Predicting 5-Year EDSS in Multiple Sclerosis with LSTM Networks: A Deep Learning Approach to Disease Progression

Backrounds: Multiple Sclerosis (MS) is a neurodegerative disease that is common worldwide, has no definitive cure yet, and negatively affects the individual's quality of life due to disease-related disability. Predicting disability in MS is difficult because of the complex nature of the disease. The primary goal of treating individuals with MS is to prevent or reduce irreversible neurological damage throughout the therapeutic course. Considering the importance of predicting disability MS in the early stage, in this study, we aimed to predict the 5th year score of the Extended Disability Status Scale (EDSS), which is used to measure disability levels in MS patients and allows for a comprehensive assessment of neurological functions. For this purpose, Long Short-Term Memory (LSTM), a special type of Recurrent Neural Network (RNN), designed specifically to analyze data and learn long-term relationships, was used in our study. Methods: The cohort consists of demographic and clinical variables of 1000MS patients, collected from two centers through the MSBase database. The variables used in the study were obtained from the first clinical diagnoses of MS patients during their visits in the first year (1st year) and from their follow-up visits 24 months later (2nd year) and 60 months (5th year). These variables were used as input vectors for training the LSTM model, and 5th year EDSS scores were predicted. Additionally, two different optimization methods were applied to improve the prediction performance of the LSTM model. The RMSE was used as a metric to determine the prediction performance of the model. Results: For the first LSTM model developed using all variables in the dataset, the RMSE on the test data was obtained as 1.46. After hyperparameter optimization and feature selection, the prediction error decreased to 1.332. In addition, according to the heat map feature selection results, age, pyramidal, cerebellar, sensory, and bowel-bladder function variables were determined as the five most important variables in predicting the 5th year EDSS. Conclusions: Our results showed the effectiveness of LSTM deep learning models in predicting EDSS scores for MS patients. Unlike existing studies, our approach integrates both static and dynamic data from MS patients, leading to accurate predictions of EDSS scores ranging from 0 to 10 with minimal prediction error.