Background:
Parkinson's disease (PD) exhibits significant heterogeneity in disease progression. This variability not only complicates treatment of patients but also necessitates larger sample sizes for clinical trials investigating novel, potentially neuroprotective drugs. Moreover, different PD progression subtypes may require different therapeutic approaches.

Objectives:
Our aim was to explain observed heterogeneity in PD by identifying distinct PD progression subtypes through an AI-based methodology. Utilizing a big data approach, our objectives were to identify, characterize, and predict these PD progression subtypes from large multimodal, longitudinal datasets. We aimed to demonstrate the generalizability of our findings by validation on several external PD cohorts, considering various data modalities, comprising clinical data, imaging data and digital biomarkers. Finally, we assessed how these findings could enhance clinical trial designs.

Methods:
To enable robust identification of PD progression types, multimodal longitudinal data from three large longitudinal cohorts were analyzed. Patients were synchronized on a common disease time scale using a latent time joint mixed-effects model (LTJMM). Progression subtypes were then identified using a Variational Deep Embedding with Recurrence (VaDER). Identified subtypes were characterized across the three cohorts using clinical scores, survival data, DaTSCANs, and digital gait assessments. Several predictive models were developed to assign patients to progression subtypes using baseline data. We conducted extensive cross-cohort validation to demonstrate the generalizability of our methodology.

Results:
A fast-progressive and slow-progressive progression subtype were identified in each cohort. These subtypes were reflected by different progression patterns of motor and non-motor symptoms, survival, response to dopaminergic medication, survival, DaTSCAN imaging, and digital biomarkers of gait assessment. The corresponding progression subtypes could be predicted with a ROC-AUC of 0.79. Further simulations showed that increasing the proportion of fast-progressive patients based on the predictive models can reduce the required cohort size of clinical trials by approximately 43%.

Conclusions:
Our analysis reveals that the heterogeneity in PD can be explained by two distinct PD progression subtypes. Both subtypes exhibit consistent progression patterns across all three cohorts. Our findings are consistent with the brain-first vs. body-first concept, which provides a biological explanation for the differences observed between the PD progression subtypes. In the future, the developed predictive models will enable clinical trials with smaller numbers of subjects by including more fast-progressive patients in these studies.