PaperlayerView exact revision on arXiv

Source-linked AI summary

Physiologically Informed Digital Auscultation for Pneumonia Detection in Long-term Care Residents

Nicholas Rasmussen, Oleg Zaslavsky, Zih-Ling Wang, Hongyu Yu, Joelle Fathi, Kaibao Nie, Amil Khanzada, Tomoko Ito

arXiv:2609.27222v1cs.SDcs.CVcs.LGeess.ASq-bio.TO

TL;DR

Prospective evidence for digital lung-sound pneumonia detection in older long-term-care residents is limited. This study evaluated supervised multichannel recordings and found strongest stethoscope-only performance under X-ray supervision, with reduced-channel deployment preserving performance.

  • Problem

    Prospective evidence for digital lung-sound diagnosis in older long-term-care residents is limited, while variable auscultation interpretation and setting-specific constraints contribute to diagnostic uncertainty.

    Source: §1 · p. 2

  • Method

    The prospective study tests standardized multi-channel digital lung sounds under radiologist-confirmed chest-X-ray and clinician diagnostic supervision, and assesses reduced-channel protocols and model attention.

    Source: §1 · p. 3

  • Results

    XRAY supervision yielded strongest, least variable stethoscope-only performance (F1 = 0.729; accuracy = 0.783), while three/four-channel deployment preserved or slightly improved performance (Test 3/4: F1 0.736/0.733; specificity 0.895/0.863).

    Source: §3 · p. 9

  • Takeaways & Limitations

    The findings support a practical reduced-site bedside protocol for long-term care that preserves discrimination while reducing variability.

    Source: §3 · p. 9

    The modest cohort (185 residents, 73 pneumonia cases) leaves subgroups beyond dementia too small for reliable estimates.

    Source: §3 · p. 10

Abstract

from arXiv · show

Pneumonia is difficult to diagnose in older long-term care residents; multimorbidity and atypical presentations obscure signs, motivating operationally efficient objective testing. We analyzed multi-channel digital stethoscope recordings from 185 Japanese residents (73 pneumonia, 112 symptomatic without), using radiologist-confirmed chest X-rays and clinician diagnoses as supervisory signals that train convolutional neural networks, multimodal fusion, and channel-based variants with time-domain Grad-CAM interpretability. Models were evaluated with repeated patient-level cross-validation showing models with X-ray supervision outperformed clinician supervision (F1 0.729, accuracy 0.783 vs. F1 0.637, accuracy 0.711). Additionally, a three-channel selection protocol maintained performance (F1 0.736; accuracy 0.803), with two mid-thoracic sites ranking highest and Grad-CAM attention overlapping adventitious sounds. These findings indicate automated multi-channel lung-sound analysis can aid long-term care pneumonia diagnosis, with X-ray supervision being more reliable than clinical, and fewer channels preserving performance while lowering acquisition times.

1 Introduction

The study addresses diagnostic uncertainty and practical barriers to pneumonia assessment in long-term care by evaluating patient-level digital lung-sound models and whether fewer recording sites can preserve diagnostic information. It also examines whether model attention aligns with respiratory phases and annotated adventitious sounds.

1 Introduction

  • LTC residents face elevated pneumonia susceptibility and severity, while atypical presentations and limited onsite testing, staffing, transport, and radiographic capabilities contribute to diagnostic uncertainty.Practical bedside diagnostics deployable within LTC facilities are needed.Source: §1 · p. 2
  • Human auscultation has limited sensitivity and variable documentation, with clinical variability, ambient noise, transient sounds, and limited prescriber oversight further constraining assessment in LTC.These limitations can contribute to diagnostic uncertainty and inappropriate antibiotic use.Source: §1 · p. 2
  • Digital stethoscopes can standardize acquisition and enable automated lung-sound analysis, but reported near-perfect accuracy may reflect dataset artifacts such as recording site, equipment, or protocol.Early studies suggest potential to reduce inter-observer variability and provide reproducible, auditable outputs.Source: §1 · p. 2
  • Evidence from prospective elderly cohorts is scarce, and prior pediatric, emergency-department, or mobile-phone studies may not reflect LTC residents’ comorbidity, cognitive impairment, immobility, or protocol constraints.These differences motivate targeted prospective cohorts and feasible recording protocols.Source: §1 · p. 2
  • LTC deployment requires patient-level validation in the target population and recording protocols that balance diagnostic yield with bedside feasibility.Multi-site recording may capture localized pathology but can burden frail or bedridden residents; reducing channels while preserving signal could improve adoption.Source: §1 · pp. 2–3
  • The study evaluates patient-level detection of pneumonia versus symptomatic non-pneumonia using radiologist-confirmed chest X-rays and clinician diagnoses, then tests recording-site contributions and interpretable attention.It also assesses whether a reduced-channel protocol can preserve performance while lowering bedside burden.Source: §1 · p. 3

2 Results

In 185 long-term-care residents, X-ray-supervised stethoscope models outperformed clinician-label models, while selected channel subsets retained strong performance. Channel rankings and expert-annotation alignment further localized informative lung-sound signals.

Physiologic Measures. / Categorical Variables.

  • Among 185 residents, clinicians classified 73 as pneumonia and 112 as non-pneumonia; VAS scores separated groups most strongly, while white sputum, thick sputum, and constant dyspnea were more common with pneumonia.Pneumonia patients had lower oxygen saturation (94.7 ± 3.67% vs. 95.7 ± 2.76%) and higher pulse rates (81.7 ± 15.5 vs. 76.6 ± 13.7 bpm); VAS scores were 34.8 ± 26.7 vs. 22.8 ± 18.4. Comorbidities showed no strong univariate associations, and tracheostomy had a borderline association.Source: Cohort Composition · p. 3, Physiologic Measures · pp. 3–4, Categorical Variables · pp. 4–5

XRAY vs. Diag Labels.

  • X-ray supervision outperformed clinician diagnostic labels across architectures; the stethoscope-only model achieved F1 0.729 and accuracy 0.783.It also had NPV 0.789, sensitivity 0.686, and PPV 0.805.Source: XRAY vs. Diag Labels · p. 5

Multimodal Fusion.

  • The fusion model achieved the highest AUC (.791) and specificity (.867), while stethoscope-only models had the lowest variability for F1, accuracy, NPV, and sensitivity.Fusion also showed low inter-split variability for AUC (SD = 0.016), specificity (SD = 0.017), and sensitivity (SD = 0.019).Source: Multimodal Fusion · p. 5, Variability Patterns · p. 5

Architecture Comparison.

  • Mean-Probability ARP–N was strongest and most stable, with F1 = 0.729, accuracy = 0.783, AUC = 0.774, and specificity = 0.851.Channel-Stacked Attention and Hierarchical MIL had weaker AUCs of 0.689 and 0.657, respectively, with lower specificity and higher variability.Source: Architecture Comparison · p. 5

Train-all, Test-subset Ablations.

  • Train-all, test-subset inference achieved F1 = 0.736 and accuracy = 0.803 with three channels, and F1 = 0.733 and accuracy = 0.791 with four.Three-channel testing had accuracy SD = 0.009 and F1 SD = 0.013; single-channel inference retained S90 = 0.332 and S80 = 0.605.Source: Train-all, Test-subset Ablations · p. 7

Matched Train/Test Ablations.

  • Matched three-channel training and testing achieved F1 = 0.730, accuracy = 0.773, and AUC = 0.765, whereas single-channel inference had sensitivity 0.761 but specificity 0.637.Two-channel training yielded F1 = 0.715, accuracy = 0.760, and AUC = 0.739.Source: Matched Train/Test Ablations · p. 7

Channel Importance.

  • Channels 3 and 4 formed the dominant ranking tier, with Borda scores of 119 and 101; Channel 5 ranked lowest at 67.Channel 3 was the modal top-ranked site, and the hierarchy was robust across seeds, folds, and label sets.Source: Channel Importance · p. 7

Attention Alignment with Expert Annotations.

  • Attention overlapped annotated acoustic regions, covering on average 74% of inspirations, 64% of expirations, 61% of rales, 82% of rhonchi, and 81% of wheezes.Alignment was strongest for the pulmonary nurse practitioner, including 82% of inspirations and 71% of rales.Source: Attention Alignment with Expert… · p. 7

3 Discussion

This prospective study shows automated lung-sound analysis can distinguish pneumonia from non-pneumonia in long-term care residents. Its discussion emphasizes label choice, reduced-channel deployment, and physiologically coherent evidence under rigorous patient-level evaluation.

Discussion

  • XRAY-supervised models were more stable and discriminative than Diag-supervised models, with the stethoscope-only model achieving F1 = 0.729 and accuracy = 0.783.Clinical diagnosis can be noisy in long-term care, while radiography provides a more reproducible, though imperfect, reference; fusion sharpened decision thresholds (AUC = 0.791; specificity = 0.867).Source: §3 · pp. 8–9
  • Three- or four-site deployment preserved or slightly improved performance after training on all six channels (Test 3/4: F1 0.736/0.733; specificity 0.895/0.863).Multi-seed Borda aggregation ranked anterior channels 3 and 4 highest; these mid-thoracic sites offer quicker, more accessible auscultation for supine residents, not equivalence to posterior auscultation.Source: §3 · p. 9
  • Downsampling to fs = 1000 Hz retained low-frequency pathology concentrated below ∼300 Hz while reducing artifacts and computational cost.Models attended primarily to rhonchi and secondarily to crackles; brief crackle transients can resemble handling or clipping, and human crackle judgments have reliability limitations.Source: §3 · p. 9
  • Strict patient-level partitioning and multi-seed evaluation yielded AUC of approximately 0.77–0.79, aligned with DeepBreath (internal AUROC 0.75, external 0.74).Public-repository accuracies of 90–100% may reflect acquisition artifacts; shortcut learning can inflate apparent performance by roughly 20%, while bias correction can reduce reported accuracy by up to 30% [25].Source: §3 · pp. 9–10
  • The study was limited by a modest, single-center cohort, a mid-study domain shift, imperfect radiographic labels, symptom-and-fever enrollment, and no external validation.Only the dementia subgroup supported interpretable subgroup analysis (N = 45; 14 positives); other subgroups were too small for reliable estimates.Source: §3 · p. 10
  • Future work should validate the anterior three/four-channel protocol in adequately powered, independent multi-site cohorts and compare it with bedside standards, particularly lung ultrasound.Clinical utility requires prospective outcome-oriented trials testing effects on unnecessary transfers, antibiotic overuse, or diagnostic delay, alongside SaMD regulatory advancement.Source: §3 · p. 10

4 Methods · 4.1 Dataset Acquisition

The study collected standardized multi-site lung sounds and clinical data from older long-term care residents with suspected pneumonia. Radiographic and clinician assessments provided distinct reference labels for a cohort of 185 patients.

4.1 Dataset Acquisition

  • The study collected lung sounds from patients with suspected pneumonia and upper respiratory symptoms between December 2025 and June 2026 under University of Tsukuba IRB approval 2151; it was not prospectively registered.Source: §4.1 · p. 10

Participants.

  • Participants were adults aged ≥80 receiving skilled nursing care in Ibaraki, Japan, with suspected infection based on respiratory symptoms and fever above 37.5°C.Cognitive impairment was exclusionary unless a legally authorized representative could consent; participants consented directly or through a representative and were not involved in study design.Source: Participants · pp. 10–11

Clinical Reference and Labels.

  • Chest radiographs were reviewed by a radiologist for XRAY labels, while clinicians assigned Diag labels under Japanese Respiratory Society guidelines.Diag labels integrated vital signs, auscultation, laboratory and microbiology findings, differential assessment, and radiographic interpretation.Source: Clinical Reference and Labels · p. 11

Digital Stethoscope.

  • Trained nurses recorded multi-site anterior lung sounds with an Eko Core 2 at a fixed 4 kHz export rate, retaining native conditioning without additional noise suppression or post-processing.A subset of 45 recordings used on-device screen recording that bypassed parts of the device pipeline; formal acoustic characterization was limited by proprietary specifications.Source: Digital Stethoscope · p. 11

Auscultation Protocol

  • Six anterior chest locations were recorded in fixed order for 15 seconds per site; early 30-second recordings were clipped to yield about 28 seconds of usable signal.Participants sat upright when possible; alternate positions and protocol deviations were documented.Source: Auscultation Protocol · p. 11

Metadata and Clinical Variables.

  • Vital signs and demographic and clinical characteristics were recorded, with acquisition logs capturing recording site, duration, patient position, and protocol deviations.Source: Metadata and Clinical Variables · p. 11

Sample Size Justification

  • 185 residents were included: 73 with pneumonia, 112 without pneumonia, and 77 with radiographic opacities; the sample was deemed sufficient without prospective power calculation.The estimated 95% CI half-width was roughly 7–9 percentage points for sensitivity and 6–8 points for specificity.Source: Sample Size Justification · pp. 11–12

4.2 Signal Processing and Feature Extraction

Recordings were standardized, segmented into overlapping windows, and converted into fixed-size log-power spectrograms for model input.

4.2 Signal Processing and Feature Extraction

  • All recordings were resampled to 1000 Hz, then de-meaned and variance-normalized per recording.Resampling used librosa’s default resampy backend with a zero-phase Kaiser-windowed sinc anti-aliasing filter and polyphase resampling.Source: §4.2 · p. 12
  • 15 s and 28 s recordings were divided into three or five overlapping 8192-sample windows with a 4096-sample hop.The final window was shifted to end at the signal endpoint, adding terminal overlap to accommodate recording length without introducing artifact.Source: §4.2 · p. 12
  • Each window became a min–max normalized 64 × 256 log-power spectrogram using an STFT with 128-sample frames and a 32-sample hop.The frequency resolution was 7.8125 Hz; the first 64 frequency bins, covering 0–500 Hz, were retained, and the time axis was fixed at 256 frames.Source: §4.2 · p. 12

4.3 Model Development

Model development used patient-level validation and class weighting with a shared ARP–N backbone, alongside channel-ranking, channel-aware, and clinical-fusion variants. Channel selection was made interpretable through cross-validated Borda aggregation, while other optimization settings were held constant.

4.3 Model Development

  • ARP–N models used patient-level splits and patient-level class weights to prevent leakage and address imbalance without resampling.The ARP–N backbone operates on log-power spectrograms; CRNN and CANN variants were not selected because they offered marginal or negative gains under comparable conditions [48].Source: §4.3 · p. 12

Borda Scoring

  • Borda aggregation summed channel rankings across 25 patient-level folds to produce a stable, seed-averaged hierarchy that was not used for outer-loop channel masking.Each channel received 6−1 points according to its rank in each fold, reducing fold-specific variability and limiting the influence of a single favorable ranking.Source: Borda Scoring · p. 13

Channel and MIL Architectures.

  • Channel-stacked and multi-instance-learning architectures were evaluated against the ARP–N mean-probability baseline.The channel-stacked model concatenated six sites into a multi-channel tensor with spatial filters; multi-instance learning treated windows as independent instances for patient-level aggregation.Source: Channel and MIL Architectures · p. 13

Hyperparameters

  • Channel-selection procedures varied in the inner loop, while optimization settings were held constant and models were optimized for the best attainable F1 score.Fixed settings included batch size, learning-rate schedule, regularization, and early-stopping criteria.Source: Hyperparameters · pp. 13–14

4.4 Annotation and Attention Map Construction

Four blinded annotators marked respiratory events and signal degradation on spectrograms, while model attention was derived from Grad-CAM maps and compared with annotator event counts. The alignment analysis was computed separately for each annotator and aggregated across feature types.

4.4 Annotation and Attention Map Construction

  • Four blinded annotators labeled respiratory events and absence events on spectrograms, with free-text notes for signal degradation.They included two nurses, a signal-domain expert, and an experienced pulmonary nurse practitioner; multiple rales within 1 second counted as one event.Source: §4.4 · p. 14
  • Grad-CAM maps from each spectrogram window were resized and stitched by averaging overlaps, then collapsed across frequency and thresholded into a binary temporal mask.An annotated region counted as attended if any temporal column in the mask overlapped it; partial overlaps counted.Source: §4.4 · p. 14
  • Attention alignment was calculated separately for each annotator by dividing attended time columns by event counts for six acoustic feature types.The analysis aggregated across annotators to assess alignment with shared perceptual regions.Source: §4.4 · p. 14

4.5 Statistical Analysis and Reporting Standards

The study described cohort characteristics and tested univariate associations separately from predictive modeling. Model performance was assessed through repeated patient-level cross-validation with diagnostic metrics and transparent reporting standards.

Cohort’s Descriptive Analysis

  • Continuous cohort variables were summarized with means, standard deviations, medians, and interquartile ranges, while categorical variables were reported as counts and proportions.Source: Cohort’s Descriptive Analysis · pp. 14–15
  • Robust two-tailed logistic regression estimated univariate p-values for continuous and categorical predictors, using numeric coercion and dummy encoding, without generating predictive performance.Source: Cohort’s Descriptive Analysis · p. 15

Model Evaluation

  • Model performance used a fixed 5 × 5 patient-level cross-validation design, yielding 25 independent test evaluations and fold-level metrics from patient-level probabilities.The evaluation included sensitivity-constrained specificity at 90% and 80% sensitivity (S90 and S80), PPV, NPV, and DeLong AUC; definitions and variability calculations were provided in Extended Data. Reporting followed TRIPOD-AI; external validation and powered subgroup analyses were infeasible, so repeated cross-validation supported robust internal evaluation.Source: Model Evaluation · p. 15

Data Availability

Raw lung-sound recordings and clinical data cannot be publicly released because of privacy and re-identification risks. A minimal de-identified derivative dataset will be available under controlled access for non-commercial research.

Data Availability

  • Raw recordings, clinical metadata, and protocols cannot be publicly released because of privacy and re-identification risks.No third-party proprietary datasets were used.Source: Data Availability · p. 15
  • The controlled-access derivative dataset will include log-power spectrograms, multimodal-model clinical variables, and preprocessing/STFT documentation.After acceptance, it will be deposited for at least five years for non-commercial research, subject to a brief proposal and ethical and legal restrictions.Source: Data Availability · p. 15
  • Editors and reviewers may request access and receive it within 10 business days after signing an agreement prohibiting re-identification, redistribution, and secondary use.Source: Data Availability · p. 15

Code Availability

The authors will provide custom code during peer review and publish the documented, versioned codebase upon acceptance. Supplementary materials include extended methodological details, analyses, and additional materials available on request.

Code Availability

  • The custom code will be shared with editors and reviewers on request during peer review, then published with documentation and versioning upon acceptance.The release will include preprocessing, spectrogram generation, model development, cross-validation, and attention-map construction, with no proprietary third-party code beyond standard open-source libraries.Source: Code Availability · p. 15

Supplementary Information.

  • Supplementary information provides methodological details, ablations, stability and subgroup analyses, while annotations, attention maps, and annotation counts are available on request.The ablations include capture fidelity and sampling-rate tuning; extended data includes mathematical formulations.Source: Supplementary Information · pp. 15–16

Competing Interests

N.R. previously received salary support from Virufy, whose founder A.K. contributed to study design and manuscript review as a co-author; the organization provided no funding or study direction, and all other authors declared no competing interests.

Competing Interests

  • N.R. previously received salary support from Virufy; founder A.K. contributed to study design and manuscript review as a co-author, but Virufy provided no funding or study direction.Virufy is a non-profit developing AI-based respiratory diagnostic tools. It had no role in data collection, analysis, interpretation, or manuscript preparation.Source: Competing Interests · p. 17
  • All other authors declared no competing financial or non-financial interests.Source: Competing Interests · p. 17
Loading 2609.27222v1…