The relevancy scores shown on articles are generated by our machine learning (ML) algorithms that analyze the content of each paper. These predictions help identify papers that are most likely to be relevant for Multiple Sclerosis (MS) research and treatment.
We plan to extend our ML Models to work with more areas other than MS.
ML Prediction Scores
Each badge represents a prediction score from our ML models, ranging from 0 to 1:
- 0.0 - 0.3: Low relevance
- 0.3 - 0.7: Moderate relevance
- 0.7 - 1.0: High relevance
Only the most recent prediction from each algorithm is shown for each article.
Our Algorithms
We use three machine learning models to analyze scientific papers:
- LightGBM: A gradient boosting framework using tree-based learning algorithms.
- LSTM model: A long short-term memory neural network for sequence analysis.
- PubMed BERT model: A BERT-based model trained on biomedical literature.
Each model examines the title and abstract of the papers to determine relevance.
As with all machine learning algorithms, the accuracy depends on the training data and the expertise of whoever annotated the relevance labels.
It is important to mention that our training data was annotated by a patient and not by someone with formal knowledge in neurology. This is a major caveat that we are actively trying to overcome.
Last Updated: 2026-05-18
