(SLB-029) Early Identification of Prodromal Symptoms for Multiple Sclerosis Risk Prediction Using Machine Learning Approaches

Background: The onset of multiple sclerosis (MS) is often preceded by a prodromal phase characterized by early symptoms that may aid in identifying high-risk individuals.

Objectives: This study evaluates the predictive value of prodromal symptoms using electronic health records (EHR) and claims data, highlighting the benefits of incorporating symptom-based terms from EHR into predictive models.

Methods: Optum’s de-identified Market Clarity database, linking EHR with claims data, was analyzed to identify MS patients in the United States from 2016 to 2021, with a retrospective look-back period of three years. Patient selection was based on at least one inpatient or two outpatient claims occurring 90 days apart, using ICD-9-CM and ICD-10-CM codes. A total of 16 comorbidities and 23 signs and symptoms were included from claims and EHR data. Logistic regression (LR), random forest (RF), and XGBoost (XG) models were applied to evaluate associations between prodromal symptoms and MS incidence.

Results: Results from 9,000 patients showed that incorporating EHR-based symptom data significantly improved model accuracy. Predictive performance, measured through AUC scores, ranged from 81% to 92% when EHR symptoms were included, compared to 58% to 70% when using claims data alone. Ophthalmic issues, bladder dysfunction, and mental health conditions emerged as key predictive features. Analysis of time-segmented data revealed that integrating EHR symptoms led to a 40% and 27% AUC gain at the 24–36 months and 0–6 months pre-diagnosis intervals, respectively.

Conclusions: These findings highlight the crucial role of early symptom detection in MS risk prediction. Leveraging machine learning models with EHR-based symptom data demonstrates the potential to improve early identification of at-risk individuals, thereby enhancing timely intervention strategies for MS diagnosis and management.