Combining electrocardiographic measurements may improve identification of dogs requiring echocardiographic assessment. This retrospective diagnostic accuracy study derived a three-category algorithm in 150 dogs and evaluated it in 50 separate validation dogs. Three observers independently measured lead II P-wave duration and the V1 Morris index, with observer A providing the primary analysis. Left atrium-to-aorta ratio (LA/Ao) ≥1.6 was the primary outcome; normalized left ventricular internal diameter in diastole (LVIDDN) ≥1.7 was evaluated independently as a secondary outcome. Thresholds from 114 reference dogs were 48 and 52 ms for P duration and 3.04 mV·ms for Morris. Positive triage comprised intermediate and high categories. Primary-outcome sensitivity and specificity were 90.6% and 50.0% in derivation and 100% and 85.7% in validation. Validation sensitivity had a 95% confidence interval of 88.1%–100%. The ordinal algorithm area under the curve was 0.961 in validation, compared with 0.868 for P and 0.844 for Morris; paired differences were not significant after Holm adjustment. Secondary-outcome validation sensitivity and specificity were 88.0% and 60.0%. The algorithm provided complementary detection of left atrial enlargement and reproducible measurements across observers, supporting further evaluation as an adjunct to echocardiographic triage.

Study population

Reporting was guided by STARD 2015 and STROBE-Vet recommendations (Bossuyt et al. 2015; Cohen et al. 2016; O’Connor et al. 2016). This retrospective diagnostic accuracy study included 150 dogs in a derivation cohort and 50 dogs in a separate validation cohort. Derivation examinations were performed at the veterinary diagnostic imaging center NUDIVE, Federal University of Piauí (UFPI), Brazil, between July 2024 and June 2025. Eligibility required a 12-lead electrocardiogram (ECG) and complete echocardiographic examination obtained on the same day. All 200 dogs were included in the principal analyses. The sample comprised the available eligible records; no prospective sample-size calculation was documented. American College of Veterinary Internal Medicine (ACVIM) stages were recorded using the consensus framework (Keene et al. 2019). The study was approved by the UFPI Animal Use Ethics Committee (CEUA/UFPI; protocol 824/2024), and written owner consent was obtained.

Electrocardiographic acquisition and measurements

ECGs were recorded with a digital 12-lead INCardio X system with dogs in right lateral recumbency without sedation, at 50 mm/s and 10 mm/mV, for 3 min. V1 placement followed the precordial lead system described by Santilli et al. (2019). Three observers, designated A, B, and C, independently measured the ECGs while blinded to the other readings and clinical and echocardiographic findings. When baseline instability, artifacts, or indistinct wave boundaries prevented reliable measurement, other readable regions of the same tracing were examined under the same blinding conditions.

Each stored ECG value represented the automatically generated mean of three measurements. Individual measurements were not retained; therefore, within-tracing repeatability could not be assessed. P-wave duration in lead II was expressed in milliseconds. The Morris index was calculated as the duration of the terminal negative P-wave component in V1 multiplied by its absolute amplitude, in mV·ms. Measurement of the terminal negative component and calculation of the index are illustrated in Fig. 1. P-wave termination was referenced to the isoelectric T–P baseline when the P–R segment was sloped. In derivation, the PR interval was calculated as lead II P duration plus the PR segment, and the Macruz index as P duration divided by the PR segment. PR and Macruz were ancillary variables and did not enter the algorithm.

Echocardiographic reference

One echocardiographer, blinded to ECG results, performed the reference examinations using a GE Vivid iq v203 system. LA/Ao was measured in early diastole in the right parasternal short-axis view at the aortic valve level, following the ACVIM approach (Keene et al. 2019). LA/Ao ≥1.6 defined the primary outcome of left atrial enlargement. Left ventricular internal diameter in diastole (LVIDd), LVIDDN, and peak early diastolic transmitral E velocity were measured by the same echocardiographer. LVIDDN was calculated as LVIDd in centimeters divided by body weight in kilograms raised to 0.294 (Cornell et al. 2004). LVIDDN ≥1.7 defined the secondary outcome, independently of LA/Ao. A single reference examination was shared by all ECG observers for each dog.

Data management and algorithm derivation

Records were matched by animal identifier. Checks included completeness, duplicate identifiers, agreement of shared echocardiographic measurements, consistency of derived variables, and the relationship between terminal V1 duration and total P duration in lead II. Each dog contributed one stored value per variable and observer. No principal measurement was imputed. Observer A provided the primary diagnostic analysis; B and C were evaluated separately without averaging ECG readings across observers.

The reference subgroup comprised derivation dogs in ACVIM stages A or B1 with LA/Ao <1.6. Empirical type 7 quantiles from observer A defined the 70th and 90th percentiles of P duration and the 70th percentile of Morris. The resulting decision rule assigned score 2 when P ≥52 ms or when P ≥48 ms and Morris ≥3.04 mV·ms; score 0 required P <48 ms and Morris <3.04 mV·ms; all other combinations received score 1 (Fig. 2). Categories were termed low, intermediate and high and represented increasing electrical evidence of enlargement, rather than predicted future events. Scores 1–2 defined positive triage, equivalent to P ≥48 ms or Morris ≥3.04 mV·ms. Score 2 alone was assessed as a more selective operating point. All thresholds were fixed for validation, other observers and the secondary outcome.

Statistical analysis

Continuous characteristics were summarized using medians and interquartile ranges; counts and percentages described categorical variables. Diagnostic performance included sensitivity, specificity, positive and negative predictive values, accuracy and likelihood ratios. Proportions received exact Clopper–Pearson 95% confidence intervals (CIs) (Clopper and Pearson 1934); likelihood-ratio intervals used the log approximation when estimable. Receiver operating characteristic analysis used the ordinal score and continuous P and Morris values. Areas under the curve (AUCs), CIs, and within-cohort paired comparisons were calculated using DeLong covariance estimates (DeLong et al. 1988). Between-cohort results were described separately.

Paired proportions were compared within McNemar’s framework (McNemar 1947), using two-sided exact conditional tests for sensitivity among outcome-positive dogs, specificity among outcome-negative dogs, and accuracy among all dogs. This implementation controls the nominal type I error rate but can be conservative and have reduced power (Fagerland et al. 2013). Holm adjustment was applied to the three marker pairs within each cohort, observer, outcome, and metric (Aickin and Gensler 1996). Associations between categories and outcomes were assessed with the Fisher–Freeman–Halton exact test; pairwise category comparisons used Fisher exact tests with Holm adjustment. B and C underwent the same analyses against the shared reference. Paired observer comparisons used corresponding three-comparison families. All tests were two-sided; adjusted p <0.05 indicated statistical evidence within the specified family.

Between-observer agreement of stored P and Morris means was assessed using the two-way absolute-agreement intraclass correlation coefficient for a single observer value, ICC(A,1) (McGraw and Wong 1996; Koo and Li 2016). Confidence intervals were estimated using 3,000 dog-level bootstrap samples, with complete observer sets retained during resampling (Efron 1979). Bland–Altman bias and limits of agreement were calculated for each observer pair (Bland and Altman 1986). Linear-weighted kappa assessed ordinal scores (Cohen 1968), and unweighted kappa assessed triage decisions (Cohen 1960). Friedman and paired Wilcoxon tests examined observer differences (Friedman 1937; Wilcoxon 1945).

An exploratory stratified five-fold derivation check re-estimated thresholds within each training reference subgroup and applied them to held-out dogs. This internal check evaluated the development procedure within the available derivation sample, consistent with the distinction between apparent performance and performance assessed by resampling (Steyerberg et al. 2001). Additional analyses included ancillary PR/Macruz discrimination, descriptive Spearman correlations, and exploratory net benefit of fixed referral strategies (Vickers and Elkin 2006; Vickers et al. 2019). Supplementary methods, tables and figures provide complete specifications and results.[1]

Analyses were performed in Python 3.12.14 with NumPy 2.3.5, pandas 2.2.3, SciPy 1.17.0, and scikit-learn 1.8.0; figures were produced with Matplotlib 3.10.8. Code, exact package versions, and the random seed (20260910) are provided with the data.

[1] Supplementary methods, Tables S1–S8, Figures S1–S4 and the data and code archive accompany this article.