Background: Progressive supranuclear palsy (PSP) is a rare neurodegenerative disorder that is difficult for primary care physicians to recognize due to its progressive nature and similarities to other neurologic disorders. This case-control study aimed to identify clinical features observed in general practice associated with a subsequent diagnosis of PSP. Methods: We analyzed a de-identified dataset of 152 PSP cases and 3,122 matched controls from electronic medical records of general practices in Germany. We used a random forests algorithm based on machine learning techniques to identify clinical features (medical conditions and treatments received) associated with pre-diagnostic PSP without using an a priori hypothesis. We then assessed the relative effects of the features with the highest importance scores and generated multivariate models using clustered logistic regression analyses to identify a subset of clinical features associated with subsequent PSP diagnosis. Results: Using the random forests approach, we identified 21 clinical features associated with pre-diagnostic PSP (odds ratio ≥2.0 in univariate analyses). From these, we constructed a multivariate model comprising 9 clinical features with ~90% likelihood of identifying a subsequent PSP diagnosis. These features included known PSP symptoms, common misdiagnoses, and 2 novel associations, diabetes mellitus and cerebrovascular disease, which are possible modifiable risk factors for PSP. Conclusion: In this case-control study using data from electronic medical records, we identified 9 clinical features, including 2 previously unknown factors, associated with the pre-diagnostic stage of PSP. These may be used to facilitate recognition of PSP and reduce time to referral by primary care physicians.
Keywords: case-control study; electronic medical records; epidemiology; general practice; machine learning; progressive supranuclear palsy.
Copyright © 2021 Kwasny, Oleske, Zamudio, Diegidio and Höglinger.