Objective: To identify Parkinson’s disease (PD) subtypes that demonstrate a specific progression pattern, using study data from the Parkinson’s Progression Markers Initiative (PPMI) cohort.

Background: Subtyping is an important problem in PD research, and previously proposed subtypes have been based upon particular recognizable but limited characteristics, for example, motor (tremor dominant (TD) postural instability gait disorder (PIGD); intermediate), cognitive (no impairment; mild impairment; dementia), mood (anxiety; depression; depression-anxiety; normal), RBD, et cetera. We aimed to identify comprehensive PD subtypes across different domains, using data-driven methodologies.

Methods: We concatenated records for each subject into a sequence according to their associated timestamps. Then we trained a recurrent neural network model called Long Short Term Memory (LSTM) on those sequences and embedded them into a latent homogeneous sequence space. Patient similarity was evaluated on those latent sequence embeddings, and PD subtypes were identified through clustering with the learned patient similarities.

Results: 466 patients with idiopathic PD were included. Three distinct subtypes were identified based upon data from baseline to 6 years follow up. Subtype I comprised 43.1% subjects, with average age 58.8 years, characterized by moderate motor decay (average Hoehn and Yahr (H&Y) scale 1.41 to 1.88) but stable cognitive ability. Subtype II comprised 22.9% subjects, with average age 61.9 years, characterized by mild functional decay in both motor and non-motor symptoms (average H&Y scale 1.52 to 1.66, average MoCA score 27.26 to 27.09). Subtype III comprised 33.9% of the patients, with average age 65.3 years, characterized by rapid progression of both motor and non-motor symptoms (average H&Y 1.61 to 2.15, MoCA score 26.63 to 24.41). Subtypes I and II contain more TD than PIGD subjects at baseline, while Subtype III was dominated by PIGD subjects., and had the greatest conversion from TD to PIGD, and from no cognitive impairment to mild cognitive impairment. Moreover, all subjects with dementia were in Subtype III.

Conclusions: Machine learning methods demonstrate great potential for detecting comprehensive PD subtypes across both motor and non-motor modalities, and also capture temporal dynamics of progression patterns.

« Back to 2018 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/identification-of-comprehensive-pd-subtypes-using-ppmi-study-data-with-recurrent-neural-networks/