Source-linked AI summary
Clinical radar heart sounds with echocardiographic phenotypes of ventricular dysfunction and valvular disease
Jinye Li, Qingchao Chen
TL;DR
Public radar resources have not paired cardiac recordings with echocardiographic labels of ventricular dysfunction and valve disease. This study presents a paired radar–echocardiography resource and evaluates reduced-LVEF classification, achieving a mean outer-test AUROC of 0.770±0.138 in patient-level five-fold evaluation.
Problem
Public radar collections described here lack echocardiographic labels of ventricular dysfunction or valve disease, leaving radar-recorded cardiac mechanical activity unpaired with those clinical phenotypes.
Source: §1 · p. 2Method
Results
A mean outer-test AUROC of 0.770±0.138 was achieved for reduced LVEF in patient-level five-fold evaluation, while technical validation retained 95.93% of nominal signal duration.
Source: §1 · p. 2Takeaways & Limitations
The paired resource provides a foundation for investigating how cardiac mechanical waveforms relate to ventricular systolic function, regional wall motion and valvular abnormalities.
Source: §1 · p. 2
Abstract
from arXiv · showhide
Radar heart sounds capture cardiac mechanical activity without contact, while echocardiography characterizes ventricular and valvular function. Paired measurements connect these signals to structural heart disease, yet such data remain sparsely represented in public resources. Here we present 537 radar heart-sound recordings paired with one echocardiographic report for each of 162 patients. The cohort covers three overlapping phenotypes: left ventricular ejection fraction (LVEF) below 50%, regional wall-motion abnormality and at least moderate valvular regurgitation, present in 38, 43 and 39 patients, respectively. Original complex signals, processed radar heart sounds and temporal quality segmentation support signal-to-phenotype analysis. Technical validation retains 95.93% of signal duration and confirms reproducible waveform reconstruction. Patient-level five-fold evaluation achieves a mean outer-test AUROC of $0.770\pm0.138$ for reduced LVEF, identifying mechanical information associated with ventricular systolic dysfunction. This resource supports research linking non-contact cardiac mechanical signals to echocardiographic structural and functional phenotypes.
1 Background & Summary
Public radar collections have not paired recordings with echocardiographic labels of ventricular dysfunction or valve disease. This paper presents a clinically paired resource and reports patient-level identification of reduced LVEF from radar heart sounds.
1 Background & Summary
- Echocardiography characterizes cardiac structure and function by quantifying LVEF, identifying regional wall-motion abnormalities and grading valvular regurgitation.The cited passage describes echocardiography as a first-line imaging method and clinical reference standard.Source: §1 · p. 1
- The paired measurements address the missing link between radar-recorded cardiac mechanical activity and clinically characterized structural abnormalities.Existing public radar resources include recordings paired with other physiological references but were acquired from healthy volunteers without echocardiographic labels of ventricular dysfunction or valve disease.Source: §1 · p. 2
- The resource pairs 537 radar heart-sound recordings with one retained echocardiographic report for each of 162 patients.The three nonexclusive phenotypes are LVEF below 50%, regional wall-motion abnormality and at least moderate regurgitation, present in 38, 43 and 39 patients, respectively.Source: §1 · p. 2
- Patient-level five-fold evaluation achieved a mean outer-test AUROC of 0.770±0.138 for reduced LVEF, with mechanical information associated with systolic dysfunction.Technical validation retained 95.93% of nominal signal duration and confirmed reconstruction of processed radar heart sounds from original complex samples.Source: §1 · p. 2
2 Methods
The study recruited patients undergoing clinical cardiac assessment and collected repeated supine radar measurements, pairing each patient’s recordings with a latest echocardiographic assessment. The acquisition system used FMCW radar and a defined signal-processing path to produce cardiac vibration time series.
2.1 Participants and clinical assessment
- The cohort comprised 162 patients with numerical LVEF and binary regional wall-motion and valvular classifications, contributing 537 radar recordings paired with 162 clinical assessments.Patients ranged from 18 to 89 years; mean age was 63.22 years (SD 13.00), and LVEF ranged from 15.0% to 80.1%.Source: §2.1 · p. 2
2.2 Radar system and acquisition protocol
- The evaluated radar processing used the third receive channel’s complex I/Q samples as input to heart-sound extraction.The system combined a Texas Instruments IWR1443 radio-frequency board with a DCA1000EVM acquisition board; TX0 was active and TX1 and TX2 were disabled.Source: §2.2 · pp. 2–3
- Each recording comprised 60,000 frames at a 1-ms frame period, producing a 1,000-Hz slow-time series over 60 seconds.Each chirp contained 512 complex ADC samples per receive channel at 9.121 MS/s.Source: §2.2 · p. 3
- Patients rested supine while the radar faced the anterior chest from approximately 50–80 cm, and the workflow selected a range cell before phase extraction, filtering, smoothing and quality review.The Figure 1 caption describes maximum-magnitude range-cell selection, outlier removal and manual review; offline reconstruction uses the final selected cell.Source: §2.2 · p. 4, Figure 1 · p. 4
2.3 Extraction of radar heart sounds
Radar heart-sound extraction transforms complex I/Q measurements into a signed vibration waveform through range-cell selection, phase processing, filtering and smoothing. Echocardiographic measurements define the clinical phenotypes used to characterize ventricular and valvular dysfunction.
2.3 Extraction of radar heart sounds
- The extraction pipeline removes each chirp’s fast-time mean, computes a 512-point Fourier transform and selects the candidate range cell with the greatest slow-time mean magnitude.One selected cell is fixed for phase tracking throughout a recording; waveform reconstruction uses the final selected cell, including cases with historical manual adjustment.Source: §2.3 · p. 6
- The selected-cell phase is unwrapped and converted using the retained 79-GHz conversion constant, then filtered at 16–80 Hz and smoothed to produce the signed radar heart sound.Automated outlier removal and manual review mark excluded intervals while maintaining the original time axis.Source: §2.3 · p. 6
2.4 Echocardiographic measurements and phenotype definitions
- Reduced LVEF is defined as LVEF < 50%, while regional wall-motion abnormality is a binary phenotype and valvular disease requires at least moderate regurgitation of any valve.Mild regurgitation alone and stenosis alone are negative under the valvular definition, and the three phenotypes are nonexclusive.Source: §2.4 · p. 6
2.5 Pairing of radar and echocardiographic measurements
The study pairs repeated radar recordings with each patient’s latest echocardiographic assessment and evaluates signal quality, processing reproducibility and reduced-LVEF classification. Patient identity is kept intact across data partitions for patient-level evaluation.
2.5 Pairing of radar and echocardiographic measurements
- All repeated radar measurements from a patient are paired with that patient’s latest echocardiographic assessment, with a maximum three-calendar-month interval between measurements.Clinical summaries use one observation per patient, whereas signal-level experiments use individual radar recordings.Source: §2.5 · p. 7
2.6 Offline validation experiments and statistical analysis
- Technical validation examines 537 recordings for continuity, retained exposure, spectral sensitivity and processing reproducibility using reviewed interval masks on the original time axis.Reconstruction reproducibility is assessed in 12 deterministically selected recordings using Pearson correlation and normalized RMSE of normalized waveform shape.Source: §2.6 · p. 7
2.7 Patient-level evaluation of reduced LVEF
- The reduced-LVEF experiment uses echocardiographic LVEF as its clinical reference and defines the positive target as LVEF < 50%.The model combines the RHS waveform, Hilbert envelope and logarithmic STFT representation, with multiple-instance learning aggregating windows within each recording.Source: §2.7 · pp. 7–8
- Five-fold cross-validation groups patients so that all recordings and windows from one patient remain in a single partition, with model selection on validation data and final testing on held-out patients.Performance measures are patient-level AUROC, PR-AUC, macro-averaged F1 and accuracy.Source: §2.7 · p. 8
2.8 Distribution and overlap of clinical phenotypes
The 162-patient cohort contains three overlapping echocardiographic phenotypes, with subgroup demographics and LVEF variation summarized at the patient level.
2.8 Distribution and overlap of clinical phenotypes
- The cohort included 38 patients with reduced LVEF (23.46%), 43 with wall-motion abnormalities (26.54%), and 39 with valvular disease (24.07%).Figure 2 summarizes the distributions and intersections of these nonexclusive labels.Source: §2.8 · p. 8, Figure 2 · p. 9
- 111 of 162 patients (68.52%) were aged at least 60 years; age-band counts from youngest to oldest were 7, 19, 25, 54, 48, and 9.The age bands were 18–39, 40–49, 50–59, 60–69, 70–79, and 80 years or older.Source: §2.8 · p. 8, Figure 2 · p. 9
- Clinical LVEF was below 40% in 20 patients and 40% to below 50% in 18; 26, 52, and 46 patients fell in the successive higher bands.Those higher bands were 50% to below 60%, 60% to below 70%, and at least 70%.Source: §2.8 · p. 8, Figure 2 · p. 9
- Mean ages in the reduced-LVEF, wall-motion, and valve groups were 62.21, 64.19, and 68.08 years; their mean LVEFs were 37.65%, 50.64%, and 54.87%, respectively.These phenotype-positive subgroups overlap within the same patient population.Source: §2.8 · p. 8
- Of 162 patients, 56 had one target, 23 had two, six had all three, and 77 had none of the three target labels.Six patients with reduced LVEF also carried all three labels; 17 also had wall-motion abnormality and 16 had valvular disease.Source: §2.8 · p. 8, Figure 2 · p. 9
3 Data Records
Each radar measurement is represented by complex signal data, a processed radar heart sound, and time-aligned quality segments, with repeated recordings linked to patient phenotypes.
3 Data Records
- Three complementary records—the original complex signal, processed radar heart sounds (RHS), and temporal quality segments—link acquired I/Q data to vibration waveforms and analysis intervals.The complex signal and processed RHS are supplied in MATLAB format; quality segments retain their positions on the original time axis.Source: §3 · p. 9, Table 3 · p. 10
- The complex representation contains 60,000 slow-time observations and 512 complex fast-time samples per observation, supporting re-extraction with alternative range-selection or filtering strategies.Source: §3 · p. 10
- Each RHS is a continuous 60,000-sample signed waveform at 1,000 Hz; retained and excluded intervals remain aligned to the original 60-second axis.The segmentation supports consistent retained-exposure calculations and quality-aware spectral features.Source: §3 · p. 10
- Repeated recordings share each patient’s clinical phenotype vector, supporting within-patient signal analysis and comparisons across overlapping clinical groups.Source: §3 · p. 10
4 Technical Validation
Technical checks characterize retained signal coverage, spectral sensitivity, waveform reconstruction, and patient-level clinical evaluation. The section reports robust signal continuity and close reconstruction agreement while defining the clinical reference and validation scope.
4.1 Signal continuity and quality-aware exposure
- 95.93% of nominal signal duration was retained, with a median retained duration of 58.77 s and a minimum of 40.04 s.Quality exclusions affected 330 recordings (61.45%); edge exclusions for spectral analysis are additional to these coverage estimates.Source: §4.1 · p. 10
4.2 Spectral sensitivity to quality masking and window length
- Quality masking changed the 16–80-Hz centroid by a median absolute 0.213 Hz, with a maximum change of 7.828 Hz.The reported differences show that temporal quality segmentation can materially affect energy-weighted spectral features.Source: Figure 3 · p. 11, §4.2 · p. 12
- Increasing the analysis window from 2.048 to 4.096 s changed the centroid by a median absolute 0.073 Hz, while first- versus last-half centroids had Spearman correlation 0.815.Window-length sensitivity was typically smaller than masking-associated change or within-record variation, although individual recordings differed.Source: §4.2 · p. 12
4.3 Reconstruction from raw complex radar samples
- All 12 sampled RHS and reconstructed normalized waveforms had Pearson correlation above 0.999999.The reconstruction comparison also reported normalized RMSE from 2.65×10^-10 to 1.23×10^-6; comparisons were centered and RMS-normalized.Source: §4.3 · p. 12, Figure 5 · p. 14
4.4 Clinical reference and patient-level phenotypes
- The clinical reference uses one latest echocardiographic assessment per patient, with reduced LVEF defined by a strict 50% threshold.The cohort includes 38 reduced-LVEF patients, 43 with regional wall-motion abnormality, and 39 with at least moderate valvular regurgitation.Source: §4.4 · p. 12
4.5 Discrimination of reduced left ventricular systolic function
- Table 4 reports signal and functional validation; for LVEF discrimination, AUROC summarizes the five outer test folds in the 162-patient study cohort.Source: Table 4 · p. 14
5 Usage Notes
The usage notes emphasize quality-aware signal handling, patient-level evaluation, and preserving the study’s nonexclusive clinical phenotype definitions.
5 Usage Notes
- Use windows wholly contained in retained intervals and specify a minimum retained-duration criterion before supervised modelling.The supplied masks and retained-duration summaries provide the temporal information needed to apply that criterion.Source: §5 · p. 15
- Preserve nonexclusive phenotype definitions: reduced LVEF uses the strict 50% threshold, and valvular disease means at least moderate regurgitation of any valve.The 77 patients without target labels form the three-target-negative group; report-level methods, age, sex, and coexisting phenotypes are relevant covariates.Source: §5 · p. 15
- Keep each patient’s recordings together across training, validation, and test partitions to align evaluation with patient-level echocardiographic targets.Report patient-level denominators, label overlap, missingness, and class-specific performance with uncertainty.Source: §5 · p. 15
6 Code Availability
The study software supports radar acquisition and the signal-processing and validation workflow.
6 Code Availability
- The software implements radar control, complex-signal processing, range selection, phase-based RHS extraction, quality segmentation, and technical validation experiments.Evaluated operations and parameters are specified in the Methods.Source: §6 · p. 15
- The acquisition and analysis software is maintained with the study materials.The described implementation covers both radar control and downstream signal analysis.Source: §6 · p. 15
- The study materials include software for the technical validation experiments as well as the acquisition and analysis workflow.The Methods specify the evaluated operations and parameters.Source: §6 · p. 15
7 Data Availability
The radar measurements are maintained in original and processed forms, with temporal quality segmentation and patient-level echocardiographic phenotypes.
7 Data Availability
- Original complex-signal MATLAB data and processed RHS MATLAB data are maintained by the investigators.Each measurement is associated with its patient-level echocardiographic phenotype.Source: §7 · p. 15
- Temporal quality segmentation is maintained alongside the radar measurement data.The measurements retain their association with the corresponding patient-level echocardiographic phenotype.Source: §7 · p. 15
- Each radar measurement is associated with the corresponding patient-level echocardiographic phenotype.The maintained data include both original complex signals and processed RHS data.Source: §7 · p. 15
A Supplementary Information
The supplementary figures and methods describe radar acquisition timing and the phase-based extraction of radar heart sounds, followed by quality-aware spectral sampling.
A Supplementary Information
- The RHS reconstruction derives the waveform from complex RX3 data by selecting a range cell, extracting and unwrapping its phase, then converting and filtering it.Processing uses a 79-GHz amplitude conversion constant, forward–backward filtering and Savitzky–Golay smoothing.Source: A Supplementary Information · p. 17, Figure S2 · p. 18
- Spectral analysis preserves the original time axis and includes only complete windows contained in reviewed retained intervals, excluding one second at each recording edge.No fragments are concatenated; the first and last second are additionally excluded for spectral eligibility.Source: A Supplementary Information · p. 17, Figure S2 · p. 18
- One chirp per 1-ms frame yields 60,000 temporal observations over 60 s, with four receive channels recorded in the confirmed 1TX4RX configuration.The evaluated receive-channel representation contains 60,000 by 512 complex I/Q samples; the sampled aperture is distinguished from the full FMCW ramp.Source: Figure S1 · p. 17
- Periodic Hann windows of 2,048 and 4,096 samples at 1,000 Hz, each with 50% overlap, define the spectral-analysis weighting functions.The reviewed intervals mark retained segments in blue and exclusions in gray.Source: Figure S2 · p. 18