Heterogeneous disease
Parkinson’s manifests across dopaminergic, cholinergic, cognitive, and motor domains. Capturing heterogeneity demands multimodal evidence and robust harmonization.
Our methodology for progressive Parkinson’s intelligence
Progressive AI harmonizes imaging, diffusion, biospecimen, wearable, and clinical data into interpretable latent spaces and clinician-facing tools. Four domain thrusts build high-quality modality pipelines, while two workflows transform those signals into decision-ready intelligence.
This page details how the program is structured, the workflows we operate, and how thrusts and workflows reinforce one another.
Why this approach
Episodic clinical scales
Clinic visits miss day-to-day variation. Imaging, diffusion, and wearables fill the gaps and provide quantitative anchors for precision interventions.
Actionable translation
Insights must reach neurologists with context. Our workflows produce interpretable, auditable outputs that connect data science to bedside action.
Program architecture
Domain thrusts
Modality-focused teams deliver reliable feature extraction for imaging, diffusion, wearable-cognitive phenotyping, and mechanism inference. Each thrust enforces rigorous quality control, provenance tracking, and quantitative validation.
Visit the Domain Thrusts page for full briefs, figures, and cohort specifics.
Operational workflows
Workflow 1 builds generative latent spaces that align modalities; Workflow 2 delivers interpreted insight to clinicians via ActionIntel, Free Motion, and sensor-clinical correlation dashboards.
Explore both on the Workflows page.
Workflow snapshots
Workflow 1 — Generative Latent-Space Modeling
Harmonizes imaging, diffusion, and clinical data into interpretable latent spaces that seed state estimators and policy agents.
Workflow 2 — Clinician-Centered Decision Pipeline
Delivers multimodal inferences to neurologists via interactive visualization, case review, and cohort monitoring tools.
Thrust contributions
Thrust 1 — Imaging Biomarkers
Automated DaT-SPECT and T1 MRI pipelines generate harmonized biomarker matrices that sharpen CSFSAA stratification and feed downstream agents.
Thrust 2 — Diffusion + Clinical Co-Clustering
Free-water–corrected diffusion metrics align with clinical scales inside a compositional Bayesian co-clustering framework, revealing severity-aligned patient subtypes.
Thrust 3 — Wearable & Cognitive Phenotyping
Wearable IMU data, cognitive batteries, and demographics combine to predict CSFSAA status and uncover multimodal phenotypes.
Thrust 4 — Mechanism Inference
Large-scale statistical testing links multimodal biomarkers to putative biological pathways, grounding discoveries in mechanism-centric hypotheses.
Progressive pipeline overview
- Signal harmonization: Align clinical timelines, imaging acquisition dates, and sensor sampling rates to a single canonical calendar.
- Representation learning: Construct latent spaces via multimodal co-clustering and contrastive encoders that retain interpretability.
- Policy training: Apply progressive reinforcement learning with patient-specific constraints, leveraging recent advances in pointwise control on Hamiltonian manifolds.
- Interactive deployment: Stream inference to ActionIntel, enabling scenario testing, counterfactual analysis, and shared decision-making.
Validation checkpoints
- Imaging pipelines cut manual processing from ~330 minutes to ~12 minutes per subject, enabling cohort-scale DaT-SPECT + T1 analysis.
- Multimodal co-clustering with diffusion metrics yields severity-aligned subtypes and feeds Workflow 1 latent modeling.
- Wearable and cognitive fusion predicts CSFSAA status with AUC 0.93 ± 0.07, complementing invasive assays.
- Mechanism inference verifies 36 claims across genetic, dopaminergic, cholinergic, gait, and cross-pathway domains with reproducible code.
Dig deeper
Read the detailed thrust briefs, walk through each workflow, or explore background context on Parkinson’s neurobiology. Complementary resources live on the Thrusts, Workflows, and Background pages.