The living twin

One living model per patient, organized by a port-Hamiltonian core

AI4PD is designed to maintain a single longitudinal model per patient and to update it as evidence arrives: a single intake, each clinic visit, a continuous wearable stream, or DBS sessions only. Its dynamics are organized by a port-Hamiltonian core — an interpretable structured prior, not a literal-physics claim.

H — an energy-like storage function

A model-internal latent used as a compact summary of control margin. It is not a measured biological reserve and has no validated clinical referent; its values are meaningful only relative to the fitted model.

J — interconnection geometry

How the motor, autonomic, and cognitive subsystems exchange influence.

R — dissipation

The structural term that makes the storage function decay rather than persist. It is a modeling device for irreversibility, not a measurement of a patient's neurodegeneration rate.

Therapy ports

Where medication, rehabilitation, and stimulation enter the dynamics.

Parkinson's has no conserved physical energy; the structure is an inductive bias, not a recovery of real physics. What it buys is concrete: bounded, stable forecasts under missing modalities, and interventions expressed as port inputs rather than hidden coefficients. We state the identifiability assumptions explicitly and test whether the structure earns its place with planned ablations against calibrated neural-ODE and unstructured baselines.

Supporting methods: PHAST port-Hamiltonian forecasting; perennial learning.

Inputs

Four method thrusts turn raw evidence into features the twin can use

Multimodal evidence has to become analysis-ready features before the twin can reason over it. Four method thrusts do this. Each is a standalone brief with full methods, figures, and references.

• Imaging. DaT-SPECT and T1 MRI run through automated atlas-constrained segmentation and MNI coregistration; per-ROI features fuse with clinical and molecular measures. Brief →
• Diffusion and clinical co-clustering. Free-water-corrected DTI is fused with UPDRS and MoCA scores in a robust variational co-clustering framework that aligns microstructural and clinical manifolds into observational subtypes. Brief →
• Wearable and cognitive phenotyping. APDM Opal IMUs over sway, TUG, and dual-task walk yield 50+ gait and arm-swing features, combined with cognitive batteries. These help mainly when fused: the gait IMU substudy is roughly 93–100 patients and the integrated risk model is internally validated on n=204 with no external cohort, so treat it as hypothesis-generating. Brief →
• Mechanism inference. Statistical testing across harmonized PPMI and LRRK2-consortium cohorts associates biomarkers with candidate pathways. These are FDR-controlled, cross-sectional associations and hypotheses for follow-up, not causal claims, and cluster labels are not applied to individual patients. Brief →

See all four thrust briefs →

Reasoning

Modality-specialized agents with evidence arbitration

Collaborative-filtering signals — what the model has learned across many patients — sharpen the plausible state space for one new patient, while every candidate stays grounded in that patient's own data, mechanism, literature, and clinical context.

• Diagnostic agents specialize by modality: they interpret EHR, imaging, biomarkers, wearables, clinical history, medication context, and DBS/LFP signals where available. Each is intended to return a state estimate, a differential where appropriate, and explicit uncertainty and evidence gaps; single-modality differentials are contributory, not standalone.
• Therapy-planning agents reason over medication, neuromodulation, rehabilitation, monitoring, and safety.
• A dynamic AI-knowledge network — literature, biomedical mechanisms, population histories, and clinical EHR knowledge — continually informs the agents.
• Evidence arbitration compares and reconciles the agents' claims, resolves conflicts, filters unsafe or inconsistent options, ranks what remains by supporting evidence, and reports uncertainty and remaining evidence gaps.

A research direction

Planning DBS changes in the twin (not yet validated)

At the circuit level, Parkinson's motor signs are associated with pathological beta-band hypersynchrony in the subthalamic–cortical loop, and desynchronization is the therapeutic principle behind deep brain stimulation. Reactive and adaptive DBS have already reached the clinic. This is background motivation, not an AI4PD result.

AI4PD's intended contribution is a planning step — simulate candidate stimulation changes inside the patient-specific twin before any patient setting is changed — a prospective capability still subject to clinical validation. Whether a patient-specific twin can faithfully predict an individual's response to a given stimulation change is itself unproven; any in-twin simulation is a hypothesis to be tested against observed clinical response, never a basis for changing settings on its own. The twin does not set DBS candidacy or stimulation targets and does not prescribe. Framing intervention as controlled movement through structured patient state is a research direction (Hamiltonian control), not a present clinical feature.

What the clinician sees

Decision support, returned to the clinician

Using the same population map from the agents above, a new patient can be placed within the population and matched to statistically similar prior cases for context. Latent similarity is not clinical equivalence, and a neighbor's trajectory is not a prognosis for the index patient; retrieved cases are illustrative context, with representativeness and uncertainty reported. Per-patient subgroup assignment is a validation target, not a delivered output. The intended clinician-facing outputs below are validation targets for a system under development, not delivered guarantees:

Current patient state

Motor, cognitive, and functional estimates from EHR, imaging, biomarkers, wearables, medication history, and DBS/LFP signals when available; autonomic estimates are an aspirational layer not yet backed by a dedicated validated result.

Differential-diagnosis support (planned, not validated)

Intended ranked support for Parkinson's versus look-alikes (MSA, PSP, DLB) with uncertainty and missing-evidence flags. This is contributory to, not a replacement for, expert assessment; discriminating PD from these look-alikes is difficult even for specialists, and no cited result yet evaluates this differential.

Progression forecast (research aim, not validated)

The twin is being designed to estimate longitudinal trajectories. Specific endpoints such as 12-month MDS-UPDRS-III change and time-to-motor-fluctuation are pre-registered validation targets, not demonstrated predictions; the current evidence base is cross-sectional, individual-level prediction has not been established, and any forecast must report ON/OFF medication state and is not a substitute for clinical follow-up.

Medication and therapy planning

Levodopa-response support, rehabilitation planning, and ranked options filtered through safety agents.

DBS planning support (planned, see above)

Twin-side simulation of candidate stimulation changes before any patient parameter is altered.

Follow-up and evidence gaps

Personalized next steps for monitoring, reassessment, and uncertainty reduction.

Interactive view: Parkinson Viz dashboard · full walkthrough on the clinician workflow page.