Background: Parkinson's disease exhibits great heterogeneity in the course of the disease. This heterogeneity complicates the treatment of patients and increases the number of subjects needed for studies to test new, potentially neuroprotective drugs. In addition, different types of disease progression may require different therapeutic approaches.

Objectives: Identification, characterization and prediction of Parkinson's progression types based on multimodal, longitudinal data from Parkinson's disease patients.

Research question: Can novel AI methods from the field of machine learning be used to robustly identify Parkinson's progression types? Can the progression types of patients be predicted from baseline data? How generalizable are the results across different cohorts and in relation to different data modalities such as motor and non-motor symptom scores, imaging and digital biomarkers?

Methods:
To enable robust identification of PD progression types, we analyzed multimodal longitudinal data from three large PD cohorts. Patients were synchronized on a uniform time scale of disease progressionusing a latent-time joined mixed model (LTJMM). Subsequently, progression types were identified usingvariational deep embedding with recurrence (VaDER). The identified subtypes were characterized using clinical scores, DaTSCANs, and digital gait assessments across the three cohorts. Several predictive models were developed to assign patients to progression subtypes based on baseline data. We performed a comprehensive cross-cohort validation.

Results:
In each cohort, a fast-progressing and a slow-progressing progression type was identified. Different progression patterns existed with respect to motor and non-motor symptoms, survival, response to dopaminergic medication, DaTSCAN imaging and digital biomarkers of gait assessment. The corresponding progression types could be predicted from baseline data with a ROC-AUC of 0.79. Further simulations showed that increasing the proportion of rapidly progressive patients based on the predictive models can reduce the required cohort size of clinical trials by approximately 43%.

Conclusions:
Our results show that the heterogeneity of Parkinson's disease can be explained by two different Parkinson's progression types. Both subtypes showed consistent patterns of progression across all three cohorts. Our results are largely consistent with the brain-first vs. body-first concept, which could thus provide a biological explanation for the differences between the subtypes. The predictive models developed will enable future clinical trials with a smaller number of subjects by including more rapidly progressing patients in these studies.