Retract
Screen your paper’s retraction risk & pattern

How to use Retract

1. Paste only the abstract into the text box (no title, no keywords, no full text).
2. Choose a model: Auto (recommended), Cardiovascular, Oncology, General medicine, or AI/ML & Digital Health (if enabled).
3. Choose a threshold (slider). Higher threshold = stricter flagging.
4. Click Check.
5. Get result: the score is the probability (%) that the abstract shows patterns similar to recently retracted original research papers in the selected domain. (In Auto mode, if the router is uncertain, it may fall back to a configured default branch.)
6. Cross-domain check shows how the same abstract scores across the available domain models.
7. Most similar journals shows the top journal-style matches with percentages (currently 21 journals are added):
   • NEJM
   • The Lancet
   • BMJ
   • Circulation
   • JAMA
   • Nature portfolio
   • Nature Medicine
   • Science
   • Cell
   • Frontiers in Psychology
   • Journal of Clinical Oncology
   • European Heart Journal
   • PLOS Medicine
   • EJVES
   • BMC Medicine
   • BMC Public Health
   • BMC Cancer
   • BMC Infectious Diseases
   • MDPI Cancers
   • MDPI Journal of Clinical Medicine
   • Cureus
8. Use the AI explanation to strengthen your abstract: review the highlighted signals, apply the suggested clarity improvements, and (if provided) compare the rewrite suggestion with your original text. (Always verify factual accuracy; do not add results that are not in your data.)

Ads are shown to free users. You can remove ads for this app by spending 10 credits (Account → Remove ads).

About Retract

Retract is a lightweight screening tool that estimates how strongly an abstract resembles recently retracted biomedical papers. It provides domain models and (optionally) a journal-style classifier.

• Cardiovascular model: fine-tuned BERT text-classification model trained on 600 retracted cardiovascular research articles (2021–2026) from the Retraction Watch Database, using cardiovascular, vascular, interventional, venous, dialysis, stroke, coronary, thrombosis-related keywords. The control set consists of top-level publications matched by similar keywords from PubMed.
• Oncology model: fine-tuned BERT text-classification model trained on 1200 retracted oncology abstracts (2021–2026) from the Retraction Watch Database, collected using oncology keywords such as cancer, carcinoma, leukemia, myeloma, and melanoma. The control set consists of matched oncology publications from PubMed from the same years (2021–2026).
• General medicine model: fine-tuned BERT text-classification model trained with the same approach as the other domains: retracted abstracts were extracted from the Retraction Watch Database using general/internal medicine keywords, and a matched control set was collected from PubMed from the same years. Keywords include (as used in the extractor): “hepatitis OR cirrhosis OR pancreatitis OR endoscopy OR colonoscopy OR ulcer OR reflux OR irritable bowel OR gallstones OR cholecystitis OR appendicitis OR diverticulitis OR hemorrhoids OR constipation OR diarrhea OR pulmonary OR alcoholic OR ascites OR encephalopathy OR gastroenteritis OR colitis OR Crohn OR polyps OR drainage OR abscess OR depression OR anxiety OR traumatic OR schizophrenia OR bipolar OR substance OR opioid OR insomnia OR dementia OR epilepsy OR migraine OR pain OR fibromyalgia OR obsessive OR eating OR autism OR transplantation OR immunosuppression OR sepsis OR hypothyroidism OR hyperthyroidism OR osteoporosis ...”
• AI/ML & Digital Health model: a separate fine-tuned BERT text-classification model trained on ~700 retracted AI/ML & digital-health abstracts (2021–2026) from the Retraction Watch Database, collected using keywords such as ai, ml, artificial intelligence, machine learning, deep learning, computer vision, ehr, iot, blockchain, virtual reality, and extended reality. The control set consists of matched publications from PubMed from the same years (2021–2026). This branch is used only if it is enabled and configured on the server.
• Journal-style model: a multi-class BERT classifier trained on ~4,200 abstracts (2024–2026) across 21 journals (200 abstracts per journal). It estimates which journal’s writing style the abstract most resembles and returns the top 5 most similar journals with match percentages.

Inspiration: Retract was inspired by the BMJ methodological study doi:10.1136/bmj-2025-087581. The authors report that a similar screening model is integrated into the online submission systems of three journals from a major publisher and is being used to screen cancer-related manuscripts.

The screening models are hosted on Hugging Face. In Auto mode, Retract uses a deterministic router to choose the most likely domain (cardiovascular vs oncology vs general medicine; optional AI/ML branch if enabled). If the router is uncertain, it can fall back to a configured default branch. The journal-style model is separate and is used only when enabled on the server.

AI explanation: After screening, an LLM can provide a short interpretation of why the abstract may score higher or lower, plus practical clarity improvements. If the similarity score is ≥ 20%, it also proposes a rewrite suggestion (without adding new facts). If the abstract appears outside the selected domain, a scope note may warn that interpretation can be less reliable.

Model training: AutoTrain (Text Classification).
Validation metrics (cardiovascular): loss 0.346 · F1 0.862 · precision 0.955 · recall 0.785 · AUC 0.920 · accuracy 0.875
Validation metrics (oncology): loss 0.200 · F1 0.954 · precision 0.960 · recall 0.948 · AUC 0.984 · accuracy 0.956
Validation metrics (general medicine): loss 0.259 · F1 0.911 · precision 0.885 · recall 0.939 · AUC 0.973 · accuracy 0.910
Validation metrics (AI/ML & digital-health): loss 0.364 · F1 0.852 · precision 0.943 · recall 0.777 · AUC 0.925 · accuracy 0.867
Validation metrics (journal-style): loss 0.071 · F1(macro) 0.985 · accuracy 0.985

Part of ML in Health Science © .

Version

• 1.5 (08.02.2026): Added General medicine model + updated About/metrics.
• 1.4 (08.02.2026): Added journal-style model (top-21 “Most similar journals” match).
• 1.3 (07.02.2026): Added AI/ML & Digital Health branch support.
• 1.2 (04.02.2026): Added multi-domain routing (Auto / Cardiovascular / Oncology) and oncology screening branch. Warm-up supports multiple endpoints.
• 1.1 (02.02.2026): Added AI explanation (LLM interpretation + scope note for out-of-field abstracts + rewrite suggestion when score ≥ 20%).
• 1.0 (31.01.2026): Initial release (abstract screening + MLHS shared login/credits + ad-free entitlement).

If you report a bug, include a screenshot + browser/device details.