazard ratio analysis with a known status gate, a separately rated recorded analysis, and competing risks kept off a Kaplan-Meier complement. Two rules in that structure still do not determine a unique analysis: which time supplies the 95 percent counts, and which network holds a cause-specific versus a subdistribution hazard ratio. Those are paragraph fixes, as are the four minor edits. With those edits and the administrative conditions completed, the methods are acceptable for registration.

VERDICT: MINOR REVISION