To investigate the effectiveness of modern machine learning techniques at predicting 2-year disease outcomes for multiple sclerosis (MS) patients with only simple clinical signals.

Previous attempts at MS forecasting used simpler models and had only been successful with access to MRI data, which can be costly and time-consuming to obtain. Modern deep neural networks process sequences of clinical signals from years of standard medical appointments. We explored the efficacy of these intricate models at predicting disease outcomes in a data-impoverished setting.

Rhode Island Hospital provided data which, after filtering out subjects that had no diagnosis date and less than three years of data, included 3,848 visits of 705 unique patients. We trained three different machine learning ensembles comprised of neural networks and traditional machine learning methods. The first of these ensembles predicted future expanded disability status scale (EDSS) scores, while the other two ensembles labeled patients as ‘worsening’ or ‘non-worsening.’

The regression ensemble accomplished a mean absolute error of 0.51 points and an accuracy of 50.0% on test data. The first of the classifier models accomplished 63.0% accuracy, with 33.3% positive predictive value (PPV) and 23.3% sensitivity for the ‘worsening’ class. The final classifier ensemble achieved an accuracy of 67.6%, with 32.4% PPV and 35.9% sensitivity for the ‘worsening’ class. The top predictors were EDSS score, disease duration, and MS type, while previous trends were found to be insignificant.

In cases where MRI data is unavailable, this research shows the strength of a powerful neural network at providing insight into an MS patient’s future. Recurrent architectures bolster traditional ML models by finding different representations of patient data that can better pick out progressive cases. However, statistical power requires more data outside of clinical assessment. Future work will involve building a complimentary network that processes MRI data.