An Algorithm to Predict New Magnetic Resonance Imaging Lesions to Support Improved Disease Surveillance in Multiple Sclerosis

This paper presents a machine learning algorithm trained on electronic health records of 1,045 multiple sclerosis patients that accurately predicts new MRI lesions to help clinicians personalize the frequency of disease surveillance imaging.

The Gist

Imagine you have a car that sometimes develops a small, hidden engine trouble light. You know it might happen, but you don't know when.

In the world of Multiple Sclerosis (MS), doctors use MRI scans (like high-tech X-rays for the brain) to check for these "trouble lights," which are called lesions. These lesions are tiny spots of inflammation that show the disease is active.

Currently, the rule is: "Check the car every 6 to 12 months, no matter what." This is safe, but it's also a bit like checking your car's oil every single day just in case. It's expensive, time-consuming, and can be stressful for the patient (especially if they are afraid of the tight MRI machine).

The Problem:
We don't know which patients actually need a check-up right now, and which ones are so stable that they could wait a year or two without worrying. We are doing "one-size-fits-all" monitoring, which wastes resources and causes unnecessary anxiety.

The Solution: The "LESION" Algorithm
The researchers at Johns Hopkins built a smart computer program (an algorithm) named LESION. Think of it as a weather forecast for the brain.

Instead of guessing when the storm (new lesions) will hit, the algorithm looks at the patient's "climate data" to predict if a storm is coming soon.

How does the weather forecast work?
The algorithm asks the patient and doctor a few simple questions, similar to checking the barometer:

• The History: Did you have a storm (relapse) recently? Were there clouds (lesions) on the last scan?
• The Shield: Are you wearing a heavy raincoat? (Are you on a strong medication that protects the brain?)
• The Driver: How old is the car? (Patient's age)
• The Mood: How is the driver feeling? (Fatigue, anxiety, depression scores)

The Magic Prediction:
The algorithm crunches these numbers and gives a simple answer: "High Risk" or "Low Risk."

• High Risk: "Hey, the storm clouds are gathering. Let's get that MRI scan ASAP to catch any new trouble."
• Low Risk: "The sky is clear, the engine is running smooth, and you have a great raincoat on. You probably don't need a scan right now. Let's wait and check again in a year."

What did they find?
They tested this on over 1,000 patients.

• The "weather forecast" was surprisingly accurate (about 80% accurate in predicting the future).
• It correctly identified that 68% of the patients were "Low Risk." This means nearly 7 out of 10 people might have been able to skip their next MRI without missing any dangerous new problems.
• It correctly caught the 72 patients who did have new lesions, ensuring they got the care they needed.

Why does this matter?

1. Money: An MRI can cost thousands of dollars. Skipping unnecessary ones saves the healthcare system a fortune.
2. Time & Stress: Patients don't have to sit in the loud, scary machine if they don't need to. It's a huge relief for people with claustrophobia or back pain.
3. Better Care: It helps doctors focus their attention on the patients who actually need it, rather than treating everyone the same.

The Bottom Line:
This isn't a robot replacing the doctor. It's a smart assistant that gives the doctor a "second opinion." It says, "Based on the data, this patient looks very stable. Are you sure we need to scan them today?"

The goal is to move from guessing when to scan, to knowing when to scan, making MS care more personal, less expensive, and less stressful for everyone involved.

Note: This is a new tool that is still being tested and hasn't been fully approved for daily use in every hospital yet, but it shows a very promising future for personalized medicine.

Technical Summary

1. Problem Statement

Multiple Sclerosis (MS) patients require regular Magnetic Resonance Imaging (MRI) scans to monitor disease activity (new lesions) and treatment response. However, current clinical guidelines for scan frequency are non-individualized, leading to:

• Unnecessary imaging: Many stable patients undergo frequent scans that yield no new clinical information.
• Resource strain: High costs (approx. $3,500 per scan), limited access in certain geographic regions, and long wait times.
• Patient burden: Physical discomfort (claustrophobia, mobility issues) and financial strain (copays).
• Lack of precision: Existing recommendations do not account for specific patient subgroups (e.g., older patients, those on high-efficacy therapies) who may have lower risks of new inflammatory activity.

The authors aim to develop a risk-based algorithm to predict the presence of new MRI lesions in a subsequent scan, enabling personalized surveillance intervals.

2. Methodology

Data Sources and Cohort

• Population: 1,131 adults with MS treated at the Johns Hopkins MS Center between 2017 and 2025.
• Inclusion Criteria: $\ge$2 clinic visits and $\ge$1 MRI scan.
• Exclusion Criteria: Patients who switched to a more aggressive Disease-Modifying Therapy (DMT) since their last visit (to avoid confounding baseline changes).
• Demographics: 3:1 female-to-male ratio; mean age 48 (SD 12.3); 80% White, 19% Black, 1% Asian.
• Outcome: 8.8% (100 patients) developed new lesions on their follow-up MRI.

Data Integration & Preprocessing

The algorithm, named LESION (Longitudinal, Equitable Systematic Imaging Operations for Neuroimmunology), integrates data from two sources aligned to an index visit date:

1. MS Performance Test (MSPT): iPad-based assessments. The feature set was reduced to patient-reported outcomes (Anxiety, Depression, Fatigue, and Patient-Determined Disease Steps [PDDS] score) to maximize generalizability outside specific research consortia.
2. MS Smartform (EHR): Clinical data including relapse history, DMT class (categorized as "traditional/moderate" vs. "aggressive/high-efficacy"), and prior MRI lesion counts.
   • Note: MRI records were linked if within a 3-month window of the assessment. Physician assessment of lesions was prioritized over radiology reports when both were available.

Model Architecture

• Algorithm: Logistic Regression.
  • Rationale: Chosen for its interpretability, ease of deployment, and ability to build clinical trust, despite the availability of more complex machine learning models.
• Validation: 5-fold stratified cross-validation (80% training / 20% testing).
• Input Features:
  • Prior MRI activity (presence of new lesions).
  • Relapse history (time since last relapse, number of relapses in prior 2 years).
  • DMT class (Traditional vs. Aggressive).
  • Patient demographics (Age, Sex).
  • Patient-reported outcomes (Anxiety, Depression, Fatigue, PDDS).
  • Continuous variables were z-standardized.
• Threshold Selection: The operating threshold was optimized for sensitivity (to avoid missing active disease) while maintaining high specificity. A threshold of 0.08 was selected in collaboration with MS neurologists.

3. Key Results

Performance Metrics

At the 0.08 probability threshold:

• Area Under the Curve (AUC): 0.80.
• Sensitivity: 0.72 (Correctly identified 72 out of 100 patients with new lesions).
• Specificity: 0.75 (Correctly identified 772 out of 1,031 patients without new lesions).
• Risk Stratification: The model identified 68% of the cohort as low-risk (True Negatives), suggesting they could potentially undergo extended surveillance intervals without missing new inflammation.

Feature Importance

The model identified specific predictors for new lesions:

• Positive Predictors (Higher Risk):
  • Presence of new lesions on the prior MRI (Coefficient $\approx$ +2).
  • Number of relapses in the past two years.
• Negative Predictors (Lower Risk):
  • Use of aggressive/high-efficacy DMTs (Coefficient $\approx$ -2).
  • Older age.
• Non-Significant Features: Visit date, sex, fatigue, anxiety, depression, and PDDS score did not reach statistical significance in this specific model configuration.

4. Key Contributions

1. Development of LESION: A clinically interpretable logistic regression model specifically designed to predict new MS lesions using routinely collected clinical data.
2. Personalized Surveillance Strategy: Demonstrated that a significant portion of patients (68%) could safely extend their MRI intervals, moving away from "one-size-fits-all" guidelines.
3. Interpretability Focus: Prioritized a model that provides clinicians with specific feature contributions (e.g., "low risk due to age and high-efficacy meds") to support decision-making and trust, rather than a "black box" prediction.
4. Equity Considerations: Intentionally excluded race as an explicit classifier variable to enhance robustness and avoid encoding bias, while planning to monitor subgroup performance during future deployment.

5. Significance and Future Directions

• Clinical Impact: The tool acts as a "second opinion" to support clinicians in extending MRI intervals for stable patients, reducing unnecessary costs and patient burden (e.g., for those with claustrophobia).
• Resource Optimization: Allows healthcare systems to redirect MRI resources toward high-risk patients who truly need frequent monitoring.
• Implementation Feasibility: The algorithm relies solely on data available in standard Electronic Health Records (EHR) and routine clinic visits, requiring no specialized imaging pipelines or advanced computational infrastructure.
• Limitations & Next Steps:
  • The study is a retrospective prototype; prospective validation is required.
  • The current threshold (0.08) is study-specific and may need recalibration.
  • Future work aims to incorporate biomarkers (e.g., serum neurofilament light chain) and test the model in external cohorts (e.g., MS PATHS, DISCO) to ensure generalizability.
  • User-centered design is being considered to ensure clinicians can view the specific reasoning behind the algorithm's recommendation.

Conclusion: The LESION algorithm successfully stratifies MS patients by risk of new lesion activity, offering a data-driven pathway to personalized, cost-effective, and patient-centric disease surveillance.