Feeding study
Study design
The study (project nr. 69780) was overseen and approved by the Institutional Animal Welfare and Ethical Review Body and conducted under the authority of the UK Animals (Scientific Procedures) Act 1986.
The eleven dogs included in the study were selected from the colony at Waltham Petcare Science Institute following behavioural assessments and health-screening, including assessment of blood hematology, biochemistry and urinary health parameters, as well as a cardiac ultrasound scan conducted on conscious dogs by a board-certified cardiologist. Only dogs with health parameters within normal ranges and without signs of cardiac disease were considered for inclusion. One male dog was excluded from the study due to signs of cardiac changes observed during screening. A replacement could not be identified, and the Control groups was therefore limited to five dogs. All study dogs were neutered adult Labrador retrievers, three males and eight females, from a total of seven litters, aged between 2.5 to 7.7 years (mean 5.8 years; median 5.5 years). In a longitudinal study design, the dogs were randomized into one of two diet groups – Control or Test – balanced for sex, age, body weight (BW), familial affiliation (litter mates) and energy requirements. The Control group comprised five and the Test group six dogs with mean body weights of 25.4 (range 22.4-28.8) and 25.3 (range 21.5-33.9) kg and mean energy intakes of 1311 (range 925-1947) and 1105 (range 875-1698) kcal per day, respectively, at the start of the study. Due to time and resource constraints, a pre-feed period of 2.5 week period was implemented in which all dogs were fed the Control diet for the purpose of gaining baseline measurements. Subsequently, the dog groups were transitioned to the respective dry format experimental diets over a period of 5 days, and thereafter fed exclusively the respective diets for 30 days.
Housing and feeding
All dogs were housed according to diet group to minimize risk of feeding error and potential impact of coprophagia. The amount of diet provided was according to each dog’s estimated metabolizable energy requirements (MER) needed to maintain ideal BW and body condition score (BCS) according to the 9-point scale [44]. Dogs received one morning meal supplying 60% and an afternoon meal supplying 30% of MER with the remaining diet allowance of 10% as training rewards. Water was freely available to dogs at all times. Within their diet groups, the study dogs were exercised and socialized as normal and according to institutional practices.
Diets
Single batches of two extruded dry format diets were specifically formulated and manufactured for this study (Mars Petcare, Franklin, Tennessee, USA). The Control diet contained ingredients commonly used in conventional diets and did not contain pea, lentil or flaxseed products: ground wholegrain corn, poultry by-product meal, corn gluten meal, animal fat, meat and bone meal, soybean meal, ground wholegrain wheat, brewers rice, chicken by-product meal, dried beet pulp, methionine, vitamins and minerals. The Test diet, which was specifically formulated with less conventional ingredients compared to the Control diet, contained kangaroo, split peas, red lentils, green lentils, sunflower oil, flaxseed, pea fibre, methionine, vitamins and minerals. The inclusion level of the legumes split peas, red lentils, and green lentils was 20% each, with a total of 60%, on an as fed basis. Both were formulated for compliance with AAFCO [24] and NRC [25] recommended allowance for adult dogs. The nutrient composition analysis of both diets (Table 1; as-fed basis) was carried out in parallel by Eurofins (USA) and according to the Association of American Feed Control Officials guidelines [24].
Cyanogenic glycoside (linamarin, linustatin, neolinustatin) content of the diets were determined by HPLC-UV (UpScience, France).
Sampling and analyses
At baseline and days 3, 14 and 28 following transition onto the trial diets, fasted blood (8.5 mL; at least 12 h since last meal) was drawn from the jugular vein. Samples were kept on ice until processing and/or analysis was performed (within 60 min of sampling). Free-catch (at baseline and days 3, 14 and 28) and pooled 4-d urine collections before baseline and days 27-30 were also sampled. Any samples that could not be analysed immediately were stored at − 80 °C for a maximum of 2 years.
The differing days for blood, free-catch urine and pooled urine sample collections at the end of the trial were necessary for logistical reasons, and therefore reference to samples from both day 28 and 30 are made to reflect this. Heart ultrasound scans were performed to screen all dogs for trial inclusion, while due to unforeseen circumstances (cardiologist’s illness on the scheduled second visit), scans were only performed on six of the 11 dogs at the end of trial.
Standard hematology was measured in EDTA-dosed blood samples with a Mythic cell counter (Woodley Veterinary Diagnostics, Horwich, UK). Plasma was separated from lithium heparinized blood by centrifugation (1999 g for 10 min at 4 °C) and samples were analysed for standard biochemistry (Olympus AU480 Biochemistry analyser; Olympus Europe GmbH). The instruments were maintained according to the manufacturer’s schedule. Performance of each machine was verified by analysis of manufacturer’s QCs prior to and post-sampling. In addition, the AU480 is further verified by inclusion on an external quality assurance program (RIQAS, Randox).
Ionized Ca was measured in lithium heparinized whole blood within 30 min of collection by an ion-selective electrode of a blood gas-analyzer (Stat Profile Prime® Critical Care Analyser, Woodley Equipment Company LTD, Horwich, UK). Performance of the instrument was verified by analysis of the manufacturers’ tri-level quality control material before and after sample measurement.
Bone-specific ALP was analysed using the BAP MicroVue™ Quidel enzyme linked immunosorbent assay (ELISA; TECO Medical Group) according to the manufacturer’s instructions. Frozen (stored at − 80 °C) samples were sent to IDEXX Bioanalytics (Ludwigsburg, Germany) for NTproBNP analysis (ELISA).
For taurine analysis, plasma stored at − 80 °C was defrosted before analysis, frozen whole blood samples underwent two freeze thaw cycles and were diluted with deionized water to promote cell lysis, while fresh and pooled urine samples were diluted 1:5 with sample diluent buffer (5% SSA). For all sample types, sample processing buffer (5% SSA+ 500μMol/L Norleucine+ DGAA) was added 1:1, mixed using a Vortex Whirlimixer and allowed to incubate at room temperature (RT) for 20 min. Sample was then centrifuged at 7000 g, for 5 min at RT. Supernatant was removed and filtered with a Whatman Mini-Uniprep Syringeless 0.2 Filter. Analysis was carried out in duplicate on the same day as sample processing on a Biochrom 30+ Amino Acid Analyzer (Biochrom Ltd).
EHR interrogation
Banfield® Pet Hospitals are one of the largest networks of primary care veterinary hospitals within the United States with over 1000 practices across 42 states. Banfield® maintains a centralized electronic medical record system that stores information from the millions of patient visits since the mid-1990’s. Data were extracted from the anonymised EHRs of client-owned dogs visiting Banfield® Pet Hospitals between January 2, 2018, and December 31, 2019. The dataset was fully anonymised by removing any client and animal details that might be personally identifiable. Records were filtered to only include pets between age 4 and 16 years (inclusive) and with both veterinary visit and laboratory test data (hematology and biochemistry), this pairing was defined as laboratory results 90 days before or after a visit. This filtered data set contained 39,574 medical records for 32,127 unique dogs.
This data extract contained Banfield® diagnosis codes, breed, weight, gender, neuter status and age. The laboratory test data included for hematology were RBC, HCT, HGB, MCV, MCH, MCHC, RDW, WBC, lymphocyte %, monocyte %, eosinophil %, PLT, and MPV; and for clinical biochemistry total protein, albumin, globulin, ALT, ALP, creatinine, blood urea nitrogen (BUN), cholesterol, glucose, Ca, inorganic P, and TBILI. Drift and variability in certain analytes (RBC, HCT, HGB, MCV, MCH, MCHC and RDW) can be observed in hospital analysers. To mitigate this variability and minimize drift in analysers used, hospitals are instructed to use a regular programme of deep cleaning, calibration, and ongoing monitoring. However, caution should always be exercised when interpreting results involving those analytes.
Prior to case-control matching, imputation was used to infill missing data from records that contained at least one missing blood analyte value, this accounted for 0.47% of records. To do this, the IterativeImputer function from SKlearn package v0.24.2 was applied, which uses a BayesianRidge regressor. To ensure that imputed values are biologically valid, a minimum imputation threshold of 0 has been provided as a parameter. This threshold ensures that no imputed values below 0 can be returned by the algorithm.
Pets were identified as cases using the diagnoses terms “Canine Dilated Cardiomyopathy” or “Congestive Dilated Cardiomyopathy” in the Banfield® database. Records of patients with DCM were matched with healthy control pets without DCM (attending for a routine wellness visits) using propensity score matching. The Pymatch v0.3.4 (GitHub - benmiroglio/pymatch) Python library was used to create matched cases and controls using breed, gender (including neutered status; data not shown) and age (rounded to the nearest integer year). Subsequently, the case-control dataset consisted of 840 records, 420 for cases and 420 for controls. The two distributions showing the most common breeds are depicted in Fig. 3A. In Fig. 3B, a kernel density estimate (KDE) shows the distribution of ages.
Statistical analysis
Feeding study
Comparisons within all measures generated were tested at an overall significance level of 5%. Linear mixed effects models were fit to the data for all measures. Fixed effects were diet group, time point (day) and their interaction, and animal was the random effect. Log10 transformation of the response was applied if it appeared necessary from visual inspection of the residual distribution. Means with 95% confidence intervals are reported for all combinations of diet group and day (and post prandial time point where appropriate). Within each diet group, comparisons between all post baseline time points and baseline are reported with confidence intervals and p-values. The differences over time were then compared between diet groups and are also reported with confidence intervals and p-values. Comparisons for measures that were log10 transformed are reported as fold changes, whereas those without a transformation are reported as differences on the measured scale. All analyses were performed using R version 3.4.3 (2017-11-30).
EHR interrogation
Statistical differences between pets with DCM (cases) and healthy pets (controls) within the laboratory test data were identified by applying generalized additive models (GAMs) and generalized additive mixed models (GAMMs), a branch of linear modelling that can capture nonlinear patterns within data.
The step-wise process used to model each analyte, where smoothing has been achieved using thin plate spline smoothing, was performed as follows. Data were extracted for the required analyte and erroneous zero values excluded (orphaned paired data removed also). A Basic GAM model was generated using R3.6.1 [45] and the R package mgcv [46] to predict the value of each analyte using smoothed age, without random effects or the DCM term. This step was applied to investigate whether the residuals improved if the variable was log transformed (using natural log) and to also identify outlying observations (residual > 4, which were subsequently removed). A plot of the data for each analyte at this stage, together with loess smoothed trend lines for each value of the DCM flag, can be found in Fig. 2. This was plotted using the R package ggplot2 [47].
A GAMM model was fit, again using mgcv, with DCM as main effect and a smooth interaction between age and DCM Flag, and random term for the matched pair (main model). Additionally, for each analyte a secondary model was also fit, the same as main model but with Sex and an Age*Sex interaction, and the improvement of fit compared to the main model assessed. As no analyte showed a significant improvement in fit, no Sex effects were included in the final models.
For each analyte, the most significant p-value out of the individual p-values for the DCM Flag main effect and the p-values for the smoothing terms involving the DCM flag (from the main model) were taken as a conservative bound on the overall DCM p-value. Graphs were also produced for each analyte, again using ggplot2, showing the estimated difference between a positive and negative DCM flag by age, shown in Fig. 1. To evaluate whether a variable was statistically significant, a Bonferroni correction was used when interpreting p-values and constructing confidence intervals. Therefore, p-values < 0.05/25 = 0.002 were considered statistically significant, where 25 represents the number of blood analytes.