Left ventricular M‐mode prediction intervals in 7651 dogs: Population‐wide and selected breed‐specific values

Abstract Background Echocardiography is a common method to measure heart size in dogs. The heart dimensions are influenced by body weight (BW) and potentially by breed. Objectives To establish BW‐dependent prediction intervals (PIs) of the left ventricular (LV) linear dimensions in a population of dogs of many breeds in multicenter environment, and to identify breeds deviating from these intervals. Dogs Seven thousand six hundred and fifty‐one dogs. Methods Retrospectively, data from heart screens conducted between 2009 and 2016 were included. Cardiac dimensional PIs were generated using allometric scaling including all nonsighthound dogs and values were compared to previously published PIs. The values measured in dogs of respective breeds, including sighthounds, were then compared to the overall nonsighthound PIs to identify deviant breeds. The interobserver‐variability of the measurements was determined using the explained residual variance. Results Prediction intervals for the nonsighthound dogs were in agreement with previously published cardiac PIs, although the upper limits of the generated PIs of our study were slightly below those currently applied (except the interventricular septum in systole and the left ventricular free wall in diastole below 10.0 kg and 15.0 kg, respectively). Values measured in the nonsighthound breed Newfoundland deviated for most dimensions. Most of the sighthound breeds analyzed had greater cardiac dimensions, with the exception of the Irish Wolfhound. Conclusion and Importance Findings of our study reinforces the value of BW‐dependent PIs for cardiac dimensions in dogs and suggest that these PIs are valid for most nonsighthound breeds, but not the sighthound breeds.

method to measure the dimensions of the cardiac chambers and walls.
Thus, it provides important information for establishing diagnosis and assessing disease severity. 2 Cardiac dimensions are frequently measured using M-mode (MM) and two-dimensional (2D) echocardiography, where the left ventricular (LV) dimensions are measured using either method. 3 The left ventricle (LV) dimensions are either measured in the right parasternal long or short axis views. 3 When comparing LV MM values to 2D measurements, better agreement between short axis MM and 2D measurements has been found. 4 When comparing the longitudinal and short axis view, Schober and Baade found that some measurements vary depending on the measuring plane, but this systematic difference was less than 5%, except for the interventricular septum in systole. 4 When generating normal reference ranges for cardiac linear dimensions in dogs, the influence of body size on the echocardiographic measurements must be taken into account because of the wide range in the body size in dogs. 5 Body weight (BW) is frequently used as a surrogate for body size and can be used in statistical regression models with a potentially nonlinear relationship between BW and the cardiac variables of interest. Based on such regression models, reference ranges can be derived as the prediction intervals (PIs) for specific BW values.
In veterinary medicine, so far, clinically usable BW-dependent reference values for normal dogs have been established using ratiobased indexing 6 as well as allometric scaling (scaling to body mass) 5 ; the latter being a regression approach. The currently commonly used reference ranges in dogs were generated using allometric scaling. 5 Although these PIs generated are frequently used in clinical practice, our study had some limitations, the most important ones being that it only included 494 dogs, which, for this purpose, is a comparably small population, and more than 40% of the dogs were of a sighthound breed. Several studies have suggested that certain sighthound breeds have, compared to other breeds, different dimensional and functional cardiac variables. [7][8][9][10][11][12] In addition, there are many other single-centered studies reporting of breed specific reference ranges in sighthound and nonsighthound breeds.  Because the studies were single-centered and because the findings were not compared to those of other breeds or a general dog population, it is not known if these specific ranges are truly deviant and if they are applicable in the multicenter environment, that is, have generalizability.
The aim of our study was to generate BW-dependent PIs of the LV linear dimensions in a large population of dogs of many breeds in multicenter environment, and to identify breeds deviating from these intervals.

| Dogs
Data of healthy dogs of different breeds (Table S1)  Owner related data, signalment, auscultatory, ECG, and echocardiographic variables were retrieved.

| Grouping of dogs
Dogs were grouped into 3 groups: All nonsighthound dogs, immaculate nonsighthound dogs, and breeds including sighthound dogs.

| All nonsighthound dogs
According to the classification system of the CC, dogs with no or minimal cardiovascular abnormalities are to be considered healthy. Dogs

| Immaculate nonsighthound dogs
A subgroup of the all nonsighthound dog group consisted of dogs with no remarks on the echocardiogram. Dogs in this group had to meet the following requirements: Dogs must have been free of signs of disease and the physical examination did not reveal important or relevant abnormalities.

| Breeds including sighthound dogs
This group included dogs of sighthound breeds and dogs of nonsighthound breeds with >80 dogs. The dogs were considered to be healthy according to the classification system of the CC, as they, like the dogs in group all nonsighthound dogs, exhibited no or only minimal cardiovascular changes. Thus, they had to be free of clinical signs of disease and none or, at the most, mild insufficiencies of the atrioventricular or the semilunar valves were found on the color-Doppler echocardiogram. Additionally, the anatomy of the ventricular outflow tracts had to be normal and the maximal aortic and pulmonic flow velocity had to be within physiologic limits, respectively.

| Echocardiography
The echocardiographic examination was performed using a suitable ultrasound device, which provided the examiner with various Doppler techniques (spectral-and color-Doppler) as well as simultaneous ECG recording. However, only the MM measurements were available for analysis within the scope of our study. The LV dimensions were measured either in the right parasternal long or short axis view. 36 View of acquisition was noted in the report. The left ventricular septal and lateral wall thicknesses (interventricular septum in diastole and systole [IVSd and IVSs], respectively), left ventricular free wall in diastole and systole (LVWd and LVWs, respectively) and the left ventricular diameter were measured in systole and diastole (LVDd and LVDs, respectively). The diastolic measurements were performed at the beginning of the QRS complex; the systolic measurement was timed at the shortest distance between septum and lateral wall. The measurements were performed using the leading-edge technique. 36 The variable fractional shortening (FS) 37 was then calculated automatically.. 36 The investigators were instructed to use breed-specific values, if available.
If no specific values had been published for the respective breed, the investigators were encouraged to interpret the results at one's own discretion, excluding all kinds of possible pathologic changes. Therefore, the PIs of Cornell et al 5 were used as an aid to interpret results, but not as a definite criterion, because this would not take potential breed-specific differences into account.

| Statistical analysis
Groups analyzed separately included all nonsighthound dogs and immaculate nonsighthound dogs. The reference ranges were derived as PIs, which were estimated on the basis of statistical regression models. In the general outlier detection analysis, all variables were checked for possible outliers. In accordance with standard statistical practice, values outside the interval "upper/lower quartile ± triple interquartile range" were identified as outliers 38 and then thoroughly inspected and excluded if necessary.
This evaluation was performed separately for each breed where at least 50 dogs were available. Variables with skewed distributions were logarithmized in a preprocessing step to ensure a more symmetric distribution. In total, 197 measurements were removed from the analyses.
A simple linear regression model was estimated for each transformed variable, with the logarithmically transformed BW as the sole predictor. 39 The estimated coefficients of the model, namely the model intercept and the weight effect, then constituted the variables a and b of the corresponding allometric scaling model, respectively.
The appropriateness of these allometric scaling models was evaluated by estimating an additive regression model (GAM) for each variable, including a smooth, potentially nonlinear effect of BW to the power of the previously obtained variable b. 39 Visual inspection of the linearity of the fitted prediction line was used to assess the adequacy of the allometric scaling model. This way we assured that the linear effect structure was a reasonable assumption for the observed relationship.
Model estimation was performed based on the 2 data sets, the all nonsighthound dogs and the immaculate nonsighthound dogs. To compare these groups, the average difference between all nonsighthound dogs and immaculate nonsighthound dogs in their upper limits of the reference values across all variables were estimated. In order to do this, a GAM was used, including a dummy variable for the group and a nonlinear effect of BW. 39 The latter was estimated based on a P-spline basis, the model was fitted using the function gam from the package mgcv 39 in the statistical software R. 40 The PIs were generated in 2 differing ways: For variables that showed only a very weak association with BW, the values are pres- To determine whether certain breeds were within the generated PIs or not (deviant breeds), a cut-off value of 10.0% above or below those PIs generated in our study was used. This means that a breed was identified as deviant breed if more than 10.0% of the measurements of dogs of this breed were above or below the corresponding PI.
Interobserver-variability was analyzed using an additive mixed model for each variable, with the examiner effect as random intercept and a nonlinear effect of BW. The model was estimated only based on the breed of Boxers (from the group breeds including sighthound dogs), since they constitute the largest group within the sample and to prevent biased results caused by comparing different breeds. The proportion of the error variance (residual variance [RVa]) explained by the examiner effect was then used as measure for interobservervariability. 41 The analyses were performed using the statistical software R 3.5.0. 40

| Comparison of groups and generation of weight-dependent 95% PIs
We observed a strong association between BW and all dimensional cardiac variables, which were accurately described by allometric scal-  Figure S1). Furthermore, no relevant differences between the groups could be detected when the mean differences between the upper  limits of the all nonsighthound dogs and immaculate nonsighthound dogs were inspected (Table S3). Hence, further analyses were conducted using the group all nonsighthound dogs only. Consequently, constants for the indexation of the echocardiographic measurements (Table 2) as well as 95% PIs (Table 3) were presented for each variable using respective group.
F I G U R E 1 Scatter plots with superimposed regression lines (solid lines) and 95% prediction intervals (broken lines) by breed for the LVDd as a function of body weight for breeds in the nonsighthound breed group (A) and for breeds in the sighthound group (B). The plots show how observed values fit inside the 95% PIs. LVDd, left ventricular diameter in diastole; PI, prediction interval

| Analysis of breeds including sighthound dogs
The results of the analyses regarding the breed distribution (percentage above and below the respective PIs) are summarized in Table 4 Owczarek Nizinny had less than 10% of the observations above or below the PIs generated. In each of the breeds, Golden Retriever and Hovawart 1 variable could be detected where more than 10.0% of the measurements were above or below the corresponding PI; in the Golden Retriever, 14.1% of the measurements were above the PI of LVWs and in the Hovawart, 10.4% of the measurements were above the PI of LVDs.
In contrast, in Newfoundlands, more than 10.0% of the measurements were below the generated PIs for several variables. Figure 1A shows the results of the analysis for the LVDd dimension of nonsighthounds.
Regarding the sighthound breeds, the following could be In the Borsoi breed, however, only for LVDs a high percentage of observations (11.1%) was found to be above the upper limits of the PIs. In the cases of the Italian Greyhound and the Silken Windsprite, all measurements were less than 10% above or below the PIs.
In contrast, for Irish Wolfhound, none of the variables had a high percentage (>10.0%) of measurements above the PIs, and only 1 variable had a high percentage of observations below the PI (IVSs, 12.0%). Figure 1B shows the results of the analysis for the LVDd dimension of nonsighthounds in graphical form.  shown that the body size or the BW have a clinically relevant impact on most dimensional echocardiographic variables. 5,7-9,18,20,25,26,33,44-48 5,7-9,18,20,25,26,33,44-48 However, most of these studies were single centered. Thus, when transferring the results of these studies to the general dog population, their applicability might be limited. In contrast, our study represents a multicentered environment, which consequently leads to a high generalizability.

| Interobserver-variability
Furthermore, not only BW, but also the somatotype (eg, breeds of different sizes but uniform athletic physiques, like sighthounds) and breed have to be considered when interpreting echocardiograms. Sighthounds are known to exhibit several peculiarities in the dimensions and functional variables of the heart. For example, Greyhounds have, compared to other dogs, both larger LV dimensions and increased wall thickness of the IVS and LVW. 7,[9][10][11] Comparable results have been reported in Whippets, 7,8,12 as well as in Deerhounds. 22 The exact reason for these findings has not yet conclusively been determined, but selection for athletic 7 capacity and increased blood viscosity 11 have been suggested as possible causes.
Thus, to take these characteristics of sighthounds into account, we removed the sighthounds from the study population (all nonsighthound dogs) to generate the BW depended PIs.
In our study, we found a strong association between the logarithmically transformed echocardiographic measurement and BW for all cardiac dimensions. Association was the strongest for LV chamber dimensions in diastole and systole (R 2 = 0.766 and 0.680, respectively).
If allometric scaling is used to normalize cardiac dimensions to BW, volumes should theoretically linearly relate to BW, cross-sectional areas should be proportional to BW 2/3 , and linear dimensions should linearly relate to BW 1/3 . 5 As cardiac chamber diameters and wall thicknesses are 1-dimensional measured variables (straight lines), the scaling exponent of the allometric formulas should theoretically approximate onethird. In our study, the exponents of all cardiac dimensions approximated this theoretical value of one-third and ranged between 0.247 and 0.346 (Table 2). Furthermore, it was found that the cardiac wall thicknesses had lower values for the scaling exponents (constant b) compared to the other 2 measurements-an observation which is consistent with the results of the publication of Cornell et al. 5 When comparing the generated PIs of our study with those currently provided by Cornell et al, both differences and similarities were found. 5 The estimates for allometric constants and PIs for cardiac wall thicknesses were very similar between the 2 studies, but notable differences could be observed for the LV dimensions. These differences were also evident in The results of our study also have an impact on PIs in smaller dogs. For example, in the "EPIC study", 49  An advantage of our study was that the validity of the calculated PIs was confirmed by validating the generated values via the GAM.
The GAM is a nonparametric model, which can be used to validate other statistical models (such as allometric models) for accuracy. In contrast to "model-driven" methods, which only allow for a response curve limited by the underlying statistical method, the GAM is "datadriven" and actually represents the functional response curve of the data. 39 The degree of fitness for other models can therefore be vali- With regard to the breed distribution, the following could be seen: For most breeds, the measurements of the nonsighthounds were largely within the generated PIs, with the exception of Newfoundlands, where a significant number of measurements were below the PIs. In the case of the sighthounds, the measurements were above the generated PIs for most breeds, especially in the LV chamber dimensions. As an exception to this, the Irish Wolfhound could be detected, where the measurements were in higher percentages below the generated PIs. Concerning the sighthound breeds, our study showed that in the Afghan, the Saluki, the Whippet, and the other sighthound breeds, the measurements of several dimensions-especially the LV chamber dimensions-were clearly above the cut-off limit of 10.0%. Therefore, as generally assumed for sighthound breeds, higher values of the LV dimensions can be presumed for these breeds.
With the Irish Wolfhound, which had a large number of measurements (n = 837), an interesting finding was observed. In all dimensions, at least 7.4% (and up to 12.0%) of the measurements were below the generated PIs. This suggests that the Irish Wolfhound is different from other sighthound breeds, such as the Afghan or the Saluki.
Therefore, breed specific PIs 30 are indicated in this breed as well.
In contrast, the situation is less clear for the Borsoi, the Italian Greyhound, and the Silken Windsprite. Although for the Borsoi, 11.1% of the measurements were above the PI for LVDs, also 11.1% of the measurements for LWVs were below the respective PI. Furthermore, for the Italian Greyhound and the Silken Windsprite, all measurements were even within the generated PIs. This could be due to the low numbers of animals in these breeds (Borsoi, n = 9; Italian Greyhound, n = 13; Silken Windsprite, n = 25); on the other hand, however, it must also be considered that the assumption that the cardinal dimensions of all sighthound breeds are above the general dog population cannot be applied in principle. Therefore, these 3 breeds in particular should be investigated more intensively in follow-up studies with larger study populations.

| LIMITATIONS
Most of the limitations of this article relate to its retrospective nature.
The data used were generated over a period of 7 years by many investigators with different experience and by means of different types of ultrasound equipment. Especially due to these heterogeneous individual and technical aspects, a certain degree of variability cannot be avoided. However, it must also be stressed that this heterogeneity also leads to a higher generalizability, as all causes of variability were included, and this is a clear advantage of our study.
In addition, the standardized examination protocol of the CCs contained a certain degree of freedom of decision for the examiner.
For example, the LV measurements could be performed in either the longitudinal or in the short axis, 36  However, breed specific PIs might also be indicated in the Great Dane and the Hovawart. With regard to the sighthounds, it can be concluded that specific PIs are indicated in these breeds as most of the sighthound breeds deviated strongly from the generated PIs.