coreSR.AI
TITAN SR animated ensemble logoT-mark whose horizontal crossbar continuously redistributes between two colors, suggesting a multi-model PICO ensemble (PubMedBERT + BiomedBERT members).TITAN SR
Learn

Systematic Review Screening

Title and abstract screening with the three-member TITAN dual3 ensemble

CoreSR uses BioClinical ModernBERT-large full-context and BioLinkBERT-large head/tail members to screen thousands of references against your selected framework’s eligibility criteria. The framework and field meanings carry from EviTrace or can be set on the screening batch.

How It Works

The screening pipeline in six stages — from literature import to structured data extraction.

1ImportReferencesRIS / CSV / Excel2Set EligibilityCriteriaChoose frameworkEnter criteria3TITAN SRdual3 · 3 modelsfull + head + tailsigmoid meanisotonic calibration4ReviewResultsInclude / ExcludeOverride decisions5Full-TextReviewUpload PDFsTITAN + GLASS6DataExtractionExtract dataMulti-arm support

AI Screening Engine

TITAN dual3: three complementary biomedical language-model views, equal-weighted and isotonically calibrated under one versioned deployment contract.

BioClinical ModernBERT-large

The full-context member reads the complete criteria-conditioned title and abstract in a 2048-token window. It preserves long-range eligibility evidence that shorter biomedical encoders would otherwise truncate.

1 full-context member

BioLinkBERT-large head + tail

Two independently trained 512-token members reserve criteria and title context, then read opposite ends of the abstract. The head view captures methods and study design; the tail view retains results and conclusions.

2 anchored members

Ensemble Scoring

  • Equal-weighted sigmoid-mean of the ModernBERT full-context member and BioLinkBERT head/tail members. Output probabilities are isotonically calibrated against the held-out validation set.
  • Each reference receives a calibrated probability from 0 to 1 of meeting your selected framework’s eligibility criteria.
  • The decision threshold is resolved server-side from your Target Sensitivity setting — references at or above it are included, below are excluded. There is no manual probability cutoff to tune for new runs.
  • TITAN is the sole model used for title/abstract scoring.
  • The same exact dual3 ensemble can run on the workstation GPU, the managed RunPod GPU backup, or the server CPU fallback; responses with a different version or incomplete membership are rejected.

Historical Validation Evidence

Published results (Journal of Clinical Epidemiology, 2026) for the earlier pre-dual3 TITAN-SR configuration. They provide historical context but do not validate the live dual3 replacement ensemble.

MetricValue
Historical training corpus762,934 citation-label pairs from 19,787 completed reviews
Historical external validation22 temporally independent Cochrane reviews; 142,504 records
Historical internal test performanceMedian specificity 0.902 at 99% recall across 3,434 test reviews
Historical external AUCMedian AUC 0.976 on temporally held-out reviews
Historical ASReview comparisonTITAN-SR won 18 of 22 external head-to-head reviews
Historical LLM comparisonOutperformed Claude Sonnet 4, GPT-4o, DeepSeek-V3, Gemini 2.0 Flash, and GPT-4o-mini in the published evaluation
Applicability to live dual3Historical evidence only; these metrics do not validate the replacement ensemble

Note: Do not transfer these historical benchmark values to dual3. The live ensemble uses separately fitted calibration and operating points, and performance still varies by topic and prevalence.

Step-by-Step Guide

A detailed walkthrough of the complete screening workflow.

1

Create a Screening Batch

Give your batch a descriptive name (e.g. "SGLT2 inhibitors in HFpEF — May 2026"). Each batch is an independent screening project.

Tip: You can create multiple batches for different research questions or search updates.
2

Import References

Upload RIS, CSV, or Excel files exported from any bibliographic database (PubMed, Ovid, Scopus, EndNote, etc.). Search itself is handled by your reference manager — CoreSR ingests the export.

Tip: RIS files from PubMed, Ovid, and Scopus are automatically parsed. For CSV, ensure your file has "title" and "abstract" columns.
3

Define Eligibility Criteria

Choose PICO (intervention), PECO (exposure), PICOTS (prognosis), PIRD (diagnostic), CoCoPop (prevalence), PICo (qualitative), or PCC (scoping), then complete the fields shown for that framework. Prognostic reviews also capture prediction timing and setting.

Tip: The selected framework and field meanings carry from EviTrace into screening. You can also add study-design and free-text inclusion or exclusion criteria.
4

Configure Screening Settings

Set your Target Sensitivity — the fraction of truly relevant studies you want to keep. The default (0.995) is the high-recall setting for Cochrane-style comprehensive screening; the server resolves the matching calibrated decision threshold automatically.

Tip: For rapid reviews where specificity matters more than catching every relevant study, lower the Target Sensitivity — each mark shows the expected specificity at that setting.
5

Run AI Screening

Click "Screen" to start TITAN dual3 (BioClinical ModernBERT-large full context plus BioLinkBERT-large head and tail views, equal-weighted and isotonically calibrated). Each reference is scored against your framework-specific inclusion criteria and classified at the threshold your Target Sensitivity resolves to. Progress is shown in real time. CoreSR can fail over between workstation GPU, RunPod GPU, and CPU while enforcing the same exact ensemble contract.

Tip: Screening runs in the background. You can close the tab and return later.
6

Review and Override Results

Browse results sorted by confidence score. Each reference shows its title, abstract, AI decision, and confidence. Override any decision with a single click.

Tip: Focus manual review on references whose calibrated score sits just above the threshold — that's where the AI is least certain. The Review Results tab can sort references by distance-from-threshold.

Supported Formats

Import references from files exported by your reference manager.

FormatExtensionSource
RIS.risPubMed, Ovid, Scopus, EndNote
CSV.csvAny spreadsheet
Excel.xlsx, .xlsAny spreadsheet

Video Tutorial Coming Soon

A walkthrough video demonstrating the complete screening workflow is in production. Check back soon.

Frequently Asked Questions

Common questions about AI-powered screening.

How accurate is the AI screening?
The 2026 TITAN-SR validation paper (Journal of Clinical Epidemiology) reports training on 762,934 citation-label pairs from 19,787 completed reviews and external validation on 22 temporally independent Cochrane reviews. Those figures describe the historical pre-dual3 configuration and are not validation of the live dual3 ensemble. The active model has its own held-out calibration and operating points, but performance still varies by topic and prevalence; manually review borderline references and audit exclusions.
Can I adjust the sensitivity/specificity trade-off?
Yes. The Target Sensitivity control in Screening Settings is the single knob — you tell CoreSR what fraction of truly relevant studies to keep, and the server resolves the matching calibrated decision threshold for you. The default (0.995) is the high-recall setting recommended for Cochrane-style comprehensive reviews; lowering it buys specificity (fewer false positives to triage) at the cost of recall. Each mark shows the expected validation-fold specificity at that sensitivity.
What happens if the AI misses a relevant study?
Every reference is retained in the batch regardless of the AI decision. You can sort by confidence score and manually review low-confidence exclusions. The Review Results tab lets you override any AI decision with a single click.
How long does screening take?
Screening speed depends on the active serving tier and server load. As a conservative end-to-end estimate, a warm GPU path handles about 200–500 references per minute, so 2,000 references can finish in under 10 minutes. The emergency CPU fallback is deliberately slower (about 30 references per minute in current measurements). You can navigate away and return later; screening runs in the background.
Can I screen in languages other than English?
The TITAN ensembles (v6 for paid and trial accounts, dual3 on the free CPU tier) use BioClinical ModernBERT-large and BioLinkBERT-large checkpoints trained primarily on English biomedical literature. Titles and abstracts in other languages may be screened with reduced accuracy. For non-English reviews, we recommend choosing a higher Target Sensitivity and manually reviewing more references.
How is this different from Covidence or Rayyan?
CoreSR runs a criteria-conditioned biomedical ensemble on CoreSR-managed infrastructure rather than relying on keyword matching or a general-purpose LLM. Requests route from the workstation GPU to the managed RunPod GPU backup and then CPU failover, with every tier required to load the same versioned artifact for the generation that scores the batch. This keeps decisions reproducible while CoreSR integrates screening, full-text review, and structured data extraction in one workflow.
Can I export my results?
Yes. You can export included references, excluded references, or the full batch as CSV. The export includes titles, abstracts, authors, DOIs, confidence scores, AI decisions, and any manual overrides.
Can multiple reviewers screen the same batch?
Currently, screening batches are owned by the user who created them. Multi-reviewer support with inter-rater reliability metrics is on the roadmap. In the meantime, you can export and share results via CSV.
Does my data leave the server?
During title/abstract screening, citation text is sent only to CoreSR-managed TITAN infrastructure, not to a third-party general-purpose LLM API. It may be processed on the workstation GPU, the managed RunPod GPU backup, or CPU failover; every tier must load the same bound artifact for the generation that scores the batch. Full-text screening and data extraction are separate, reviewer-controlled stages that can use GLASS on extracted document text.
What if the AI screening engine is unavailable?
CoreSR automatically tries the same bound ensemble generation on its eligible tiers (workstation GPU, managed RunPod GPU, then CPU for TITAN v6; the CPU-only legacy service for the free dual3 tier). A backend reporting a different version or incomplete membership is rejected. If no compatible tier is ready, you will see a notification and the batch will not start; CoreSR never silently switches screening to an unrelated model or general-purpose LLM.

Ready to start?

Create your first screening batch and let the AI handle the heavy lifting.

References

1. Pitre T, Zeraatkar D, Granton J, Busse J, Rochwerg B, Guyatt G. A biomedical BERT ensemble (TITAN-SR) outperformed active-learning and large language model chatbots for systematic review title and abstract screening: development and external validation. J Clin Epidemiol. 2026;112514. doi:10.1016/j.jclinepi.2026.112514

2. Gu Y, Tinn R, Cheng H, et al. Domain-specific language model pretraining for biomedical natural language processing. ACM Trans Comput Healthcare. 2022;3(1):1-23.

3. Moher D, Liberati A, Tetzlaff J, et al. Preferred reporting items for systematic reviews and meta-analyses: the PRISMA statement. BMJ. 2009;339:b2535. doi:10.1136/bmj.b2535

4. Higgins JPT, Thomas J, Chandler J, et al. Cochrane Handbook for Systematic Reviews of Interventions. 2nd ed. Wiley; 2019.

CoreSR — AI-powered evidence synthesis