Leakage-aware supervised machine-learning classification of normal and abnormal calf lung-sound recordings: a proof-of-concept study

Digital auscultation with supervised machine learning (ML) may support objective calf lung-sound interpretation. This proof-of-concept study evaluated supervised ML using a curated multicenter dataset and leakage-aware validation. Digital lung-sound recordings were collected from pre-weaned dairy and dairy-crossbred calves in Canada, Italy, and Sweden, using a standardized digital auscultation protocol and were independently reviewed by two observers, with discrepant initial classifications resolved by consensus. Of 1,360 recordings, 424 met the predefined audio-quality criteria; 39 were subsequently excluded under the acoustic inclusion criteria and five duplicate copies were removed, yielding 380 unique recordings for analysis. The 380 retained recordings were each assigned one of five single-label consensus acoustic categories: Normal sounds, Crackles, Wheezes, Increased Breath Sounds (IBrS), and Increased Bronchial Sounds (IBS). Recordings were partitioned into 304 training and 76 held-out internal test recordings, with all recordings from the same calf (ear-tag identifier) kept in the same partition. Feature sets and tree-based ML algorithms were evaluated using five-fold stratified group-aware cross-validation. The primary task distinguished Normal from Abnormal recordings, with Abnormal comprising recordings assigned to Crackles, Wheezes, IBrS, or IBS. Two exploratory single-label multiclass tasks were evaluated secondarily: a three-class task comprising Normal sounds, Adventitious sounds (Crackles and Wheezes), and Increased sounds (IBrS and IBS), and a five-class task retaining the five individual consensus acoustic categories. On the held-out internal test set, a 94-feature non-amplitude Extra Trees classifier achieved accuracy of 0.816, balanced accuracy of 0.814, macro-F1 of 0.802, area under the ROC curve of 0.930, and average precision of 0.963; sensitivity and specificity for Abnormal recordings were 0.820 and 0.808. Exploratory three- and five-class macro-F1 values were 0.607 and 0.470, respectively. In this curated multicenter dataset, supervised ML showed promising discrimination between Normal and Abnormal calf lung-sound recordings under group-aware validation, whereas finer single-label acoustic-category classification remained less reliable. These findings warrant external validation of automated lung-sound analysis as an objective aid for respiratory screening and health monitoring in pre-weaned dairy calves.

Authors

Institutions

Publication Details

Journal
Frontiers in Animal Science
Published
2026-11-18
DOI
https://doi.org/10.3389/fanim.2026.1966658
Primary Topic
Phonocardiography and Auscultation Techniques
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Leakage-aware supervised machine-learning classification of normal and abnormal calf lung-sound recordings: a proof-of-concept study

Antonio Boccardo, Cristina Cravana, Elisa Gazzola, Francesco Rubino et al.
Frontiers in Animal Science
Phonocardiography and Auscultation Techniques
article

Leakage-aware supervised machine-learning classification of normal and abnormal calf lung-sound recordings: a proof-of-concept study

Antonio Boccardo, Cristina Cravana, Elisa Gazzola, Francesco Rubino, Salvatore Ferraro, Giuseppe Tardiolo, Sébastien Buczinski
article en

Abstract

Digital auscultation with supervised machine learning (ML) may support objective calf lung-sound interpretation. This proof-of-concept study evaluated supervised ML using a curated multicenter dataset and leakage-aware validation. Digital lung-sound recordings were collected from pre-weaned dairy and dairy-crossbred calves in Canada, Italy, and Sweden, using a standardized digital auscultation protocol and were independently reviewed by two observers, with discrepant initial classifications resolved by consensus. Of 1,360 recordings, 424 met the predefined audio-quality criteria; 39 were subsequently excluded under the acoustic inclusion criteria and five duplicate copies were removed, yielding 380 unique recordings for analysis. The 380 retained recordings were each assigned one of five single-label consensus acoustic categories: Normal sounds, Crackles, Wheezes, Increased Breath Sounds (IBrS), and Increased Bronchial Sounds (IBS). Recordings were partitioned into 304 training and 76 held-out internal test recordings, with all recordings from the same calf (ear-tag identifier) kept in the same partition. Feature sets and tree-based ML algorithms were evaluated using five-fold stratified group-aware cross-validation. The primary task distinguished Normal from Abnormal recordings, with Abnormal comprising recordings assigned to Crackles, Wheezes, IBrS, or IBS. Two exploratory single-label multiclass tasks were evaluated secondarily: a three-class task comprising Normal sounds, Adventitious sounds (Crackles and Wheezes), and Increased sounds (IBrS and IBS), and a five-class task retaining the five individual consensus acoustic categories. On the held-out internal test set, a 94-feature non-amplitude Extra Trees classifier achieved accuracy of 0.816, balanced accuracy of 0.814, macro-F1 of 0.802, area under the ROC curve of 0.930, and average precision of 0.963; sensitivity and specificity for Abnormal recordings were 0.820 and 0.808. Exploratory three- and five-class macro-F1 values were 0.607 and 0.470, respectively. In this curated multicenter dataset, supervised ML showed promising discrimination between Normal and Abnormal calf lung-sound recordings under group-aware validation, whereas finer single-label acoustic-category classification remained less reliable. These findings warrant external validation of automated lung-sound analysis as an objective aid for respiratory screening and health monitoring in pre-weaned dairy calves.

Frontiers in Animal ScienceVol. 7
University of Messina (IT), Universitat Oberta de Catalunya (ES), University of Milan (IT), Swedish University of Agricultural Sciences (SE), Ruđer Bošković Institute (HR), Université de Montréal (CA)
Good health and well-being
Phonocardiography and Auscultation Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon