---
title: 'mCardiacDx: Radar-Based Arrhythmia Diagnosis'
url: https://www.emergentmind.com/topics/mcardiacdx
type: topic
---

# mCardiacDx: Radar-Based Arrhythmia Diagnosis

Searching arXiv for the cited mCardiacDx paper and closely related cardiac-AI context papers.
mCardiacDx is a millimeter-wave radar-based, contactless system for arrhythmia monitoring and diagnosis that is designed to recover heartbeat-related waveforms from reflected chest signals and use them for heart-rate estimation, RR-interval analysis, heart-rate-variability computation, and binary healthy-versus-arrhythmia classification [2508.02274]. Its defining premise is that prior contactless cardiac sensing methods were developed mainly for healthy subjects and degrade in arrhythmia because reflected radar signals exhibit both spatial instability across range bins and temporal inconsistency relative to ECG-defined heartbeat timing. The system addresses these two failure modes through a signal-localization module called Precise Target Localization (PTL) and a reconstruction network called HPR-Net, then feeds reconstructed heart pulse waveforms into downstream monitoring and diagnostic analysis [2508.02274].

## 1. System definition and diagnostic scope

mCardiacDx is framed as a contactless alternative to conventional electrocardiogram and wearable monitoring for arrhythmia surveillance, particularly in settings where continuous observation is desirable but contact sensors are inconvenient or uncomfortable [2508.02274]. The system does not attempt to replace ECG as a gold standard; rather, it uses simultaneously recorded ECG as supervision and reference while seeking to infer clinically useful cardiac timing information from radar reflections alone [2508.02274].

The diagnostic scope reported for mCardiacDx is deliberately narrow and well specified. It supports four outputs: heart rate, RR intervals, six time-domain heart-rate-variability metrics, and binary arrhythmia diagnosis, where the two classes are healthy and arrhythmia [2508.02274]. This makes it a monitoring-and-triage framework rather than a full electrophysiologic interpretation engine. A plausible implication is that its current design is best suited to identifying abnormal rhythm burden rather than subtype-level rhythm taxonomy.

The paper’s problem formulation is clinically motivated by the fact that arrhythmia can precipitate blood clots, stroke, and sudden cardiac death, and by the observation that many existing radar cardiac systems assume a stable reflection pattern that does not hold in arrhythmic patients [2508.02274]. In this respect, mCardiacDx differs from conventional ECG-based mobile diagnosis systems such as the wavelet-plus-Bayesian architecture for Normal, PVC, PAC, and Myocardial Infarction classification, which remain contact-based and operate directly on electrical recordings [1904.07976]. It also differs from cine-MRI segmentation-and-diagnosis pipelines, where the signal source is imaging-derived anatomy and function rather than chest-wall micro-motion [1708.01141].

## 2. Radar sensing setup and signal representation

The hardware implementation uses a Texas Instruments AWR1642BOOST mmWave radar, a DCA1000 real-time capture adapter, and a Shimmer3 ECG kit for ground truth [2508.02274]. The radar operates in the 77–81 GHz band with 1 transmitter and 4 receivers, chirp period 50 µs, idle time 150 µs, ADC sampling rate 6000 ksps, 256 samples per chirp, and receive gain 30 dB [2508.02274]. ECG is recorded simultaneously at 500 Hz using left arm, right arm, left leg, and right leg electrodes [2508.02274].

Participants are seated 25 to 55 cm from the radar in a typical office environment containing ambient WiFi, LTE, and Bluetooth activity, and are instructed to breathe normally while minimizing voluntary movement [2508.02274]. The system therefore targets contactless observation under relatively controlled but not shielded conditions. The paper explicitly limits its evaluation to the idle or seated state, which indicates that large body motion remains outside the validated operating regime [2508.02274].

The radar return is represented as a two-dimensional channel impulse response matrix, with fast time corresponding to range and slow time corresponding to temporal evolution across chirps [2508.02274]. Each fast-time sample is a range bin associated with a spatial chest-wall region. Chest-wall vibration is extracted from radar phase according to

$$
d(t) \propto \angle \mathrm{CIR}(t) = \frac{\lambda}{4\pi}\phi(t)
$$

where $\lambda$ is wavelength and $\phi(t)$ is the measured phase [2508.02274]. This defines the core signal model of mCardiacDx: heart-related chest micro-motion is encoded in reflected-signal phase.

A useful contrast with other noninvasive cardiac platforms is that ECG systems such as the Bayesian network classifier pipeline begin from electrophysiologic waves and explicitly measure $P$, $QRS$, $T$, $PR$, $ST$, and $QT$ features [1904.07976], whereas mCardiacDx infers beat timing only indirectly through mechanical motion. That distinction is central to both its promise and its limitations.

## 3. Spatial disruption and Precise Target Localization

The paper’s first methodological contribution is the claim that arrhythmia causes “spatial disruption,” meaning informative cardiac reflections are no longer confined to one stable range bin but can disperse across multiple unstable neighboring bins [2508.02274]. Healthy-subject methods that track a single dominant bin are therefore mismatched to arrhythmic reflection structure [2508.02274].

PTL is designed as a dynamic bin-selection strategy rather than a static localization rule. Let

$$
R \in \mathbb{C}^{\text{num\_samples} \times \text{num\_chirps}}
$$

denote the complex channel impulse response matrix, and let

$$
M = |R|
$$

be its magnitude [2508.02274]. PTL first identifies an initial target bin using a Most Common Bin procedure,

$$
t \gets \mathrm{MostCommonBin}(M)
$$

then expands to a neighboring search window of breadth $w_b$,

$$
\text{first\_bin} = t - \left\lfloor \frac{w_b}{2} \right\rfloor
$$

$$
\text{last\_bin} = t + \left\lceil \frac{w_b}{2} \right\rceil - 1
$$

and extracts the corresponding magnitude submatrix [2508.02274]. Slow time is then processed in windows of $w_t$ chirps; within each window,

$$
M' = M[:, i:i+w_t-1]
$$

and the selected range bins are updated by

$$
s_{i:i+w_t-1} \gets \text{first\_bin} + \mathrm{MostCommonBin}(M')
$$

[2508.02274].

This mechanism allows mCardiacDx to re-estimate the dominant reflection locus over time instead of assuming fixed spatial stability. The independent ablation Baseline versus Baseline+PTL demonstrates that PTL alone improves waveform reconstruction, heart-rate estimation, RR-interval estimation, and diagnosis, especially for arrhythmia patients [2508.02274]. For arrhythmia, DTW improves from 5.92 to 3.78, HR MedAPE from 9.10% to 4.95%, RR MedAPE from 8.42% to 5.21%, recall from 0.75 to 0.83, and accuracy from 0.85 to 0.89 [2508.02274]. This indicates that spatial localization is not merely a preprocessing convenience but a central algorithmic component.

## 4. Temporal disruption and HPR-Net

Even after PTL, the paper argues that arrhythmia creates “temporal disruption”: reflected signals become less consistently aligned with ECG-defined beat timing [2508.02274]. The authors quantify this using cross-correlation, reporting normalized ZNCC of 1.0 for healthy subjects and 0.54 for arrhythmia patients [2508.02274]. This motivates HPR-Net, whose purpose is to reconstruct heart pulse waveforms from temporally inconsistent multi-bin radar signals.

Before HPR-Net, mCardiacDx extracts phase and magnitude from the PTL-selected bins and applies a 0.2–50 Hz bandpass filter to remove low-frequency drift and high-frequency noise while preserving respiration and cardiac motion [2508.02274]. It then computes a second-order derivative of phase,

$$
\phi''_t = \frac{(\phi_{t-3} + \phi_{t+3}) + 2(\phi_{t-2} + \phi_{t+2}) - (\phi_{t-1} + \phi_{t+1}) - 4\phi_t}{16h^2}
$$

to emphasize higher-acceleration cardiac motion over slower respiration [2508.02274]. The three signals ultimately fed to HPR-Net are bandpass-filtered phase, bandpass-filtered magnitude, and second-derivative-filtered phase [2508.02274].

HPR-Net has three modules: Heart Signal Extractor, Encoder-Decoder, and Reconstructor [2508.02274]. The Heart Signal Extractor uses 1-D convolutions and residual blocks for temporal feature extraction and a Graph Attention Network-style mechanism for inter-bin feature integration. For range-bin features $\vec{h_i} \in \mathbb{R}^d$ at one time step, the updated representation is

$$
\vec{h_i'}=\sigma\left(\sum_{j=1}^{m}{\alpha_{ij}\mathbf{W}\vec{h_j}}\right)
$$

with attention coefficients

$$
\alpha_{ij} = \frac{ \exp\left( \mathrm{LeakyReLU}\left( \vec{a}^{T}[\mathbf{W}\vec{h_i}\,||\,\mathbf{W}\vec{h_j}] \right) \right) }{ \sum_{k=1}^{m} \exp\left( \mathrm{LeakyReLU}\left( \vec{a}^{T}[\mathbf{W}\vec{h_i}\,||\,\mathbf{W}\vec{h_k}] \right) \right) } 
$$

where $\sigma$ is LeakyReLU with slope 0.2 [2508.02274]. This lets the model learn which unstable range bins should influence each other most strongly.

The Encoder-Decoder is a convolutional encoder-decoder with residual blocks, transposed residual blocks, and skip connections [2508.02274]. The Reconstructor is a BiLSTM followed by dense layers and sigmoid activation, yielding the final heart pulse waveform [2508.02274]. The use of BiLSTM is specifically intended to compensate for temporal inconsistency by using both past and future context [2508.02274].

The training target is an ECG-derived heart pulse waveform, but the paper does not provide an explicit neural-network loss equation [2508.02274]. That omission is methodologically important because it limits exact reproducibility. Still, the architectural separation between spatial stabilization and temporal reconstruction is clear and is the paper’s main technical contribution.

## 5. Monitoring outputs and diagnostic performance

From reconstructed heart pulse waveforms, mCardiacDx applies peak detection to estimate heart rate and RR intervals, then computes six HRV metrics: MeanNN, MedianNN, SDNN, IQRNN, MadNN, and MadNN / MedianNN [2508.02274]. Binary diagnosis is performed by a Random Forest classifier operating on such HRV features [2508.02274]. This means that HPR-Net is not the diagnostic classifier itself; it is an upstream physiological signal reconstruction module.

The evaluation uses three tasks: waveform reconstruction, monitoring, and diagnosis [2508.02274]. Reconstruction quality is measured by DTW, monitoring accuracy by MedAPE, and diagnosis by accuracy, precision, recall, F1, and ROC/AUC [2508.02274]. The strongest reported improvements occur in arrhythmia patients rather than healthy subjects, which supports the paper’s central claim that arrhythmia-specific failure modes must be modeled explicitly [2508.02274].

The main quantitative results are summarized below.

| Task | Baseline | Baseline+PTL | mCardiacDx |
|---|---:|---:|---:|
| Healthy DTW | 4.02 | 3.90 | 2.82 |
| Arrhythmia DTW | 5.92 | 3.78 | 2.92 |
| Healthy HR MedAPE | 2.66% | 2.63% | 1.68% |
| Arrhythmia HR MedAPE | 9.10% | 4.95% | 2.94% |
| Healthy RR MedAPE | 2.73% | 2.70% | 1.71% |
| Arrhythmia RR MedAPE | 8.42% | 5.21% | 2.95% |

For diagnosis, Baseline achieves precision 0.94, recall 0.75, F1 0.83, accuracy 0.85, and ROC 0.97; Baseline+PTL improves to precision 0.95, recall 0.83, F1 0.88, accuracy 0.89, and ROC 0.97; mCardiacDx reaches precision 0.95, recall 0.91, F1 0.93, accuracy 0.93, and ROC 0.98 [2508.02274]. The confusion details reported for mCardiacDx are 1 false positive, 2 false negatives, 22 true positives, and 23 true negatives [2508.02274]. Relative to Baseline, PTL gives +11% recall, +6% F1, and +5% accuracy, while the full system gives +21% recall, +12% F1, and +9% accuracy [2508.02274].

These results position mCardiacDx as a specialized contactless arrhythmia monitoring system rather than a broad cardiac diagnosis platform. Compared with ECG-based anomaly classifiers that directly categorize Normal, PVC, PAC, and MI from waveform features [1904.07976], mCardiacDx currently performs a simpler healthy-versus-arrhythmia decision but does so without body contact.

## 6. Dataset, limitations, and relation to broader cardiac AI

The dataset description in the paper is internally inconsistent. The Introduction mentions evaluation on 48 subjects, 24 healthy and 24 arrhythmia, while the Implementation and Experimental Details section reports 210 participants: 108 healthy and 102 arrhythmia patients, ages 32–68, 60-second trials, about 20,000 heartbeats, healthy HR range 67–97 bpm, and arrhythmia HR range 55–115 bpm [2508.02274]. The later section appears to be the fuller description, but the inconsistency remains a documented limitation.

The paper also reports train, validation, and test splits separately for healthy and arrhythmia cohorts, but the listed healthy 60 / 24 / 24 and arrhythmia 56 / 22 / 24 counts sum to more than the total cohort sizes, indicating a typographical error in the split specification [2508.02274]. This affects confidence in exact evaluation protocol reconstruction.

Several additional limitations are explicit. The experiments are limited to idle or seated subjects; the system reconstructs heart pulse waveforms rather than full ECG-equivalent morphology; and the relationship between temporal misalignment and arrhythmia severity is left for future study [2508.02274]. The paper also provides no subtype-level arrhythmia breakdown, no motion-robustness evaluation, no cross-site generalization analysis, and no detailed calibration study [2508.02274]. These omissions matter because contactless sensing systems are especially sensitive to body motion, posture changes, and scene complexity.

In a broader research landscape, mCardiacDx belongs to a wider family of noninvasive cardiac AI systems that differ primarily in sensing modality and diagnostic target. Cine-MRI pipelines use segmentation-derived biomarkers for structural diagnosis [1708.01141], ECG systems use wavelet features or deep learning for rhythm or ischemia detection [1904.07976; 2009.13232], and multimodal opportunistic screening systems combine ECG with chest X-ray for cardiovascular event risk prediction [2506.19174]. mCardiacDx occupies a distinct position within this landscape: it uses mmWave radar as a contactless surrogate for heartbeat-related mechanical activity and restricts its downstream inference to monitoring and binary arrhythmia diagnosis [2508.02274].

A plausible implication is that mCardiacDx is best viewed as a modality-specific building block. Its PTL and HPR-Net design patterns—dynamic localization under spatial instability and bidirectional temporal reconstruction under misalignment—could inform other contactless physiological sensing systems, but the present system does not yet provide the disease breadth of ECG- or MRI-based cardiac AI frameworks. Its strongest evidence is the large performance gain on arrhythmia patients relative to healthy baselines, which supports the claim that contactless cardiac sensing requires arrhythmia-specific modeling rather than direct transfer from healthy-subject methods [2508.02274].

Source: https://www.emergentmind.com/topics/mcardiacdx