Adaptive Radar-Based Physiological Sensing

Research Question

How can a radar-based measurement system decide what to measure next in order to maximise information about an underlying physiological process — respiration, cardiac motion — when each measurement is noisy, the forward model is physically constrained, and real-time limits apply?

Method

This project frames physiological radar sensing as a sequential decision problem under uncertainty. The approach combines three of the four methodological pillars described on the Methods page:

• Uncertainty-aware inference — Bayesian reconstruction of continuous physiological signals (respiration rate, cardiac motion patterns) from sparse, noisy radar measurements, producing calibrated posterior distributions rather than point estimates.
• Adaptive decision-making — selecting radar measurement configurations (waveform, angle, sampling pattern) that maximise expected information gain about the underlying physiological state, framed as a constrained sequential decision problem.
• Mechanistic machine learning — encoding electromagnetic propagation physics and sensor geometry into the inference pipeline as forward-model constraints, rather than treating the radar as a black-box signal source.

Current experimental prototypes

Two controlled toy problems have been implemented to isolate the decision and inference components before scaling to full hardware:

1. 1D radar-like signal reconstruction — a Bayesian linear regression with physics-informed basis functions recovers a synthetic respiratory phase signal from noisy samples. The experiment showed that posterior variance from a global basis is noise-dominated and cannot rank candidate measurement locations by error — a boundary condition for when adaptive acquisition helps and when it does not.

2. 2D point-source tomography — an energy-based model (EBM) maintains a posterior belief over a hidden source location, and an adaptive policy selects the next projection angle to maximise information gain. Greedy policies converge 30–60 % faster early, but random selection catches up as the budget grows — unless the state dimensionality is high relative to the measurement budget, where adaptive sensing wins decisively (31.6 % at D=16, tight budget).

These prototypes are documented in two research-writing posts: Physics-Grounded ML for Adaptive Biomedical Sensing and When Adaptive Sensing Doesn’t Win.

Personal Contribution

Frederik Banis is the lead researcher on this program. He designed the adaptive sensing framework, implemented the experimental prototypes, and formulated the connection to prior energy-systems control work. The research is conducted as part of the guest researcher appointment at Hangzhou Institute of Technology, Xidian University.

Validation and Key Results

• The 1D reconstruction toy confirmed that uncertainty-driven acquisition requires an uncertainty signal that varies meaningfully across candidate measurements — a structural limitation of global bases with noise-dominated posteriors.
• The 2D tomography toy demonstrated that information-gain acquisition outperforms projected-variance acquisition and random baselines when the measurement budget is tight relative to state dimensionality.
• A scaling experiment (D = 2, 4, 8, 16) mapped the boundary conditions where adaptive sensing earns its complexity cost.

All results are from controlled toy problems with synthetic data and known physics. Transfer to real-world radar hardware remains an open question.

Outputs

• Publications: None yet. The program is ongoing; no manuscript has been submitted.
• Software: Experimental prototypes; no released software package at this time.
• Research writing: Two methods-notes posts (linked above) document the prototypes and negative results.

Collaborators

• Hangzhou Institute of Technology, Xidian University — host institution for the guest researcher appointment.

Status and Next Steps

Status: Ongoing. The adaptive sensing framework is under active development. No outcomes are claimed as completed.

Next steps:

• Extend the 1D reconstruction toy to 2D with a realistic radar forward model.
• Runtime-verify the EBM posterior for adaptive measurement selection.
• Integrate physics-informed neural network (PINN) reconstruction with EBM uncertainty estimation in a shared benchmark.
• Scale from toy problems toward hardware-relevant radar configurations.

Connection to prior work: The sequential-decision framework — choose the next measurement to maximise information gain under constraints — is the same framework that drove target-adjusted MPC for microgrid frequency control (published). The shift is from controlling physical energy systems to measuring physiological systems, but the underlying decision-under-uncertainty architecture is shared.

See the Methods page for the methodological pillars, the Research page for the research program overview, and the Applications page for the full application map.