Machine learning model can predict PIRA in first years of MS diagnosis
Study identifies factors that signal disease progression independent of relapses
Written by |
- A machine learning model predicts progression independent of relapse activity (PIRA) in early multiple sclerosis.
- Key predictors for PIRA include age at symptom onset, early disability scores, and time to first evaluation.
- This tool could help tailor treatment and improve long-term prognosis for people with multiple sclerosis.
A new machine learning tool can now help predict whether people newly diagnosed with multiple sclerosis (MS) will experience disability worsening that occurs even in the absence of relapses.
According to a recent study, these artificial intelligence algorithms can identify the risk of “progression independent of relapse activity,” or PIRA, using simple clinical data typically collected during routine doctor visits.
By analyzing factors such as a patient’s age at symptom onset, disease duration, and early disability scores, researchers were able to predict which individuals were more likely to experience PIRA within the first three years of their diagnosis.
“Our results support the feasibility of applying [machine learning] techniques to determine the prediction of PIRA in early stage [MS patients], using only data generally available in clinical practice,” researchers wrote. “These results would allow subsequent tailoring of treatment and definitely a better long-term prognosis.”
The study, “Machine Learning Analysis Applied to Prediction of Early Progression Independent of Relapse Activity in Multiple Sclerosis Patients,” was published in the European Journal of Neurology.
Understanding the progression of MS
In most people with MS, the disease is marked by relapses or flares where symptoms suddenly worsen, followed by periods of remission where symptoms ease.
To some extent, disability worsening in MS can be driven by symptoms that persist even when the relapse has resolved. However, modern research indicates that most disability progression in MS occurs due to gradual worsening that happens even when patients aren’t experiencing relapses. This is known as progression independent of relapse activity.
Currently, no reliable methods can predict the likelihood of PIRA in people newly diagnosed with MS. To address this, a team led by scientists in Italy investigated whether machine learning could be used to predict near-term PIRA risk.
Machine learning is a form of artificial intelligence that works by feeding a computer a large dataset, along with mathematical rules (algorithms) that the computer uses to identify patterns in the data. The computer can then apply those patterns to make sense of other data.
For this analysis, the researchers used clinical and demographic data from 719 people with MS who underwent routine assessments over the first three years after MS onset. After three years, 13% of the patients had experienced PIRA.
To evaluate the accuracy of their machine learning models, the researchers calculated a statistical measure called the area under the receiver operating characteristic curve, or AUC. This measure assesses how well a test distinguishes between two groups (i.e., PIRA vs. no PIRA). AUC scores range from 0.5 to 1, with higher values indicating greater accuracy.
The researchers tested multiple machine learning models. The best-performing model, a Random Forest model, achieved an AUC of 0.75, indicating relatively good accuracy. In this model, the factors most important for accurate predictions were patients’ age at symptom onset, disability scores at two years, and the lag between symptom onset and the first evaluation.
Notably, MRI data contributed comparatively little to prediction accuracy.
Refining predictions for specific patient groups
The scientists also showed that their machine learning model could achieve higher accuracy by narrowing down subsets of patients. For example, when limited to patients ages 45 or younger, the AUC was slightly higher at 0.77. And among patients who didn’t show evidence of disease activity — meaning no relapses, no MRI activity, and no disability worsening — in the first two years, the model yielded an AUC of 0.8.
“Our study supports the feasibility of applying [machine learning] techniques in predicting PIRA in [people with] MS in clinical routine, with a good level of accuracy,” the team concluded.
The scientists stressed that more research is needed to validate and refine this approach, but said that this type of machine learning analysis may one day be used to help predict outcomes and guide treatment decisions for people with MS.
