AI4PD · AI for Parkinson's Disease
Parkinson's is the fastest-growing neurodegenerative disease, and it reaches well beyond tremor: into gait and balance, sleep, mood, autonomic function, and cognition. The evidence about any one patient arrives scattered across clinic visits, imaging, biomarkers, wearables, genetics, and the electronic health record. AI4PD assembles that scattered evidence into one living, uncertainty-aware model of the patient and turns it into clinician-facing support for differential diagnosis, progression forecasting, and therapy planning. Every output goes to the physician first.
Oden Institute · The University of Texas at Austin · CVC Lab · TACC infrastructure · Supported by the Michael J. Fox Foundation
Evidence to date (retrospective)
Four preprint and in-review studies on established Parkinson's cohorts (PPMI, BioFIND, PDBP, FoxInsight). The headlines are below; full cohorts, methods, and limitations are on the Evidence page. These results are retrospective and hypothesis-generating, the calibrated substrate the twin is built on, not the mechanism claim itself.
Calibrated motor states · PPMI + BioFIND
Across 29,366 PPMI visits from 4,773 patients, motor states stay stable on average yet 25.5% of patients shift over time, and the model discloses its overconfidence (0.989 nominal vs 0.849 empirical) instead of hiding it.
Details →Trial enrichment · four cohorts
Across 62,543 patients, a 16.6% rapid-progressor subgroup declines about +4 MDS-UPDRS-III points a year, a candidate trial-enrichment signal that still needs prospective validation.
Details →Genetics + wearables · medRxiv
LRRK2 G2019S carries a 1.92× PD prevalence ratio and +4.35 motor points; wearable arm-swing asymmetry (27%) and a risk model (AUC 0.717) add scalable digital signal.
Details →Circuit imaging · bioRxiv
Imaging anchored to six brain circuits (294 PPMI) ties nigrostriatal integrity to motor burden and sensory/visuospatial integrity to cognition, FDR-controlled, with modest effect sizes.
Details →Underneath those results sits the patient-specific digital twin. Its interpretable core is a structured prior for how a patient's state evolves — functional reserve, coupling between subsystems, dissipation, and therapy ports — not a claim of physical energy conservation, and it updates as new visits, sensors, and biomarkers arrive.
On top of the twin, multi-agent diagnostic and therapy-planning agents reason over a dynamic knowledge network, surfacing disagreement, uncertainty, and gaps rather than smoothing them over. Letting a clinician simulate a candidate DBS change in the twin before changing patient settings is a prospective-validation hypothesis, not a current capability — the twin does not prescribe DBS candidacy or targets.
AI4PD is decision support, never an autonomous prescriber, and its hardest constraint is not the model. It is the multi-institution longitudinal evidence layer needed to validate it. We are looking for movement-disorders neurologists and institutions that govern longitudinal Parkinson's cohorts to partner on clinical data and prospective validation, on terms that meet your standard of proof.
