Radar can now reconstruct a 12-lead electrocardiogram (ECG, the heart's electrical trace) from millimetre-scale chest vibrations, with no electrodes on the skin. That matters for newborns, elderly patients and overnight monitoring, where adhesive sensors cause harm or go unworn. These systems work by translating radar into the "vocabulary" of a model pretrained on hundreds of thousands of ordinary ECGs. The same design creates a hazard. When the radar carries little information about some part of the waveform, the pretrained ECG prior fills it in, and the result looks like a clean, plausible trace. A clinician who sees ST-segment depression in a radar-derived Lead III cannot tell whether the chest motion actually showed it or the decoder invented it.
We propose a per-finding evidence test that answers exactly that question. For each clinical finding in a reconstructed ECG, we build a counterfactual ECG with the finding removed. Both versions pass through a physically structured forward model that predicts what the radar should have recorded. We then ask whether the observed radar favours one version over the other. If the radar cannot tell them apart, the finding is labelled "not observable from this sensor," however confident the reconstruction looks. The explanation is not a heatmap over model internals. It is a check against the physics of what the sensor can see.
Core Claim
Current contactless cardiac systems report signal fidelity and classification accuracy but never say which parts of a given reconstruction are supported by the measurement. Three examples show the pattern. Contactless 12-lead electrocardiogram via deep computational radar (CardioRadar) maps beamformed radar into a pretrained codebook, a learned dictionary of ECG shapes. It reports a mean correlation of 0.75 and an RMSE of 0.09 mV, with markedly worse Lead III than Lead II. Discovering Interpretable Semantics from Radio Signals for Contactless Cardiac Monitoring aligns radar latents to an ECG latent space and visualises those factors. Atrial fibrillation detection via contactless radio monitoring and knowledge transfer outputs only an AF probability.
All three motivate the proposal, because a strong ECG prior makes the output look trustworthy exactly where the sensor is weakest. Our claim is that a log-likelihood ratio under a forward model can separate supported findings from hallucinated ones, patient by patient and finding by finding. It does so better than uncertainty or saliency computed inside the reconstruction network, because those methods share the prior that produced the hallucination.
Concretely, let $y$ be the observed radar segment, $\hat{x}$ the reconstructed ECG, and $\hat{x}_{\neg f}$ a counterfactual ECG identical except that finding $f$ is removed. Let $p_\phi(y \mid x)$ be the forward model's likelihood of radar given an ECG. The evidence score is
$$S(f) = \log p_\phi(y \mid \hat{x}) - \log p_\phi(y \mid \hat{x}_{\neg f}).$$
Averaging $S(f)$ over radar simulated from the forward model gives an observability map. That map states, before any patient is scanned, which findings in which leads the sensor can distinguish at all. Lead III ST changes are a natural candidate for low observability.
Research Program
The work has three components.
The first is the forward model, from ECG to radar. We reuse the structure of Cardio-mmFlow, which synthesises radar from ECG. It uses a mass-spring-damper model of the chest wall (the chest treated as a damped spring) whose parameters are set from age, height, weight and BMI. We keep that physical core and add a neural residual for effects the spring model misses.
This residual is the weak point. If the neural part can model radar without depending on the ECG, then $S(f)$ collapses toward zero for every finding. If it absorbs the physical dynamics, the damping and stiffness estimates lose meaning. To prevent this we adapt the penalty of Orthogonal projection-based regularization for efficient model augmentation. That method penalises the part of the neural output lying in the subspace spanned by the physical model's Jacobian, so the network learns only true residuals. It was shown only on vehicle simulation, so its transfer to cardiac mechanics is a hypothesis we test, not assume. The method constrains the design too, because its weight $\beta$ trades orthogonality against accuracy, and we must report sensitivity to it.
The second component is counterfactual editing. The candidate findings are ST depression, T-wave inversion, absent P-waves, premature beats and QRS widening. We implement removal in two ways and report both: parametric template edits to the affected beats, and substitution of nearest "normal" codes in a VQ-VAE codebook of the kind CardioRadar uses. The fixed-healthy-beat ODE of radarODE is a useful negative lesson here. Its ODE enforces a normal PQRST shape and cannot represent arrhythmias. For that reason we never use a normal-beat prior as the forward model. Physiological constraints appear only in the edit operator, where producing a normal beat is the point.
The third component is the clinician-facing output. For each finding the system gives a short structured statement: the finding, the lead, and its evidence score relative to an observability-calibrated threshold. Structuring these statements follows the inspectable reasoning steps of Visual and Domain Knowledge for Professional-level Graph-of-Thought Medical Reasoning. The key difference is that each step here is grounded in a measurement likelihood rather than generated text. That avoids the problem that free-text rationales can be plausible and still physiologically wrong.
Baselines
The naive reference point is to accept every finding in the reconstruction. Its finding-level precision equals the reconstruction's raw diagnostic precision.
The second reference point is a population lookup table: per-lead, per-finding reliability estimated on a validation set, as implied by CardioRadar's Lead II versus Lead III gap. It is cheap and captures sensor geometry. It cannot vary with the individual patient or recording.
The strongest competitor is a deep ensemble of reconstruction networks. We train five members with the CardioRadar-style architecture on the same paired data and compute each finding's agreement rate or its amplitude variance. This is the standard way to flag unreliable outputs. Every method gets the same training split, the same GPU budget for the reconstruction backbone, and the same hyperparameter search budget for its flagging threshold. We also include layer-wise relevance propagation on the reconstruction network, applied as Deep Learning for ECG Analysis: Benchmarks and Insights from PTB-XL did for contact ECG classifiers, scoring how much relevance for the finding falls on high-SNR radar regions.
For the flagging task itself, we found no reported result on Lacuna for detecting hallucinated findings in radar-reconstructed ECG. The closest reported numbers concern reconstruction and diagnosis rather than finding-level trust. CardioRadar reports mean correlation 0.75 and RMSE 0.09 mV. Chen et al. report AF F1 0.929 and premature-beat F1 0.867. Yuan et al. report AF sensitivity 0.844 and specificity 0.995.
Beating the ensemble means a higher AUROC for separating true from false reconstructed findings, with the gain concentrated in low-observability leads. The pre-registered margin is at least 0.05 AUROC on pooled findings.
Two ablations isolate the key ingredients. The first removes the orthogonal penalty ($\beta = 0$), which tests whether the regulariser is what keeps $S(f)$ informative. The second replaces the physical forward model with a fully neural ECG-to-radar network of matched size, which tests whether the physics matters or any forward model would do.
Why not just…?
Why not just use ensemble or dropout uncertainty on the reconstruction? This is the real threat. Ensembles trained on the same pretrained codebook tend to agree when they fall back on the prior, which is exactly when they hallucinate a typical-looking segment. The forward model instead asks whether the measurement discriminates, which is a different quantity from model disagreement. The planted-feature experiment below settles the question directly. If ensemble disagreement detects radar-unsupported findings as well as $S(f)$ does, the physics adds nothing and we will say so.
Why not just run saliency or LRP on the reconstruction network? LRP shows where the network looked, as in the PTB-XL analysis. It cannot show whether what the network saw actually distinguishes the finding from its absence. A network can put relevance on a clean radar window that carries no ST information at all. The comparison is part of the baseline suite.
Why not skip reconstruction and classify directly from radar, as the AF systems do? Direct classifiers such as Yuan et al.'s give strong AF numbers but no morphology. They also move the trust problem into the classifier without solving it. Our score applies to any finding a classifier would report, and we additionally run it on AF decisions to see whether it identifies false positives among the classifier's positives.
Why not just use the population lookup table? If per-recording variation in signal quality, posture and body habitus is small, the lookup table will match us. We test for this by stratifying results by BMI and by recording SNR. If our score is to justify its cost, its advantage should grow within those strata.
Proposed Experiments
The decisive experiment is a planted-feature test with known ground truth. We take PTB-XL ECGs (21,837 twelve-lead records with 71 statement labels) and synthesise radar with a separate Cardio-mmFlow-style generator. This generator is deliberately different from our forward model, so we are not scoring on our own simulator.
We plant or remove findings at controlled amplitudes and sweep radar SNR. We then reconstruct with a CardioRadar-style decoder and ask each method whether each reconstructed finding was actually present in the source ECG. The planted features are ST depression, T-wave inversion and QRS widening, which are the findings that cannot be read from inter-beat interval timing alone. Metrics are finding-level AUROC and precision at fixed recall, reported per lead. The prediction is that $S(f)$ beats ensemble disagreement mainly on Lead III and on low-amplitude findings, where the prior dominates.
On real paired data we use synchronized radar-ECG cohorts: the 9,518-outpatient set of Chen et al. or the RECG cohort of Yuan et al. if access is granted, and otherwise a new collection following the radarODE protocol, which covers normal and irregular breathing, sleep and post-exercise recordings. The contact ECG serves as the referee. A reconstructed finding is true if it also appears in the synchronized contact trace.
We check three things on this data. First, whether the observability map predicted from simulation matches the empirical per-lead false-finding rates. Second, whether $S(f)$ stays discriminative under the motion conditions where prior methods degrade. Third, whether the fitted chest damping parameters correlate with BMI once the orthogonal penalty is on, and fail to once it is off. That last result would show that the physical component remains physically meaningful rather than being absorbed by the network.
A small reader study with cardiologists then compares diagnostic accuracy on radar-derived ECGs with and without the evidence statements. It is modelled on the finding of Algorithmic transparency and interpretability measures improve radiologists' performance in BI-RADS 4 classification that explanation aids improved radiologists on hard cases. The study is secondary to the quantitative tests and is run only if they succeed.
References
- Contactless 12-lead electrocardiogram via deep computational radar
- Discovering Interpretable Semantics from Radio Signals for Contactless Cardiac Monitoring
- Atrial fibrillation detection via contactless radio monitoring and knowledge transfer
- Cardio-mmFlow: A Gaussian-Prior-Free Physics-Informed Flow Matching Framework for Electrocardiogram to mmWave Radar Synthesis
- Orthogonal projection-based regularization for efficient model augmentation
- radarODE: An ODE-Embedded Deep Learning Model for Contactless ECG Reconstruction From Millimeter-Wave Radar
- Visual and Domain Knowledge for Professional-level Graph-of-Thought Medical Reasoning
- Deep Learning for ECG Analysis: Benchmarks and Insights from PTB-XL
- Algorithmic transparency and interpretability measures improve radiologists’ performance in BI-RADS 4 classification