RESEARCH OVERVIEW Homepage

Integrated Sensing and Physical AI for Health Digital Twins

Understanding how our health changes in everyday life—and what to do next.

CORE SCIENTIFIC QUESTION

Your smartwatch can tell you when your heart rate rises—but can it tell you why, whether it matters, and what to do next?

How can we reliably infer changing human physiological states from heterogeneous, imperfect observations—and determine when those changes matter and what evidence to acquire next?

y(t) · HETEROGENEOUS, IMPERFECT OBSERVATIONS — observed ┄ missing ○ probe NEXT EVIDENCE BIOELECTRICAL BIOMECHANICAL / AMBIENT BIOCHEMICAL EMG · throat ECG · ear ACC · neck RF · WiFi SWEAT · lipid PERSONALIZED PHYSIOLOGICAL STATE x̂i(t) individual i missing missing missing missing artifact artifact now enable a channel request a reference model → sensor

Swipe left or right to explore all signals and the next evidence.

Schematic. Bioelectrical, biomechanical and ambient, and biochemical channels drop out or carry artifacts. The personalized state x̂i(t) carries uncertainty that widens where evidence is missing, and the model feeds back to request the next observation or a reference measurement.

THE MODEL

From longitudinal observations to personalized models

Every measurement reflects two processes that a personalized model must tell apart: how each person’s physiological state evolves under dynamics of their own, and how sensors turn that state into data.

STATE-SPACE VIEW · INDIVIDUAL i OVER TIME
… … ci,t ui,t fθi + ηi,t di,t di,t+1 h h xi,t xi,t+1 yi,t yi,t+1 latent state measurement

Hidden states evolve under each person’s own dynamics; every sensor reading is a noisy, condition-dependent view of the current state.

STATE DYNAMICShow each person changes
xi,t+1 = fθi(xi,t, ci,t, ui,t) + ηi,t
θi = θ + δipopulation structure + personal deviation

xi,t latent physiological state of individual i at time t · θi that individual’s own dynamics parameters · c context · u observable task or external input · η unmodeled process variation

OBSERVATION PROCESShow sensors turn that state into measurements
yi,t = h(xi,t, ci,t, di,t) + εi,t

d measurement conditions such as placement, contact, and device · ε measurement noise

WHAT EACH THRUST ASKS OF THIS MODEL
IWhich states can be recovered through h?
IIDid x change through this person’s dynamics fθi, or did c or d change the reading?
IIIWhich observation would most reduce uncertainty in x next?
THREE THRUSTS

Three questions, one loop

Recover what is hidden. Tell what a change means. Decide what to measure next.

THRUST I

Observability

What can we reliably recover?

Establish which physiological states are identifiable from integrated electrical, mechanical, biochemical, and ambient measurements.

electrical mechanical biochemical ambient x̂i Multimodal → state
x1 ≠ x2 ⇒ p(y | x1) ≠ p(y | x2)
identifiable: different states leave different evidence
THRUST II

Interpretability

When does a change actually matter?

Model individual physiological dynamics, contextual influences, measurement processes, and uncertainty to distinguish genuine state changes from observational artifacts.

observed y state x̂i context c sensor d context artifact genuine change State + context + time
p(Δxi,t ≠ 0 | yi,1:t, c, d)
is the change genuine, given context c and measurement process d?
THRUST III

Actionability

What evidence should we acquire next?

Develop uncertainty-aware AI that selects informative, low-burden observations and updates personalized physiological models.

infer probe update Infer → probe → update
a* = argmaxa I(xi,t; ya) − λ·cost(a)
trade the information gained against measurement burden
LONG-TERM VISION

Personalized Health Digital Twins

Personalized, adaptive physiological models that support more reliable health decisions.

MEASURE UPDATE PREDICT ACQUIRE EVIDENCE REFINE
MEASURE · New observations of individual i arrive.
UPDATE · The personal model departs from the population.
PREDICT · The model forecasts ahead, with uncertainty.
ACQUIRE EVIDENCE · The most informative next measurement is acquired.
REFINE · The forecast tightens.
measured now forecast individual i population model SCHEMATIC