Pinned Tweet
Which AI should screen 10,000 abstracts?
Chatbots write one word at a time. BERT reads both ways and gives a score you can tune for high recall.
TITAN-SR kept 99.8% of included studies in 22 later Cochrane reviews.
sciencedirect.com/science/ar…
coresr.ai
#SystematicReview #Epi
A new review mapped 55 quantitative bias analysis methods for summary-level data. Only 4 were rated applicable to meta-analyses of observational studies, and 9 to hazard ratios.
So when a review pools adjusted HRs, what does a line about "residual confounding" actually buy us?
Correia, Wallach et al. built applicability matrices (confounding 28, information bias 21, selection 6) and tried them on 3 published studies. Open access: doi.org/10.1186/s12874-026-0…
#SystematicReview #MetaAnalysis #EpiTwitter
New this week at CoreSR: TITAN-SR v6, ROB BERT v6 and GLASS v3, with what the evidence does and doesn't yet show for each.
Weekly update 👇
drtylerpitre.substack.com/p/…
Launching SynthesisBench: an open screening benchmark across 85 systematic reviews, 2,327 records and 26 systems. Top Youden's J: TITAN-SR v6 0.879, Claude Opus 5.5 0.831, GLASS-SR 0.827. Run by CoreSR. synthero.ai/screening
Youden's J = sensitivity + specificity − 1. Intervals: 5,000 whole-review bootstrap replicates. #SystematicReview #EvidenceSynthesis #AI
A living review is only as current as its slowest step. A new trial lands: screen it, extract it, appraise it, rerun the model, re-rate certainty.
If those steps sit in five tools, the update stalls. In CoreSR they sit in one project, so each update starts where the last ended.
TITAN screening, GLASS quote-grounded appraisal, real-R meta-analysis and NMA, native GRADE. 7-day trial at coresr.ai, from $39 CAD/mo.
#LivingReview #SystematicReview #EvidenceSynthesis
Cluster RCT in your meta-analysis, no ICC reported? You borrow one.
Bhat & Beretvas (Res Synth Methods) simulated pooling ICCs: REML vs RVE barely mattered; the variance formula did. Only the Fisher-transformation weight never gave a substantially biased pooled ICC.
One question: they simulated continuous, normally distributed outcomes. Health cluster trials often use binary ones. Does the Fisher advice carry over? The authors flag it as untested.
Open access: doi.org/10.1017/rsm.2026.100…
#MetaAnalysis #ClusterRCT #EvidenceSynthesis
When two RoB 2 raters disagree, often it isn't judgment. They read different sentences: one stopped at "analysed by ITT," the other found the dropout table.
In GLASS, every RoB call sits next to the quote it rests on. You argue about text, not memory. A human still decides.
GLASS is the quote-grounded appraisal step inside CoreSR, next to screening, extraction, real-R meta-analysis and GRADE. 7-day trial: coresr.ai
#RiskOfBias #SystematicReview #EvidenceSynthesis
RoB 2 papers usually tell us raters disagree. D'Souza et al. (Cochrane Evid Synth Methods) show what one team did about it: 49 results from 34 RCTs of remote alcohol/drug therapies, piloted together, assessed in pairs, every assumption written into a review-specific rulebook.
One rule I keep turning over: missing data imputed as relapse could still be low risk; multiple imputation could only be high, since missingness likely means use.
Fair for relapse. But if dropout differs by arm, "missing = relapse" tilts the effect too. Did it here?
D'Souza P, Macdowall W, Burchett HED, Kwan I, Sutcliffe K. Cochrane Evid Synth Methods 2026, open access. Their Table 1 is worth copying for your own protocol.
doi.org/10.1002/cesm.70107
#RiskOfBias #RoB2 #SystematicReview #EvidenceSynthesis
An NMA has no single certainty rating. Each comparison gets its own: rate direct and indirect evidence, take the higher if they agree, downgrade for incoherence if they don't, then judge imprecision on the network estimate.
CoreSR does NMA GRADE natively, beside the R model.
7-day trial at coresr.ai, from $39 CAD/mo.
#NetworkMetaAnalysis #GRADE #SystematicReview #EvidenceSynthesis
After stroke/TIA, pooling 432 estimates says 74.1% of patients stay adherent. Standardize to one definition (pharmacy data, moderate cutoff, 9–15 mo) and it's 53.7%. Persistence: 76.8% → 48.2%.
Jiang et al. (JCE): how adherence was measured drove more heterogeneity than who.
A high I² can be a definitions problem before it's a patient problem. One question: how many studies sat near the reference spec, i.e. how much of the 53.7% is model extrapolation?
doi.org/10.1016/j.jclinepi.2…
#MetaAnalysis #EvidenceSynthesis
With 30 included studies, recall moves in 3.3-point steps: miss one and you're at 96.7%. So a screener's "99% recall" only means something next to the number of includes behind it.
Small validation sets make recall look perfect or awful. Ask for the denominator.
How we report it for TITAN-SR (JCE, doi 10.1016/j.jclinepi.2026.112514): specificity at a fixed 99% recall, median 0.902 across reviews. The live CoreSR screener is TITAN dual3, which that paper doesn't validate.
7-day trial: coresr.ai
#SystematicReview #Screening