According to Gram Research analysis of a 2026 scoping review, higher disease activity at the start of treatment was the most common predictor of poor response to methotrexate in rheumatoid arthritis. The review screened 266 articles and included 14 studies that used machine learning to predict treatment response.
Methotrexate is the most widely used first treatment for rheumatoid arthritis, but many people do not respond well to it. A 2026 scoping review, reviewed by Gram Research analysis, looked at 266 articles and kept 14 studies that used machine learning, a type of computer program that finds patterns in medical data, to predict who would benefit from methotrexate. The most consistent finding was that people whose disease was very active when they started treatment were less likely to respond. Other clues included blood test results, a patient questionnaire about daily function, and some genetic markers. The review did not test treatments itself, so it cannot tell doctors what to do differently yet. Its main value is showing where future research on personalized treatment should focus.
Key Statistics
A 2026 scoping review of 14 studies, drawn from 266 articles screened in PubMed and Web of Science, found that higher baseline disease activity measured by DAS28 was the most common predictor of poor response to methotrexate in rheumatoid arthritis.
The same 2026 review reported that other clinical predictors of methotrexate response included Health Assessment Questionnaire scores, lymphocyte count, platelet count, and anti-citrullinated protein antibodies.
Of the 266 articles identified in the 2026 scoping review, only 14 met the inclusion criteria for studies using machine learning to predict methotrexate response in rheumatoid arthritis.
The Quick Take
- What they studied: Researchers wanted to know whether computer programs that learn from patient data can predict which people with rheumatoid arthritis will respond well to methotrexate, and what clues those programs rely on.
- Who participated: This was a review of previously published studies, not a study with new patients. Researchers screened 266 articles and included 14 studies. The total number of patients in those 14 studies was not reported in the review.
- Key finding: The most common warning sign of a poor response was high disease activity at the start of treatment, measured by a score called DAS28 that tracks swelling, tenderness, and inflammation. This was the strongest predictor across the clinical studies reviewed.
- What it means for you: If you have rheumatoid arthritis, your doctor already tracks disease activity, and this review suggests that a high score at the start may mean methotrexate alone is less likely to work well for you. Do not change or stop any medication based on this review. Talk with your rheumatologist about your own results and options.
The Research Details
This study is a scoping review. A scoping review is a way of mapping out what is already known on a topic. Instead of testing a new treatment, the researchers searched two large medical research databases, PubMed and Web of Science, for studies that used machine learning to predict how people with rheumatoid arthritis would respond to methotrexate.
The team read through the search results and found 266 articles. After applying their rules for what counted as relevant, 14 studies were kept. For each one, they recorded the study design, the type of computer method used, where the data came from, and what the study found. Data came from sources such as electronic medical records, routine clinical measurements, and genetic testing.
Because a scoping review does not calculate a single combined result, the researchers summarized the findings in a descriptive way. They described patterns that appeared across studies rather than pooling numbers into one estimate.
Mapping the existing studies helps show which predictors appear again and again and which ones have only been tested once. This makes it easier for future researchers to choose promising directions. It also shows where the field lacks good evidence, such as large studies that follow patients over time.
The search used two major databases and clear inclusion rules, which are strengths. However, scoping reviews do not grade each study’s quality in depth, so the conclusions depend on the quality of the original studies. The review does not report how many patients were included overall, which makes it harder to judge how much evidence sits behind each finding. Readers should view it as a map of the research landscape rather than proof that any single predictor works in practice.
What the Results Show
The clearest result was that higher disease activity at the start of treatment was the most common predictor of poor response to methotrexate in the clinical studies. Disease activity was measured with DAS28, a score that combines joint counts, a blood marker of inflammation, and a patient’s own rating of how they feel. A higher score means more active disease.
The researchers also found other clinical clues. These included scores from the Health Assessment Questionnaire, which measures how well a person can handle everyday tasks, and levels of erythrocyte folate, a form of vitamin B9 in red blood cells. Lymphocyte count and platelet count, two routine blood measures, also appeared as predictors. Antibody tests, including anti-citrullinated protein antibodies and rheumatoid factor, were linked to response as well.
Some studies looked at genetics. The review mentioned specific gene variants, such as ABCB1 rs1045642 and ABCC3 rs4793665, that may affect how the body moves methotrexate in and out of cells. These variants were reported as predictors in some studies, but genetic findings were less consistent across the review than the clinical ones.
The machine learning methods were mostly supervised, meaning the programs learned from patient records where the outcome was already known. Common approaches included random forests, which combine many simple decision trees, and logistic regression, a standard statistical method. These tools can spot patterns, but they are only as good as the data used to train them.
Beyond the main predictors, the review points to a practical idea: combining several routine measures may work better than any one alone. This is a sensible direction, but the review did not directly compare combined models against single measures, so this remains an idea to test rather than a proven result.
The genetic findings matter for future personalized care, since they could eventually help explain why some people process methotrexate differently. For now, they are not used to guide routine treatment decisions.
The finding that high disease activity predicts a weaker response fits with a long-standing view in rheumatology that patients with more active disease at the start often need additional or different treatments. The review confirms that this pattern appears consistently across machine learning studies, not just traditional statistical ones. The inclusion of genetic and blood test predictors also matches the broader push toward personalized medicine in rheumatoid arthritis.
The biggest limitation is that this review summarizes other studies rather than testing patients directly. It cannot prove that any predictor causes a poor response or that a computer model would work well in everyday clinics. The review also does not report the total number of patients across the 14 studies, so readers cannot easily judge how much evidence supports each finding.
The included studies likely differ in their patient groups, methods, and how they defined response, which makes direct comparison difficult. Some findings came from only a few studies, so they may not hold up in larger or different groups of people. Finally, the review searched only two databases, so some relevant studies may have been missed.
The Bottom Line
Research suggests that high disease activity at the start of methotrexate treatment is a common sign that the drug may not work well on its own. Confidence in this finding is moderate, because it appeared consistently across the clinical studies but the review did not measure treatment effects directly. Confidence is lower for genetic markers and for machine learning models as ready-to-use tools. Patients should continue discussing their disease activity scores with their doctor, who can decide whether a change in treatment is appropriate.
People diagnosed with rheumatoid arthritis, their families, and their rheumatology teams may find this useful for understanding why treatment responses vary. Researchers developing prediction tools will find it helpful for identifying promising predictors. People who do not have rheumatoid arthritis, or who take other medications, should not apply these findings directly to their own care.
Methotrexate typically takes several weeks to a few months to show its full effect, so doctors usually reassess response after about three months. Computer-based prediction tools are still in research stages and are not yet common in clinics. Any change in how doctors use these predictors is likely years away and would require testing in larger studies first.
Frequently Asked Questions
What predicts whether methotrexate will work for rheumatoid arthritis?
A 2026 scoping review found that high disease activity at the start of treatment, measured by the DAS28 score, was the most common predictor of poor response. Other clues included the Health Assessment Questionnaire score, blood counts, and certain antibody tests.
Can machine learning predict how well methotrexate will work for me?
Machine learning has been used in 14 studies to look for patterns that predict methotrexate response, but these tools are still in research stages. They are not yet standard in clinics, so your doctor will continue to rely on established measures of disease activity.
Why does methotrexate not work for everyone with rheumatoid arthritis?
Research suggests that people with more active disease when they start treatment are less likely to respond well to methotrexate alone. Genetic differences in how the body processes the drug may also play a role, though this evidence is less consistent.
How long does it take for methotrexate to start working?
Methotrexate usually takes several weeks to a few months to show its full effect. Doctors typically reassess how well it is working after about three months, using disease activity scores and symptom reports.
Want to Apply This Research?
- Log your disease activity score, if your doctor shares it, along with your morning stiffness time in minutes and the number of swollen joints you notice each week. Record these at the same time each week so trends are easy to see. Bring the log to each rheumatology visit.
- Use the app to set a weekly reminder to note your joint symptoms and how well you managed daily tasks, and to take methotrexate on the day your prescription specifies. Share your trends with your care team rather than adjusting your medication on your own.
- Track symptoms and routine lab results over the long term, ideally for at least six months after any treatment change. Look for patterns, such as whether higher activity scores are followed by slower improvement. Review the trends with your doctor every few months, and update the log whenever your treatment plan changes.
This article summarizes a review of previously published research and is for general information only. It is not medical advice and does not replace guidance from your doctor or rheumatologist. Do not start, stop, or change methotrexate or any other medication based on this summary. Machine learning prediction tools described in the research are not yet standard care.
This research translation is published by Gram Research, the science division of Gram, an AI-powered nutrition tracking app.