What we are working on
Active scientific work, lines that have already produced shipped tools (Grovalin, Volitiq), and proposed projects (WAVE, PULSE under the HEARTH program). Plus the statistical foundation shared across all three programs.
Research lines
Six lines of work. Some have already produced shipped tools; some sit under the HEARTH program for in-home health sensing; some are still in the question-asking phase.
A theory-driven pipeline extending Mosconi et al. (2015), testing whether feedforward (anterior cerebellum) and feedback (posterior cerebellum) motor deficits can be dissociated from standard video, not just specialised force transducers. Built on the KCART-praxis dataset (118 participants, 582 tool-use videos).
Grounded in VR cognitive-rehabilitation research showing significant improvements in contextual processing and cognitive flexibility in children with ASD (PMC3845243). Translates clinical findings into an AI-driven learning system that personalises to objects in a child’s own home.
Motor differences are documented in 34–80% of autistic individuals. This line of work operationalises digital phenotyping signals such as postural sway, atypical head movement, and movement variability (PMC6242931) as quantitative measurements from a browser webcam, without specialist hardware.
Recovering human motion, gait, falls, and respiratory rate from the perturbations a body imposes on the multipath structure of ordinary WiFi signals. Cross-modal supervision against vision pose estimators; forward-compatible with IEEE 802.11bf (Oct 2024).
Heart rate, HRV, and respiratory rate from the chromatic modulation each heartbeat imposes on skin colour, observed with an ordinary RGB camera. Adapting recent rPPG methods (PhysFormer, CodePhys, Spiking-PhysFormer) to longitudinal in-home use with explicit older-adult and disability cohorts.
Joint WiFi-CSI and rPPG sensing for continuous caregiver-oriented monitoring. WAVE penetrates walls but cannot resolve a heartbeat; PULSE resolves the heartbeat but needs the person in frame. The fusion is structurally novel for in-home eldercare and is the unique angle HEARTH can claim.
A foundation, not a feature
The product lines of AutismAILab look superficially different. Volitiq watches motor performance in the clinic. HEARTH listens to motion and pulse at home. Grovalin supports learning in context. What unifies them is not the sensor and not the population. It is the statistical layer underneath: a way of looking at a person's biomarker trajectory across time, comparing it to a population, and quantifying when something has changed.
That layer rests on three established methods, used in combination.
Trajectory analysis
Human biomarkers are not linear functions of age. Motor skill grows in childhood, plateaus across young adulthood, and declines unevenly in later life. GAMMs are the standard tool for handling this kind of curve: smooth functions of age fitted by penalized regression splines, mixed-effects to handle repeated measurements on the same subject, generalized link functions for non-Gaussian outcomes. Fitted in mgcv (R) or pyGAM (Python).
Why this matters here
The same GAMM that fits motor performance in our autism cohort (ages 7–34) for the developmental trajectory chapter of the dissertation, run on gait speed extracted by WAVE in older adults (65+), gives the aging trajectory for the HEARTH program. One method, multiple populations.
Normative modeling
Rather than ask "is the mean different between groups," normative modeling asks "where does this individual fall in the full distribution of the population at their age?" Gaussian processes fit the conditional distribution of a biomarker as a function of covariates, producing not just a fitted curve but a credible interval at every point. The modern alternative to case-control statistics, drawn from Marquand et al.
Why this matters here
Builds a population-conditioned reference for every biomarker the lab measures, from Volitiq motor scores to HEARTH gait speed and HRV. Each new measurement is positioned against the conditional distribution, not the marginal mean.
Individual deviation
From the normative model, each person carries a continuous z-score against the population norm at their age. Longitudinal change in that score (not the raw biomarker) becomes the clinical signal. This converts noisy biomarker measurements into a stable, age-controlled trajectory that can be correlated with clinical assessments (TUG, gait speed, MoCA, ADOS-2 for the autism arm).
Why this matters here
A single statistical layer that operates identically whether the upstream sensor is a clinical camera (Volitiq), a learned WiFi pose network (WAVE), or an rPPG-derived heart-rate stream (PULSE). This is the structural reason for keeping HEARTH inside the same portfolio as Volitiq rather than spinning it out.
The marginal cost of adding a new modality, conditional on the statistical layer existing, is dominated by sensing hardware and the modality-specific front-end ML pipeline, not by the analysis layer. That is the structural reason WAVE and PULSE belong inside AutismAILab rather than as spinouts.
Researchers, clinicians, and self-advocates with relevant cohorts, datasets, or methodological expertise are invited to make contact.