Detecting multiple sclerosis disease activity and progression in progress notes from electronic medical records using natural language processing and machine learning

Multiple sclerosis (MS) phenotypes provide useful disease descriptions but lack complete information regarding the continuing disease process. Disease activity and progression are meaningful modifiers of the MS phenotypes which can further guide prognosis, therapeutic decisions, and clinical trial designs and outcomes, which were not explicitly documented in patients electronic medical records (EMRs). We aimed to detect disease activity and progression in patients with MS from clinical notes in the EMR using Natural Language Processing and Machine Learning models. Using randomly selected progress notes from MS patients at the University of Rochester MS clinic, we integrated NLP and machine learning technologies to predict selected phenotype modifiers that represent disease activity and progression. The method was evaluated by the performance of both the NLP models and machine learning models, as well as the interpretability of the integrated method. We identified 460 progress notes from 287 adult MS patients. The NLP model had an average of 0.92 in precision, 0.87 in recall, and 0.89 in F-score for entity extraction. It had an average of 0.85 in precision, 0.84 in recall, and 0.85 in F-score for entity relation extraction. The sensitivities and specificities of the classification algorithms in predicting phenotype modifiers were: 67% and 93% for predicting modifier "Active", 61% and 82% for predicting modifier "Worsening", 92% and 98% for predicting modifier "Progression", 80% and 94% for predicting modifier "New MRI Lesion", respectively. We showed that the integrated method of NLP with machine learning classification is capable of detecting evidence of disease activity and clinical progression from clinical notes. The classification algorithms yielded interpretable and largely clinically relevant features (symptoms and clinical conditions) that were persistently associated with disease activity and progression. This method holds promise for facilitating the screening of MS clinical trial participants and potentially identifying early evidence of disease progression. Author SummaryDisease activity and progression of disability can be meaningful modifiers to base MS phenotypes which can further impact prognosis, therapeutic decisions, and clinical trial designs and outcomes. However, studies have shown that neither MS phenotypes nor their modifiers are consistently documented in electronic medical record (EMR) chart notes. The evidence for disease activity and progression often resides in the clinical notes, requiring manual chart review from clinical experts and increasing the difficulty of conducting clinical research. In this paper, we developed a generalized information extraction, classification and prediction pipeline, incorporating Natural Language Processing (NLP) technologies and shallow machine learning models, to detect MS disease activity and progression in clinical notes from EMR and to predict phenotype modifiers. Results demonstrated that this integrated method extracts clinically relevant information from progress notes that are persistently associated with disease activity and progression, and predicts MS phenotype modifiers with satisfactory performance, encouraging portability and interpretability. In the future, we aimed to apply the method in this study for facilitating high throughputs of MS clinical trial screening and assessing disease modifying therapy utilization based on disease modifiers.