Ambient Intelligence / Research
The science behind contactless radar monitoring, mapped to what we build and what we are entitled to claim.
Capability → Evidence → Honest bound
Height-collapse detection from 3D people-tracking. Contactless, works in any light, sees through fabric and darkness.
Bound A far-side bed-exit fall is occluded from a single room radar (the mattress blocks line-of-sight to the floor). The bed needs its own radar.
Chest-wall micro-motion via range-bin phase. Breathing is the strong, reliable channel; heart rate is the hard one.
Bound Best within ~1 m on a still subject. Limits of agreement reach +/-14 bpm under motion. Breathing-harmonic contamination of the heart band is the #1 limiter.
In-home walking speed, variability, and step length over time, turned into a per-resident baseline that forecasts falls and decline.
Bound Needs weeks of baseline before deviation is meaningful. Single-viewpoint radar has an aspect-angle bias; trend detection tolerates it, absolute clinical-grade numbers want a second orthogonal view.
Micro-Doppler gait-type and abnormality, freezing-of-gait, Parkinsonian signs, sit-to-stand kinematics.
Bound Research-stage ML. Radar gives limb-motion signatures, not camera-grade joint positions. Clinical claims need validation (SaMD scope).
Coarse mobility ADLs (sit / stand / walk / transfer) directly; finer ADLs inferred from location and duration.
Bound Fine ADLs (bathing, dressing, continence) are beyond radar's spatial resolution. The reachable picture is functional mobility plus space-use.
Respiratory pattern (Cheyne-Stokes, apnea, agonal change), restlessness, and HR trend for bed-bound, still patients.
Bound The best case for radar vitals: a still patient minimizes the motion artifact. Contactless preserves dignity. Clinical claims framed via SaMD.
On-radar TinyML (agitation, cough, seizure) trained on the deployed fleet's own data.
Bound Gated on the passive-platform data flywheel: you cannot train these without a deployed fleet generating labeled data first.
Study registry
12 rooms, multiple zones per room. Falls, vitals, and a gait baseline on a single device design.
ODAT-grant funded. Evidence and protocols feed the SaMD regulatory file.
Radar HR/BR vs reference (pulse-ox / chest strap), Bland-Altman, on still / hospice-like subjects.
Establishes the accuracy claim where radar vitals are strongest.
Longitudinal in-home gait speed and variability vs incident falls and functional change.
Tests the reactive-to-preventive thesis on our own deployment.
Method principles
The target population will not wear a pendant, press a button, or gesture. Every capability must work passively, with a person who cannot or will not act.
Absolute gait and vitals vary enormously person to person. The signal is deviation from an individual's rolling baseline, not a population threshold.
The dominant real-world error is motion. Knowing when a measurement is trustworthy matters more than squeezing the last fraction of accuracy.
Accuracy figures are paper-reported, often under controlled or stationary conditions. Treat them as directional, not as our validated performance.