Abstract
Spontaneous hypercortisolism (HC) is a common endocrine disease of senior dogs, often overlapping in selected clinical signs and hematologic and blood biochemical abnormalities with nonadrenal diseases (NADs). HC and NAD could differentially affect cortisol metabolism, which is a complex 10-enzymatic pathway process. HC might also affect blood and urine lactate levels through its effects on mitochondrial function. We aimed to differentiate between HC and NAD via a urinary cortisol metabolites and lactate panel. We prospectively recruited 7 healthy dogs and 18 dogs with HC, 15 with congestive heart failure (CHF), and 9 with NAD. We analyzed urine by gas chromatography–mass spectrometry and liquid chromatography–mass spectrometry. We normalized urinary lactate and cortisol metabolites to urine creatinine concentration, and then compared groups using a linear-mixed model and principal component (PC) analysis. A machine-learning classification algorithm generated a decision tree (DT) model for predicting HC. The least-squares means of normalized urinary 6β-hydroxycortisol and PC1 of the HC and CHF groups were higher than those of the healthy and NAD groups (p = 0.05). Creatinine-normalized urinary 6β-hydroxycortisol had better sensitivity (Se, 0.78; 95% CI: 0.55–0.91), specificity (Sp, 0.89; 95% CI: 0.57–0.99), and a likelihood ratio (LR; 7), than the Se (0.72; 95% CI: 0.49–0.88), Sp (0.89; 95% CI: 0.57–0.99), and LR (6.5) of PC1 for distinguishing HC from NAD. Lactate and dihydrocortisone had the highest decreasing node-weighted impurity value and were considered the most important features in the DT model; dihydrocortisol had no role in determining whether a dog had HC.
Spontaneous hypercortisolism (HC; Cushing disease) is a common endocrine disease of senior dogs. 25 Early diagnosis of HC can sometimes be obfuscated by overlapping clinical signs of nonadrenal diseases (NADs). 3 Systemic stress from NAD illness can induce hypothalamic-pituitary-adrenal axis (HPAA) responses that result in increased concentrations of cortisol and its metabolites.7,10,26 Activation of the HPAA in NAD is especially important because the endocrine tests used in the diagnosis of HC in dogs (i.e., the low-dose dexamethasone suppression test [LDDST] and the adrenocorticotropic hormone [ACTH] stimulation test) are premised on cortisol concentrations as a diagnostic endpoint.3,9 Consequently, to date, none of the tests available for HC reliably differentiates between NAD and HC when significant activation of the HPAA is present.
Cortisol metabolism is a complex 10-enzymatic pathway process that results in 10 corticoid metabolites (Fig. 1). 16 If NAD directly interferes with regulation of one or more cortisol metabolic pathway, it has the potential to alter the ratio between cortisol and its metabolites without an overall change in cortisol concentrations. For example, thyroid and growth hormones alter 11β-hydroxysteroid dehydrogenase 1 activity (responsible for the conversion of cortisone to cortisol) and steroid 5α/5β-reductase activity (responsible for C-19 and C-21 steroid biosynthesis). 16 Metabolic signatures unique to NAD could aid in distinguishing between NAD and HC. For example, 6β-hydroxycortisol concentrations increase in the urine of people with HC.20,23 Similarly, urinary lactate could be increased in HC as a result of disruption of mitochondrial function and lactate recycling by high concentrations of cortisol.4,12,15,19 Machine-learning (ML) algorithms in medical sciences enable identification of associations within and between arrays of biomarkers and disease output attributes, which allows identification of unique metabolic signatures for a variety of diseases.
Cortisol metabolic pathways and metabolites. The liver is the primary site of cortisol metabolism, where cortisol is reduced or oxidized and hydroxylated. The hepatic cortisol degradative reaction products become water-soluble through conjugation with sulfate or glucuronic acid to facilitate urinary excretion. The liver inactivates cortisol by disrupting the 3-keto, delta-4 double-bond structure, and the reduction of the delta-4 double bond is the rate-limiting step in cortisol metabolism. Two distinct enzymatic processes facilitate this: 1) 3α-hydroxysteroid dehydrogenase reduces the keto group leading to formation of a 3-hydroxyl group; 2) 5α-reductases (2 liver isoforms) and 5β-reductase reduce the delta-4 double bond. The kidney is the primary site of extrahepatic cortisol metabolism, where the high activity of renal hydroxysteroid 11β-dehydrogenase-2 protects the mineral corticosteroid receptor from cortisol. Adapted from Newell-Price JDC, Auchus RJ. 16 11β-HSD1, 11β-HSD2 = 11β-hydroxysteroid dehydrogenase 1 and 2, respectively.
We previously identified differences in the urinary corticoid metabolite tetrahydrocortisol between healthy dogs and dogs with HC. 18 The goals of our current study were to ascertain if additional differences in cortisol’s urinary metabolites 6β-hydroxycortisol, 6β-hydroxycortisone, 20β-dihydrocortisol, 20β-dihydrocortisone, and urinary lactate are present between dogs with HC and NAD. We hypothesized that the analysis of a panel of urinary cortisol metabolites and lactate would differentiate dogs with HC from dogs with NAD.
Materials and methods
Study overview
From June 2019 through November 2020, we prospectively enrolled 49 client-owned dogs into 4 groups: HC, congestive heart failure (CHF), NAD, and apparently healthy dogs (Table 1). The University of Illinois’ Institutional Animal Care and Use Committee approved our study (protocol 19056). Informed consent was obtained from all dog owners prior to inclusion in the study.
Demographic information of dogs included in our study of urinary lactate and cortisol metabolites.
AST = ACTH stimulation test; CHF = congestive heart failure; CKD = chronic renal disease; CM = castrated male; DM = diabetes mellitus; HC = hypercortisolism; LDDST = low-dose dexamethasone suppression test; MBD = mixed-breed dog; n = valid number of observations; NAD = nonadrenal disease; SF = spayed female.
Dogs with acute pulmonary edema as a result of congestive heart failure.
Dogs were untreated at the time of inclusion.
No significant differences in bodyweight between groups (p = 0.31); different superscript letters indicate a statistically significant difference between groups (p ≤ 0.05).
The HC dog group consisted of 18 dogs with clinical signs, as well as hematologic and biochemical results, compatible with HC, confirmed by either an ACTH stimulation test or LDDST. Localization as either pituitary dependent or adrenal dependent was then determined based on the results of one or more of the following tests: LDDST (suppressed 4-h serum cortisol concentration, or decrease in 4- or 8-h serum cortisol concentration by >50% compared to baseline in cases of pituitary-dependent HC [PDH]), plasma endogenous ACTH concentration (increased in PDH cases and decreased in the case of a cortisol-secreting adrenal tumor [AT]), and diagnostic imaging studies (i.e., symmetric adrenals supporting PDH vs. one large and one small adrenal, supporting a diagnosis of AT). None of the dogs had received previous treatment for HC. Seventeen of the 18 dogs had a confirmed or tentative diagnosis of PDH, and one dog had a tentative diagnosis of AT.
The CHF group consisted of 15 dogs with CHF. These dogs had been presented for acute left-sided CHF, secondary to dilated cardiomyopathy or chronic valvular disease, which required intensive care and stabilization. All patients had a transthoracic echocardiogram performed by a board-certified cardiologist; thoracic radiographs were used to confirm cardiogenic pulmonary edema based on a perihilar, diffuse, or focal interstitial or alveolar lung pattern. We chose to include these dogs because this group is readily defined in a clinical setting and also because we and others had shown that structural heart disease resulting in heart failure leads to marked activation of the HPAA (i.e., a positive control group).18,21
The NAD group consisted of 9 dogs that had been presented to the Small Animal Internal Medicine Service at the University of Illinois Veterinary Teaching Hospital (VTH; Urbana, IL, USA). HC was one of the main differential diagnoses in this group and the reason for referral (Table 1).
The healthy dog group consisted of 7 dogs. These dogs had unremarkable histories, clinical examinations, and CBC, biochemistry, and urinalysis results.
To our knowledge, in the 14 d prior to enrollment in our study, none of the 49 dogs had received treatment with topical or systemic glucocorticosteroids. After enrollment, a single urine sample (either a voided sample or via cystocentesis) was collected from each dog after an overnight fast, placed into a sterile container, and stored at −80°C until analysis. The urinary cortisol metabolites 6β-hydroxycortisol, 6β-hydroxycortisone, 20β-dihydrocortisol, and 20β-dihydrocortisone were measured by liquid chromatography–mass spectrometry (LC-MS). The concentration of urinary lactate was measured by gas chromatography–mass spectrometry (GC-MS). Urinary creatinine concentrations were measured by a kinetic colorimetric assay based on the Jaffé method (AU680 analyzer; Beckman Coulter). The cortisol metabolite standards 6β-hydroxycortisol (CAS 53-35-0), 6β-hydroxycortisone (CAS 16355-28-5), 20β-dihydrocortisol (CAS 116-58-5), and 20β-dihydrocortisone (CAS 116-59-6) were purchased from Steraloids.
Liquid chromatography–mass spectrometry
For the analysis of urinary cortisol metabolites, the urine sample was mixed with the internal standard d3-testosterone and β-glucuronidase (Roche), followed by solid-phase extraction cleanup (Strata-X 33-μm polymeric reverse-phase cartridges, Phenomenex; 30 mg/3 mL). The collected eluate was subject to complete dryness with SpeedVac (Thermo Fisher) and reconstituted into a 60% methanol solution before instrument injection. Samples were analyzed (5500 Qtrap LC-MS/MS system; Sciex) in the Metabolomics Laboratory of the Roy J. Carver Biotechnology Center, University of Illinois at Urbana-Champaign. Analyst v.1.7.1 (Sciex) was used for data acquisition and analysis. The 1200 series HPLC system (Agilent) includes a degasser, an autosampler, and a binary pump. The LC separation was performed on a Gemini C6-phenyl column (2 × 100 mm, 3 μm; Phenomenex) with mobile phase A (0.1 % formic acid in water) and mobile phase B (0.1% formic acid in acetonitrile). Mass spectra were acquired under electrospray ionization with the ion spray voltage of +5,000 V. Multiple reaction monitoring (MRM) was used for quantification with the transitions: 6β-hydroxycortisol m/z 379.2 → m/z 343.1; 6β-hydroxycortisone m/z 377.2 → m/z 359.2; 20β-dihydrocortisol m/z 365.2 → m/z 269.1; 20β-dihydrocortisone m/z 363.2 → m/z 163.0; internal standard d3-testosterone m/z 292.1 → m/z 256.1.
Gas chromatography–mass spectrometry
For the analysis of urinary lactate, 30 μL of urine sample was evaporated under vacuum and derivatized with 100 μL of MTBSTFA (CovaChem) and 20 μL of pyridine before instrument injection. Lactate profiles were acquired using a GC-MS system (Agilent) consisting of a 7890B gas chromatograph, 5977A MSD, and a 7693 autosampler. MassHunter WorkStation software v.B.08.00 (Agilent) was used for data acquisition and analysis. GC separation was performed on a HP-5MS capillary column (60 m × 0.32 mm I.D., 0.25 µm film thickness; Agilent). The mass spectrometer was operated in positive electron impact SIM mode at 69.9 eV ionization energy. The m/z 261 was used for lactate quantification.
Normalization to urinary creatinine
The total (i.e., conjugated + free) individual urinary cortisol metabolites and lactate were normalized by dividing their respective concentrations by the urinary concentration of creatinine to account for differences in urine concentration between dogs.
Power sample size calculation
An a priori power sample size calculation (G*Power v.3.1.9.4, https://www.psychologie.hhu.de/arbeitsgruppen/allgemeine-psychologie-und-arbeitspsychologie/gpower.html) was based on the primary outcome variable 6β-hydroxycortisol, according to a study in people. 23 The calculation assumed that the sizes of the 4 groups would be balanced to detect a difference between 6β-hydroxycortisol in HC (x̄ ± SD; 754 ± 888 ng/mL) and NAD (79 ± 43 ng/mL) groups with an α error probability of 0.05, power of 0.8, and effect size of 1.07. The required sample size was 15 dogs per group.
Statistical analysis
All data analyses were performed with SAS University Edition (SAS Institute). The data were examined for normal distribution by inspection of Q-Q plots, histogram, and by the Shapiro–Wilk test. Non-normally distributed data were natural log transformed. The descriptive statistics of normally distributed data were described by x̄ ± SD; non-normally distributed data was described by median (min., max.). Comparison of bodyweights between groups was performed with the nonparametric Kruskal–Wallis test. ANOVAs of log-transformed urinary lactate:creatinine ratio and log-transformed urinary cortisol metabolites:creatinine ratio was performed with the MIXED procedure using a linear-mixed model. The model included the fixed effect of group. Age, sex, bodyweight, and time-of-storage, and the interactions of group with age, sex, bodyweight, and time-of-storage had a nonsignificant effect on the dependent variables and were not included in the final model. The least-squares means (LSM) and SEs of each of the dependent variables for each dog group were obtained and used for comparisons between groups using the Fisher least significant difference (LSD) test and the Tukey method for post hoc pairwise comparisons as implemented in the LSMEANS option of the MIXED procedure.
Principal component (PC) analysis as implemented in the PRINCOMP procedure was used to assess for a clustering effect within the samples. Only PCs with an eigenvalue >1 were considered for inclusion in downstream analyses. ANOVAs of selected PCs were performed with the GLM procedure using a linear model. The model included the fixed effect of group, and the model’s LSM was used to compare groups. The significance level was set at p = 0.05. The data were analyzed before and after excluding the dog with an AT. Given that there were no significant differences in the results, the final analyses included all of the dogs in the HC group.
Receiver operating characteristic curve (ROC) analysis was computed in Prism 9 (GraphPad) to calculate the sensitivity (Se), specificity (Sp), and likelihood ratio (LR) of selected urinary cortisol metabolites and PCs. Cutoff values were determined from the Se and Sp that gave the highest LR. The significance level was set at p = 0.05.
Machine-learning analysis
We used decision trees (DT), a supervised-learning classification machine-learning (ML) algorithm 13 to differentiate between the HC group and the other groups. A classification algorithm determines the relation between input attributes and the output attribute to construct a model. 27 All dogs in the study were included in the training set that consisted of 5 features (i.e., 6β-hydroxycortisol, 6β-hydroxycortisone, 20β-dihydrocortisol, 20β-dihydrocortisone, and lactate; all normalized to urine creatinine concentration) and a binary outcome (yes = “HC”, no = “other”). We repeated this process by adding additional features: age, sex, time-of-storage, and bodyweight one at a time and all 4 together. The resulting model did not change, affirming that those features did not have a significant effect on the model.
To partially compensate for the small number of data points and to overcome overfitting, we used an “extreme” leave-one-out cross-validation (LOOCV) technique. Each learning set was created by taking all of the samples but one; the test set was the sample left out. Thus, we ended with 49 different training and test sets. By averaging performance seen over the 49 models, we selected the best model and the best feature set.
The Gini index of a DT node is the probability of a particular variable being classified wrongly when it is chosen randomly. The Gini index (Gini impurity) for each node was calculated by the following formula:
where pi is the probability of an object being classified to a particular class. A “feature importance” is measured across the DT nodes involving a particular feature by the decrease in node impurity weighted by the probability of reaching that node. It was calculated by the following formula: Feature importance = N_t/N × (Gini_Index – N_t_R/N_t × right_Gini_index – N_t_L/N_t × left_Gini_Index), where N is the total number of samples, N_t is the number of samples at the parent node, N_t_L is the number of samples in the left child node, and N_t_R is the number of samples in the right child node. The ML model was built using the Scikit-learn library for the Python programming language. 17
Results
Creatinine-normalized log-transformed urinary 6β-hydroxycortisol significantly differed in dogs in the HC and CHF groups compared to dogs in the NAD and healthy groups (p = 0.05; Table 2, Fig. 2A). The creatinine-normalized log-transformed urinary cortisol metabolites 6β-hydroxycortisone, 20β-dihydrocortisol, and 20β-dihydrocortisone, as well as the creatinine-normalized log-transformed urinary lactate, differed in dogs from the HC and CHF groups compared to the healthy group (p = 0.05) but not from dogs from the NAD group (p = 0.05; Table 2, Fig. 2). Neither the creatinine-normalized log-transformed urinary cortisol metabolites nor creatinine-normalized log-transformed urinary lactate differed between dogs in the HC and CHF groups (p = 0.05; Table 2, Fig. 2).
LSM ± SE creatinine-normalized log-transformed urinary lactate and total cortisol metabolites of healthy dogs, and dogs with CHF, NAD, and HC.
Different superscript letters indicate a statistically significant difference between groups within the metabolites (p ≤ 0.05).
CHF = congestive heart failure; HC = hypercortisolism; LSM = least-squares mean; NAD = nonadrenal disease.
Boxplot of the ratio of
Principal component 1 (PC1) had an eigenvalue of 3.43 and accounted for 69% of the variance between samples. PC1 LSM (± SE) in the HC (1.06 ± 0.33) and CHF (0.33 ± 0.36) groups differed significantly from those in the healthy (−2.69 ± 0.53) and NAD (−0.58 ± 0.47) groups (p = 0.05); however, some overlap existed between dogs in the HC and NAD groups (Fig. 3). The area under the curves (AUC) for the ROC of PC1 (Fig. 4A) and 6β-hydroxycortisol (Fig. 4B) were 0.8 and 0.85, respectively. For distinguishing between HC and NAD, the PC1 cutoff value of 0.25 had a Se of 0.72 (95% CI: 0.49–0.88), Sp of 0.89 (95% CI: 0.57–0.99), and a LR of 6.5. The creatinine-normalized urinary 6β-hydroxycortisol cutoff value of 0.051 had a Se of 0.78 (95% CI: 0.55–0.91), Sp of 0.89 (95% CI: 0.57–0.99), and a LR of 7.
Scatter plot of eigenvalues of principal component 1 (PC1) in healthy dogs and dogs with nonadrenal disease (NAD), congestive heart failure (CHF), or hypercortisolism (HC). The horizontal line is the zero line of PC1.
Receiver operating characteristic curves of 18 dogs with hypercortisolism and 9 dogs with nonadrenal disease.
Using the ML DT with creatinine-normalized urinary dihydrocortisone at the root node, 11 of 18 of dogs with HC were classified after 2 decision nodes (i.e., testing 2 conditions), 5 more dogs with HC were classified after 1 additional decision node, and the last 2 dogs with HC were classified after going farther down the DT (Fig. 5). Lactate and dihydrocortisone had the highest decreasing node weighted impurity value and thus were considered the most important features in the DT; dihydrocortisol had no role in determining whether or not a dog had HC.
Decision tree model for the prediction of dogs with hypercortisolism (HC) generated by a machine-learning classification algorithm using a training set of 49 dogs. Each node presents a statement about the value of a creatinine-normalized urinary lactate or creatinine-normalized urinary cortisol metabolite being less than a specific value. For example, the top root node assesses dihydrocortisone, and any value ≤0.055 will result in a true response, whereas a value >0.055 will result in a false response. A true response results in proceeding to the next node on the left, whereas a false statement results in proceeding to the next node on the right. The color of the node and the class represent the more likely diagnosis at that node (blue = Y = yes; orange = N = no). End-leaf nodes are nodes with a certain outcome, either the diagnosis or not of HC. The Gini number is the probability of a correct classification at that node; a Gini number of 0.5 implies a 50% chance of a certain classification (i.e., HC or not); a value of 0 means 100% certainty of the node’s disease classification.
Discussion
We were able to differentiate dogs with HC from dogs with NAD by analysis of selected urinary cortisol metabolites and urinary lactate. As a standalone result, urinary 6β-hydroxycortisol differentiated the HC from NAD groups with a Se of 0.78 (95% CI 0.55–0.91) and a Sp of 0.89 (95% CI 0.57–0.99). These values are similar to those reported previously for the LDDST (Se 85–100%; Sp 44–73%) and the ACTH test (Se 57–95%; Sp 59–93%). 3 However, we think that better differentiation and discrimination between groups of dogs with HC and NAD could be achieved in the future if a large cohort of dogs with HC and NAD could be assessed by ML algorithms using a panel of urinary cortisol metabolites combined with urinary lactate.
The advantages of diagnosing HC based on a single urine sample include time saving, eliminating stress from patient travel to the clinic, and increased client convenience. ML-based analysis of urinary cortisol metabolites and lactate in one panel also has the potential to be more accurate than current standards of practice because of disease-specific feedback loops that influence cortisol metabolizing pathways in HC and NAD, and mitochondrial dysfunction in cases of HC, which results in increased lactate levels. The diagnostic concept of a metabolomic approach used in our study has already been introduced in human medicine 14 ; assessment of a predetermined set of related biomarkers instead of the analysis of a single biomarker has the potential to be more accurate in the diagnosis of diseases.
As in our study, 6β-hydroxycortisol has been found to be associated with states of HC in people. In one study, cortisol and 6β-hydroxycortisol were measured in vitro in an isolated human adrenal-cell system obtained from 3 normal adrenals (healthy control), 4 non-hyperfunctioning ATs, 2 functional ATs, and 5 aldosterone-producing adenomas. In the basal state, 6β-hydroxycortisol secretion, expressed as a percent of cortisol secretion, was 0.5–2% in normal adrenal cells, 1–7% in cells from non-hyperfunctioning ATs, 2.6–3.9% in cells from aldosterone-producing adenomas, and 12–15% in cells from functional ATs. The de novo production of cortisol and 6β-hydroxycortisol after stimulation with ACTH increased in all groups in a dose-dependent manner, indicating that human adrenal 6β-hydroxylase is under ACTH regulation. The higher 6β-hydroxycortisol:cortisol ratios from non-hyperfunctioning and functioning ATs indicated that adrenocortical cells in AT upregulate 6β-hydroxylase activity. 20 Similarly, when baseline plasma 6β-hydroxycortisol and cortisol levels were compared between healthy people (controls), patients with non-hyperfunctioning ATs, and patients with a hyperfunctioning AT, 6β-hydroxycortisol levels were significantly higher in the hyperfunctioning AT group than the non-hyperfunctional AT group; the latter had significantly higher levels of 6β-hydroxycortisol than healthy controls. Following LDDST and stimulation with ACTH, there were different ratios of plasma 6β-hydroxycortisol:cortisol between these groups, suggesting that there are important variations in the induction of 6β-hydroxylase in HC compared to the other groups. 23 These studies support the idea that the induction of 6β-hydroxylase would differ between dogs with adrenal-dependent HC and dogs with nonfunctional AT and could aid in differentiation of the 2 conditions.
Across species, hepatic cytochrome P4503A (CYP3A) 6β-hydroxylase activity is the major source of 6β-hydroxycortisol. 6β-hydroxylase is primarily active in the human liver, 1 and under normal health conditions, 6β-hydroxylase has minimal activity in the adrenal gland. 20 Although the intra-individual levels of urinary 6β-hydroxycortisol in healthy humans are fairly stable, with a correlation coefficient of 0.79 for repeatedly measured samples taken on the same individuals over an 8-wk period, 8 variation is wide in urinary 6β-hydroxycortisol levels between healthy people given a wide inter-individual variation in hepatic 6β-hydroxylase activity. 8 In our study, we also found that the largest SEs for log-urinary cortisol metabolite:creatinine ratios were present in the healthy dog group. Interestingly, the smallest SEs of log-urinary cortisol metabolite:creatinine ratios were present in the HC dog group, which could imply that the large inter-individual variability of hepatic 6β-hydroxylase isoform activity in canine health is cancelled by large amounts of 6β-hydroxycortisol produced by the adrenal isoform of 6β-hydroxylase in dogs with HC. Given that all but one of the dogs with HC in our study had pituitary-dependent HC, we speculate that the high levels of cortisol in dogs with HC could induce compensatory increased hepatic and adrenal 6β-hydroxylase activities in an attempt to lower cortisol levels. However, it is important to keep in mind that previous studies indicated that there is substantial inter-species variability in cortisol-metabolizing enzymes, and this assumption needs to be tested in dogs with HC.2,22 Additionally, CYP3A that encodes 6β-hydroxylase has been shown to be inducible by xenobiotics, such as phenobarbital in Cebus albifrons monkeys and humans.5,6 Therefore, the effects of exposure to xenobiotics in canine HC and NAD would need to be determined in the development of a urinary-based 6β-hydroxycortisol assay for canine HC.
Preliminary results of nuclear magnetic resonance with subsequent PC analysis on selected urine samples 18 indicated that urinary lactate discriminated canine patients with HC from healthy controls (unpublished data) and was the rationale for ascertaining if lactate had a similar effect in our study and if it could discriminate between HC and NAD. Pathomechanistically, there is some evidence to suggest that mitochondrial cellular respiration is disrupted in states of HC, thus altering the ability of mitochondria to recycle lactate, and potentially increasing lactate levels in the blood and consequently in the urine.4,11,12,15,19,24 The results from our ML analyses also suggest that urinary lactate has an important discriminatory role in the diagnosis of HC in dogs. Future studies with a large sample size will be needed to corroborate this assumption.
Our study has a few limitations that warrant consideration. First, within the funding period, we were unable to enroll the number of dogs that we intended to recruit according to our a priori power sample size analysis. Like many other studies that employ a number of test subjects smaller than planned, the probability for a type-II statistical error increases, which implies that some of our nonsignificant results could have reached statistical significance had a larger number of dogs been recruited. Second, the information we cited of studies involving human patients and human liver and adrenal tissue slices strongly support the upregulation of 6β-hydroxylase activity in ATs. However, only one dog in the HC group had a presumptive diagnosis of AT. Therefore, we could not compare whether any of the urinary biomarkers in our study differ between dogs with PDH and AT. Nevertheless, in one study, children with PDH (n = 4) and AT (n = 4) did not differ with respect to their urinary 6β-hydroxycortisol levels, and both were significantly higher than urinary 6β-hydroxycortisol levels in age-matched healthy controls. 24 The authors of that study demonstrated that cortisol rather than ACTH upregulates 6β-hydroxylase. 24 Hence we also speculate that high levels of cortisol in dogs upregulate 6β-hydroxylase activity resulting in increased levels of urinary 6β-hydroxycortisol. Although this is not specific for HC (given that urinary 6β-hydroxycortisol was present in high levels in the dogs with CHF), we were able to discriminate between HC and NAD based on urinary 6β-hydroxycortisol levels. Third, from a ML perspective, the small number of dogs enrolled in our study forced us to perform LOOCV. Although this cross-validation decreased the probability for model overfitting, the risk remained high. Hence, we emphasize that a follow-up study with a large cohort of dogs with HC and NAD is essential for the development of a strong ML model. However, the risk of model overfitting does not negate the results that we found and their significance. In particular, we found urinary lactate to be an important component for future development of a ML urinary-based panel for the diagnosis of HC. As well, the effect of prolonged storage at −80°C on the stability of urinary cortisol metabolites and urinary lactate is largely unknown. Although not evaluated in our study, we controlled for storage during the statistical analysis, and storage was not found to have a significant effect on either the mixed-effects model or on the ML model.
