Persons and affiliations

Date

Description

Aims:
To compare two artificial intelligence (AI)-based patient stratification strategies for reducing sample sizes in randomized clinical trials (RCTs) in Parkinson’s disease (PD): (I) enrichment for a data-driven fast-progressing subtype predicted from clinical data, and (II) enrichment of patients with accelerated cognitive decline using MRI-derived brain age gap (BAG) at baseline.

Methods:
PD progression subtypes were identified from longitudinal trajectories of 409 PD patients using a combination of two AI methods and validated in two external cohorts. Individual subtypes were predicted by baseline and 1-year clinical scores, comparing different machine learning models. BAG was computed from baseline MRI in 451 PD patients from the same cohort, comparing different BAG prediction workflows. For both strategies, we simulated RCTs of a disease-modifying treatment and calculated sample sizes for an enrichment allowing still inclusion of 30% of all patients.

Results:
Two PD progression subtypes were identified, showing distinct progression patterns for several symptom domains with patterns being reproducible across all cohorts. Clinical data predicted subtypes with ROC-AUC 0.79, enabling enrichment of 47% fast-progressing PD patients compared to 18% without stratification. This yielded a 43% sample size reduction compared with unstratified scenarios (Fig. 1a), with similar results replicated in validation cohorts. Higher baseline BAG correlated with faster decline across several cognitive domains. BAG-based enrichment reduced required sample sizes by 28% for a cognitive composite endpoint (Fig. 1b) and on average by 54% for single cognitive endpoints compared with unstratified scenarios.

Conclusions:
AI-enabled enrichment using data-driven subtyping or MRI-derived BAG substantially reduces sample size requirements in clinical trials, providing promising strategies to improve future RCTs and accelerate the development of new treatments in PD.