Machine-learning algorithm as a prognostic tool in non-obstructive acute-on-chronic kidney disease in the cat

Abstract

Objectives

The aim of this study was to develop an algorithm capable of predicting short- and medium-term survival in cases of intrinsic acute-on-chronic kidney disease (ACKD) in cats.

Methods

The medical record database was searched to identify cats hospitalised for acute clinical signs and azotaemia of at least 48 h duration and diagnosed to have underlying chronic kidney disease based on ultrasonographic renal abnormalities or previously documented azotaemia. Cases with postrenal azotaemia, exposure to nephrotoxicants, feline infectious peritonitis or neoplasia were excluded. Clinical variables were combined in a clinical severity score (CSS). Clinicopathological and ultrasonographic variables were also collected. The following variables were tested as inputs in a machine learning system: age, body weight (BW), CSS, identification of small kidneys or nephroliths by ultrasonography, serum creatinine at 48 h (Crea₄₈), spontaneous feeding at 48 h (SpF₄₈) and aetiology. Outputs were outcomes at 7, 30, 90 and 180 days. The machine-learning system was trained to develop decision tree algorithms capable of predicting outputs from inputs. Finally, the diagnostic performance of the algorithms was calculated.

Results

Crea₄₈ was the best predictor of survival at 7 days (threshold 1043 µmol/l, sensitivity 0.96, specificity 0.53), 30 days (threshold 566 µmol/l, sensitivity 0.70, specificity 0.89) and 90 days (threshold 566 µmol/l, sensitivity 0.76, specificity 0.80), with fewer cats still alive when their Crea₄₈ was above these thresholds. A short decision tree, including age and Crea₄₈, predicted the 180-day outcome best. When Crea₄₈ was excluded from the analysis, the generated decision trees included CSS, age, BW, SpF₄₈ and identification of small kidneys with an overall diagnostic performance similar to that using Crea₄₈.

Conclusions and relevance

Crea₄₈ helps predict short- and medium-term survival in cats with ACKD. Secondary variables that helped predict outcomes were age, CSS, BW, SpF₄₈ and identification of small kidneys.

Keywords

Survival creatinine machine learning uremic crisis

Introduction

Chronic kidney disease (CKD) is frequently diagnosed in cats. Although it can develop at any age, it is particularly common in geriatric cats, with a reported prevalence of 15–80%.^1,2 Because of the difficulties involved in identifying early renal function loss and the frequent lack of clinical signs early in the course of the disease, CKD is often only diagnosed at an advanced stage. Thus, veterinary care is frequently only sought when acute exacerbation of azotaemia (‘acute-on-chronic kidney disease’ [ACKD]) occurs, and is sometimes responsible for a dramatic clinical picture. Ureteral obstruction (ie, obstructive kidney disease) accounts for some of these events and is associated with specific prognostic and therapeutic considerations. In the remaining cases (often referred to as intrinsic renal failure) some additional degree of parenchymal compromise is responsible for the sudden decrease in kidney function. At the time of presentation, the previous CKD stage and potential reversal of azotaemia are unknown, making it difficult to give prognostic information to the owners who will have to decide whether to engage in long and costly patient care or not.

Serum creatinine (SCr; ie, the stage of CKD) has been consistently identified as a prognostic factor in cats with stable CKD.^3–6 The CKD stage of needs to be determined in a cat with stable renal function and once prerenal factors have been corrected. However, one study found that the stage at diagnosis differed from the stage at baseline, suggesting that an initial acute component of azotaemia is present at diagnosis in the majority of cats.³ In that study, survival differed depending on whether the stage was determined at diagnosis or at baseline: for instance, stage IV cats at diagnosis had a median survival time (MST) of 103 days, whereas stage IV cats at baseline had an MST of only 35 days.³ The initial acute component of azotaemia can be caused by prerenal factors (ie, dehydration) and be volume-responsive, or it can be caused by intrinsic renal or post-renal factors. Other prognostic factors have been identified in stable CKD, including serum phosphorus (SPh), urine protein:creatinine ratio, age, packed cell volume, circulating fibroblast growth factor-23 and parathyroid hormone levels.^3–6

In acute kidney injury (AKI), SCr at diagnosis does not seem to be of prognostic importance.^7,8 The aetiology of AKI is rather related to survival, and infectious, obstructive and ischaemic causes carry a better prognosis than toxic causes.⁹ Two studies developed models to predict death in cases of feline AKI: their performances in predicting the outcome at different time points were good, with an area under the receiver operating characteristics between 0.81 and 0.86.^7,9

Information on prognosis in the case of ACKD is sparse. A recent retrospective study included 100 cats with ACKD of different causes.¹⁰ The survival rate to discharge was 58%. Age, SCr, uraemia and SPh were higher in non-survivors than in survivors, but only SPh remained significantly associated with survival in the final multivariate model. The MST of the survivors after discharge was 66 days. SCr at discharge was significantly associated with long-term survival (for every 1.0 mg/dl increase in SCr at discharge there was a 1.43-fold [95% confidence interval 1.24–1.64] increase in risk of death).¹⁰

The fact that a clinical parameter or a biological variable is correlated with survival does not imply it can predict survival. Machine learning, a branch of artificial intelligence, has been used extensively in human medicine as an aid in medical decicion-making and prognosis.¹¹ In supervised machine learning, the system is trained with instances in which the outcome (ie, what needs to be predicted) is known, to develop the algorithm. The performance of the algorithm is then evaluated using other instances.

The purpose of the present study was to use machine learning to develop an algorithm that may help in the prognosis of cats presenting with intrinsic ACKD.

Materials and methods

Animals

Abdominal ultrasonographic reports of feline patients presented to the Alliance Small Animal Clinic between 2014 and 2018 were collected and the corresponding medical files were reviewed. Cats that had been hospitalised for at least 48 h for acute clinical signs (ie, <14 days) and azotaemia were selected. Initial SCr had to be >212 µmol/l (which was the upper limit of the reference interval [RI] of the biochemistry analyser) and urine specific gravity (USG) <1.035. A SCr value also had to be available <24 h before discharge and a urine culture had to have been performed. All SCr values at diagnosis and during hospitalisation were obtained using the same in-house biochemistry analyser (Catalyst; IDEXX). The presence of CKD was ascertained in one of two ways.

Abdominal ultrasonography

At least two of the following signs had to be present: small kidney size (one kidney or both, sagittal length <32 mm); hyperechoic cortex; reduced corticomedullary demarcation; irregular contour; cortical cysts.^12,13 If a case had a combination of ultrasonographic abnormalities that was equivocal for CKD, detailed retrospective evaluation of the medical file was performed and, if ambiguity persisted, the case was excluded.

Previous history of CKD

Basal SCr had to have been previously reported to be >212 µmol/l along with decreased USG (ie, <1.035) for >3 months, whether CKD could be identified on abdominal ultrasonography or not. If azotaemia was reportedly identified at another institution, the upper limit of the RI of that institution’s analyser was used to define azotaemia.

Finally, the following exclusion criteria were applied: any post-renal cause for azotaemia, known or suspected exposure to nephrotoxicants (except for angiotensin-converting enzyme inhibitors or angiotensin II receptor blockers if used for >30 days); confirmed diagnosis or strong suspicion of renal neoplasia or feline infectious peritonitis; clinical signs not caused by azotaemia according to the clinician in charge of the case; euthanasia for financial reasons.

Data

Information was recorded on signalment, historical complaints and physical examination findings (eg, body weight, body condition score, rectal temperature, hydration status, thoracic auscultation, renal palpation). A score was retrospectively attributed to the following variables: lethargy (absent/mild to moderate/severe: 0/1/2); decreased appetite (absent/dysorexia/anorexia: 0/1/2); digestive signs (none/any severity of vomiting or diarrhoea: 0/1); rectal temperature (<36°C: 2; ⩾36.0°C and <37.5°C: 1; ⩾37.5°C and ⩽38.5°C: 0; >38.5°C and ⩽40°C: 1; >40°C: 2). The sum of these different scores was used to define a clinical severity score (CSS) ranging from 0 to 7. Whenever any values were missing, the CSS was not calculated, and the cats were not included in the final analysis.

Data relative to the available clinicopathological investigations (urinalysis and urine culture, complete blood count and biochemical panel) were also recorded. Symmetric dimethylarginine was not routinely performed in this population of cats. The following ultrasonographic data were gathered: kidney size (length in the sagittal plane); renal pelvis size; presence of nephroliths; cortical segmental lesions; dystrophic mineralisation or perirenal fluid; regularity of the urinary bladder wall; echogenicity of the urinary bladder content.

Finally, SCr values throughout hospitalisation (at admission, at 48 h and upon discharge), spontaneous feeding at 48 h, uraemic crisis aetiology and development of fluid overload during hospitalisation were recorded. As toxic and post-renal causes were excluded, the aetiology of the uraemic crisis could be ischaemic, infectious (ie, pyelonephritis), cardiorenal syndrome (cats presented with heart failure and not treated with a diuretic before the ACKD episode) or unknown. Fluid overload was suspected if congestive signs (ie, cavitary effusion, pulmonary oedema) developed during hospitalisation, along with an increase in body weight and compatible echocardiographic signs. Pyelonephritis was diagnosed when a urine culture from a pyelic sample was positive or when a urine culture from a bladder sample collected by cystocentesis was positive together with one of the following criteria: fever or leukocytosis without an alternative cause, or ultrasonographic signs of pyelonephritis (pyelectasia, echogenic urine in the pyelic cavity, pyelic or ureteral mucosal thickening). Information regarding treatment was not recorded. Treatment was not standardised given the heterogeneity of the population studied and the retrospective nature of this study. However, cats received at least intravenous fluid therapy; other treatments (eg, antiemetics, antibiotics in cases with pyelonephritis) were administered at the discretion of the clinician in charge.

Survival was calculated from the day of discharge from hospitalisation. A survival score of 0 was attributed to cats that died during hospitalisation. For cats that were still alive at the time of data collection, survival was calculated up to 29 January 2020. Information regarding survival was obtained from the medical files or by telephone interview with the owners or the treating veterinarian.

Data analysis

A supervised machine-learning system was used to develop a decision tree algorithm (in R using partykit). A decision tree is a type of algorithm that is easy to display and interpret. The system is provided with known inputs and outputs: the inputs are the variables tested to predict the outputs. The decision tree is formed by a series of branching nodes and terminates in ’leaves’, which are the outputs. The algorithm is developed using optimal discriminant analysis (ODA). ODA is a machine learning algorithm that identifies a threshold (for a continuous variable) or an assignment rule (for a categorical variable) that optimally distinguishes between the outputs.¹⁴ The optimal cut-point is determined by iterating through every value of the attribute and computing the ‘effect strength for sensitivity’ (ESS), which is mean sensitivity across classes. The maximally accurate model uses the optimal cut-point associated with the highest ESS values.¹⁴ Given the number of cases included in our study, we had to select a limited number of inputs in the model. This choice was subjective: we aimed to select meaningful and easily obtained variables that were available for the majority of our cats. The variables tested as inputs were age, body weight, CSS, small kidneys (on abdominal ultrasound), identification of nephroliths, SCr at 48 h (Crea₄₈), spontaneous feeding at 48 h and aetiology. The outputs were survival at 7, 30, 90 and 180 days after discharge.

Finally, the performance of the algorithm to predict survival was evaluated by calculating sensitivity, specificity, and positive and negative predictive values for each survival time point.

Results

Animals

Forty-six cats met the inclusion criteria during the study period. The breeds represented included domestic shorthair (n = 40), Birman (n = 3), Ragdoll (n = 2) and Siamese (n = 1). There were 21 (45.7%) castrated males, 22 (47.8%) spayed females and three (6.5%) intact females. Median age was 10.5 years (range 2.3–17.6) and median body weight was 3.6 kg (range 1.8–7.3).

Baseline characteristics

Detailed information concerning the baseline characteristics (ie, signalment, CSS and ultrasonographic abnormalities tested as inputs in the model) and other variables tested as inputs (ie, Crea₄₈, spontaneous feeding at 48h, aetiology) is provided in Table 1.

Table 1

Detailed demographics and case data included in the analysis

Case	Breed	Sex	Age (years)	BW (kg)	CSS	Small K	Nephroliths	Crea₄₈ (µmol/l)	SpF₄₈	Aetiology	Survival (days)
1	DSH	MN	9.9	3.1	5	No	No	1091	No	0	0
2	DSH	MN	9.4	4.7	3	Both	Yes	2173	Yes	0	0
3	DSH	FN	7.9	2	7	Both	Yes	1046	No	0	0
4	DSH	FN	7.6	3.4	2	One	Yes	1302	No	Pyelonephritis	0
5	Siamese	FN	16	3	6	One	No	1748	No	Pyelonephritis	0
6	DSH	FN	13.5	2.5	6	No	Yes	1001	Yes	0	0
7	DSH	MN	4	2.3	6	Both	Yes	775	No	0	0
8	DSH	MN	5.2	3.1	6	No	No	1285	No	0	0
9	DSH	FN	17.6	2.5	6	No	No	793	No	Pyelonephritis	0
10	DSH	MN	10.5	6.3	5	One	Yes	1571	No	0	0
11	Birman	MN	14.1	3.6	3	One	No	612	Yes	0	1
12*	DSH	F	−	3.1	1	No	Yes	583	Yes	Pyelonephritis	1
13	DSH	MN	13.8	3.4	6	No	Yes	1651	No	0	2
14	DSH	MN	11.2	4.6	5	No	Yes	414	Yes	0	3
15	DSH	FN	15.6	3.9	6	One	No	1111	No	0	4
16	DSH	MN	14	4.6	6	No	No	754	Yes	0	4
17	DSH	MN	10.2	4.8	3	No	No	919	Yes	Pyelonephritis	7
18*	DSH	MN	6.6	3	−	One	Yes	660	Yes	0	9
19	DSH	FN	9.9	2.3	4	One	No	1017	Yes	Pyelonephritis	12
20	DSH	MN	15.4	6.9	3	No	No	819	Yes	0	19
21*	DSH	MN	12.7	5.5	−	One	Yes	1399	No	0	20
22	DSH	MN	17	7.3	0	No	Yes	218	Yes	Cardiorenal syndrome	26
23*	Birman	F	17.5	1.8	3	Both	No	−	Yes	Pyelonephritis	32
24	DSH	MN	7.4	2.8	5	No	Yes	636	No	0	34
25	DSH	MN	12.6	4.8	1	No	Yes	1039	Yes	0	39
26	Birman	MN	15.3	3.7	3	One	No	347	Yes	0	44
27	DSH	MN	17.5	4.1	5	One	No	1013	Yes	0	50
28	DSH	FN	13.1	2.3	3	Both	Yes	501	Yes	0	53
29	Ragdoll	FN	5.2	4	6	One	Yes	240	Yes	0	82
30	DSH	FN	3.6	2.4	2	No	Yes	395	Yes	0	93
31	DSH	FN	11.3	5.5	5	No	Yes	985	Yes	0	161
32	DSH	FN	11.8	3.9	3	One	Yes	496	Yes	0	183
33	DSH	MN	5.4	3.6	4	One	Yes	824	Yes	0	187
34	DSH	FN	17.3	5.4	4	No	Yes	283	No	0	322
35	DSH	FN	6.1	3.7	5	One	Yes	209	No	Pyelonephritis	356^†
36	DSH	FN	9.5	4	6	One	Yes	305	Yes	Pyelonephritis	560^†
37	DSH	FN	10.9	4.1	2	One	No	176	Yes	0	627^†
38	DSH	FN	15.8	2.8	1	One	Yes	279	Yes	Pyelonephritis	672^†
39	DSH	FN	7.3	2.5	3	One	Yes	389	Yes	0	700^†
40	DSH	MN	7.3	6.3	5	One	Yes	244	No	Pyelonephritis	738^†
41	DSH	FN	10.5	3.8	1	Both	No	325	Yes	0	861
42	Ragdoll	FN	6.7	3.5	3	One	No	1398	Yes	0	985^†
43	DSH	MN	9.3	5.4	4	One	Yes	300	Yes	Pyelonephritis	1065^†
44	DSH	FN	2.3	1.8	3	One	No	548	Yes	Pyelonephritis	1148^†
45	DSH	FN	5.8	3.4	4	One	Yes	670	Yes	0	1448^†
46	DSH	F	8	2.8	5	No	No	262	Yes	Pyelonephritis	1886^†

Cases are classified according to survival (surviving <7 days, at least 7, 30, 90 or 180 days)

Not included in the analysis because of missing data

†

Cat still alive on completion of the study

BW = body weight; CSS = clinical severity score; small K = small kidneys as determined by abdominal ultrasonography; Crea₄₈ = serum creatinine 48 h after admission to hospital; SpF₄₈ = spontaneous feeding after 48 h in hospital; DSH = domestic shorthair; MN = male neutered; FN = female neutered; F = female intact

CKD identification

USG was measured on a urine sample obtained before starting fluid therapy in all but nine cats.

CKD was identified ultrasonographically in all cats. In nine cats, azotaemia had been previously identified for >3 months; all these cats also had ultrasonographical abnormalities consistent with CKD.

Aetiology

Pyelonephritis was diagnosed in 14 cats based on the predefined criteria. No cat with a positive urine culture had subclinical bacteriuria. Cardiorenal syndrome of suspected type 1 (acute impairment of the cardiac function causing acute kidney injury) was identified in one cat presented with hypertrophic cardiomyopathy, pleural effusion and azotaemia; this cat was not receiving any treatment at the time of diagnosis. The aetiology of the uraemic crisis could not be determined in the remaining 31 cats.

Hospitalisation and survival

Median SCr at admission, at 48 h and upon discharge were, respectively, 755 µmol/l (n = 46 cats; range: 267–1688), 670 µmol/l (n = 45 cats; range 176–2173) and 392 µmol/l (n = 36 cats; range 133–1487). Median hospitalisation time was 4 days (range 2–10).

Thirty-six cats survived the hospitalisation period and 24 were alive 1 month after discharge. Of the cats that did not survive the hospitalisation period, one died shortly after seizures and the remaining nine cats were euthanased. The overall median survival time was 33 days (range 0–1886).

Prediction of survival

As data on four cats were lacking, only 42 cats were included in the final model.

Of the tested variables, Crea₄₈ was the one that best distinguished between survivors and non-survivors for all time points.

Thirty-three cats had a Crea₄₈ <1043 µmol/l and 26 of them were alive 7 days after discharge. Nine cats had a Crea₄₈ ⩾1043 µmol/l and only one of them was alive 7 days after discharge. The median Crea₄₈ of cats that survived at least 7 days was 496 µmol/l (range 176–1399); for those that did not, Crea₄₈ was 1068 µmol/l (range 414–2173).

Eighteen cats had a Crea₄₈ <566 µmol/l and 16 and 13 of them were alive 30 and 90 days after discharge, respectively. Twenty-four cats had a Crea₄₈ ⩾566 µmol/l and only seven and four of them were alive 30 and 90 days after discharge, respectively. The median Crea₄₈ was 389 µmol/l (range 176–1398) in cats that survived at least 30 days, and 1009 µmol/l (range 218–2173) in cats that did not. The median Crea₄₈ was 325 µmol/l (range 176–1398) in cats that survived at least 90 days, and 960 µmol/l (range 218–2173) in cats that did not.

Thirteen cats had a Crea₄₈ <389 µmol/l and 10 of them were alive 180 days after discharge. Twenty-nine cats had a Crea₄₈ ⩾389 µmol/l. Of these, 22 were >7.1 years old and 1/22 survived at least 180 days, while seven cats were <7.1 years old and 4/7 survived at least 180 days. The median Crea₄₈ in cats that survived at least 180 days was 305 µmol/l (range 176–1398) and it was 952 µmol/l (range 218–2173) in those that did not.

If Crea₄₈ was excluded from the analysis, the algorithms presented a more complex branching pattern (Figures 1 –4).

Figure 1

Decision tree algorithm to predict 7-day survival when age, body weight (BW; in kg), clinical severity score (CSS), identification of small kidneys by ultrasonography, presence of nephroliths, spontaneous feeding at 48 h and aetiology were included in the analysis but serum creatinine at 48 h was not. White boxes are the nodes of the decision tree and black boxes are the leaves (outputs)

Figure 2

Decision tree algorithm to predict 30-day survival when age (years), body weight, clinical severity score (CSS), identification of small kidneys by ultrasonography, presence of nephroliths, spontaneous feeding at 48 h and aetiology were included in the analysis but serum creatinine at 48 h was not. White boxes are the nodes of the decision tree and black boxes are the leaves (outputs)

Figure 3

Decision tree algorithm to predict 90-day survival when age (years), body weight, clinical severity score, identification of small kidneys by ultrasonography, presence of nephroliths, spontaneous feeding at 48 h (SpF₄₈) and aetiology were included in the analysis but serum creatinine at 48 h was not. White boxes are the nodes of the decision tree and black boxes are the leaves (outputs)

Figure 4

Decision tree algorithm to predict 180-day survival when age (years), body weight, clinical severity score, identification of small kidneys by ultrasonography, presence of nephroliths, spontaneous feeding at 48 h and aetiology were included in the analysis but serum creatinine at 48 h was not. White boxes are the nodes of the decision tree and black boxes are the leaves (outputs)

Performance of the algorithms

Sensitivity, specificity, and positive and negative predictive values of the different algorithms are reported in Tables 2 and 3.

Table 2

Performance of the algorithm in predicting survival when age, body weight, clinical severity score, identification of small kidneys by ultrasonography, presence of nephroliths, serum creatinine at 48 h, spontaneous feeding at 48 h and aetiology were included in the analysis

	Sensitivity	Specificity	PPV	NPV
7-day survival	0.96	0.53	0.79	0.89
30-day survival	0.70	0.89	0.89	0.71
90-day survival	0.76	0.80	0.72	0.83
180-day survival	0.93	0.78	0.70	0.95

PPV = positive predictive value; NPV = negative predictive value

Table 3

Performance of the algorithm in predicting survival when age, body weight, clinical severity score, identification of small kidneys by ultrasonography, presence of nephroliths, spontaneous feeding at 48 h and aetiology were included in the analysis but serum creatinine at 48 h was not

	Sensitivity	Specificity	PPV	NPV
7-day survival	0.81	0.87	0.92	0.72
30-day survival	0.83	0.79	0.83	0.79
90-day survival	0.76	0.76	0.68	0.83
180-day survival	0.80	0.78	0.67	0.88

PPV = positive predictive value; NPV = negative predictive value

Discussion

Our study showed that in this population of cats, Crea₄₈ was the best predictor of short- and medium-term survival in intrinsic ACKD. It was such a strong predictor that it had to be removed from the analysis for algorithms with other variables to emerge. These algorithms were more complex but still predicted survival quite well.

Crea₄₈ was chosen over SCr at admission as an input in the machine-learning system as the prerenal component of azotaemia, which is usually present at admission, can cause overestimation of the true degree of renal dysfunction. The fact that Crea₄₈ is such a strong predictor of the outcome is consistent with the results of previous studies on CKD but not with what has been reported for feline AKI. In a recent study on ACKD, SCr at discharge was also associated with long-term survival.¹⁰ It has recently been suggested that progressive CKD may be a slow form of AKI or may be the result of several episodes of AKI.¹⁵ ACKD could thus be a feature of progressive CKD. AKI is characterised by a potential for reversal, although it has been reported that >50% of the cats that survive an AKI episode show persistent azotaemia.⁸ It is possible that the potential for renal cellular repair is reduced when CKD is already present.¹⁰ In our study, overall survival was short, in line with the recent report on ACKD,¹⁰ but the range was very wide. This could be explained by sustained renal injury, as the inciting cause of the ACKD could not be identified in the majority of cats and its potential specific treatment could not be implemented. Furthermore, the cases of pyelonephritis were treated according to current guidelines, but the success of the treatment is difficult to ascertain and a subclinical chronic pyelonephritis may have persisted in some cats despite adequate treatment and follow-up care.

Among the variables that emerged in the algorithms generated when Crea₄₈ was excluded from the analysis, the CSS was also a good predictor of survival. It included rectal temperature, which has already been linked to disease severity and outcome in acute azotaemia in cats.^7,9,16 The aetiology of uraemic hypothermia is poorly understood, but it is speculated that uraemic toxins may cause cellular hypometabolism.¹⁷ Digestive signs (ie, vomiting and, less frequently, diarrhoea) are also common in azotaemic cats, even if other causes may coexist in an individual patient. In addition to respiratory signs and ataxia, vomiting was also associated with survival in a study on AKI in cats.⁹ The point of a disease severity score is to provide objective scoring of the disease clinical activity, which could be used in research to evaluate the efficacy of a treatment, to predict outcome and to help the stratification of cases.¹⁸ To validate such a score, it would be necessary to demonstrate a correlation between the score and a known disease activity marker. In AKI, SCr would probably be the most suitable marker for the purpose. On the contrary, many cats with CKD display very few – if any – clinical signs, with sometimes moderate-to-marked azotaemia; this is believed to be due to the progressive nature of the decline in renal function and the time allowed to compensate for it. Therefore, the best marker for disease activity in the particular case of ACKD remains to be identified. Finally, the clinical information (ie, lethargy, anorexia) included in the CSS and their scoring were retrospectively determined, which led to inherent imprecision.

Age was also associated with outcome, with fewer older cats surviving in most of the algorithms. Age has already been reported to be a negative prognostic factor in both CKD and AKI in cats.^4,7 Prevalence of CKD increases with age; ageing has also been proposed as a contributing factor to CKD development.¹⁹ Older cats may have had a more advanced stage of CKD before ACKD, thereby explaining its association with poorer outcomes. This could not be ascertained here as the previous CKD stage was documented in only 9/46 cats.

Body weight helped predict survival for the 7-day survival period. Weighing >3.65 kg or <2.95 kg was associated with survival. The reasons for this result are not clear. It may be that the cats were of different sizes and their weight did not reflect their overall body condition. Body condition and especially muscle condition scores would likely be better indicators of the cats’ overall physical condition. In this study, they were not consistently available in the medical files and unfortunately could not be included in the model.

Spontaneous feeding after 48 h in hospital was associated with survival in the 90-day survival period algorithm. This might be explained by a more rapid reversal of azotaemia and less severe clinical consequence of the uraemic crisis in the cats that were able to spontaneously eat in hospital. Identification of small kidneys on abdominal ultrasound was surprisingly associated with survival in the 180-day survival algorithm. One would expect small kidneys to be associated with a more advanced disease, and the reason for this finding is not clear.

Our study had several limitations. Firstly, its retrospective nature led to inherent imprecision in the collection of some data or some missing data in the medical files. Secondly, the CSS had to be calculated retrospectively from information retrieved from the medical records. Furthermore, this clinical scoring system was not previously validated. Prospective studies are therefore needed to validate a clinical score that could be used consistently in future studies. Thirdly, the performances of the algorithm were evaluated using the same cases as those used to train it, whereas it would have been ideal to test the algorithms on another population of cats (ie, cats from a different institution). The algorithm performed well despite the limited number of cases; however, had we had access to more cases, we would probably have been able to test more variables as inputs and to design a more finely tuned algorithm that performed even better. Lastly, treatment and follow-up were not standardised, which – at least in theory – may have influenced our results.

Conclusions

The use of a machine-learning system on the data of this small-scale study identified Crea₄₈ as the most useful prognostic indicator in cases of feline ACKD. Our promising results now need to be confirmed in a larger-scale study in which more potential predictive variables are tested to develop a more finely tuned and more efficient algorithm.

Footnotes

Accepted: 3 February 2021

Author note

This paper was presented in part at the 2019 ECVIM Forum.

Conflict of interest

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Ethical approval

This work involved the use of non-experimental animals only (including owned or unowned animals and data from prospective or retrospective studies). Established internationally recognised high standards (‘best practice’) of individual veterinary clinical patient care were followed. Ethical approval from a committee was therefore not specifically required for publication in JFMS.

Informed consent

Informed consent (either verbal or written) was obtained from the owner or legal custodian of all animal(s) described in this work (either experimental or non-experimental animals) for the procedure(s) undertaken (either prospective or retrospective studies). For any animals or humans individually identifiable within this publication, informed consent (either verbal or written) for their use in the publication was obtained from the people involved.

ORCID iD

Mathieu R Faucher

References

Lulich

Osborne

O’Brien

, et al. Feline renal failure: questions, answers, questions. Compend Cont Ed Pract Vet 1992; 14: 127–152.

Marino

Lascelles

BDX

Vaden

, et al. Prevalence and classification of chronic kidney disease in cats randomly selected from four age groups and in cats recruited for degenerative joint disease studies. J Feline Med Surg 2014; 16: 465–472.

Boyd

Langston

Thompson

, et al. Survival in cats with naturally occurring chronic kidney disease (2000–2002). J Vet Intern Med 2008; 22: 1111–1117.

Geddes

Elliott

Syme

. Relationship between plasma fibroblast growth factor-23 concentration and survival time in cats with chronic kidney disease. J Vet Intern Med 2015; 29: 1494–1501.

King

Tasker

Gunn-Moore

, et al. Prognostic factors in cats with chronic kidney disease. J Vet Intern Med 2007; 21: 906–916.

Syme

Markwell

Pfeiffer

, et al. Survival of cats with naturally occurring chronic renal failure is related to severity of proteinuria. J Vet Intern Med 2006; 20: 528–535.

Lee

Chan

Hsu

, et al. Prognostic factors and a prognostic index for cats with acute kidney injury. J Vet Intern Med 2012; 26: 500–505.

Worwag

Langston

. Acute intrinsic renal failure in cats: 32 cases (1997–2004). J Am Vet Med Assoc 2008; 232: 728–732.

Segev

Nivy

Kass

, et al. A retrospective study of acute kidney injury in cats and development of a novel clinical scoring system for predicting outcome for cats managed by hemodialysis. J Vet Intern Med 2013; 27: 830–839.

10.

Chen

Dunaevich

Apfelbaum

, et al. Acute on chronic kidney disease in cats: etiology, clinical and clinicopathologic findings, prognostic markers, and outcome. J Vet Intern Med 2020; 34: 1496–1506.

11.

Awaysheh

Wilcke

Elvinger

, et al. Review of medical decision support and machine-learning methods. Vet Pathol 2019; 56: 512–525.

12.

Bragato

Borges

Fioravanti

MCS

. B-mode and Doppler ultrasound of chronic kidney disease in dogs and cats. Vet Res Commun 2017; 41: 307–315.

13.

Paepe

Bavegems

Combes

, et al. Prospective evaluation of healthy Ragdoll cats for chronic kidney disease by routine laboratory parameters and ultrasonography. J Feline Med Surg 2013; 15: 849–857.

14.

Linden

Yarnold

. Modeling time-to-event (survival) data using classification tree analysis. J Eval Clin Pract 2017; 23: 1299–1308.

15.

Cowgill

Polzin

Elliott

, et al. Is progressive chronic kidney disease a slow acute kidney injury? Vet Clin North Am Small Anim Pract 2016; 46: 995–1013.

16.

Fults

Herold

. Retrospective evaluation of presenting temperature of urethral obstructed male cats and the association with severity of azotemia and length of hospitalisation: 243 cats (2006–2009). J Vet Emerg Crit Care (San Antonio) 2012; 22: 347–354.

17.

Kabatchnick

Langston

Olson

, et al. Hypothermia in uremic dogs and cats. J Vet Intern Med 2016; 30: 1648–1654.

18.

Hayes

Mathews

Kruth

, et al. Illness severity scores in veterinary medicine: what can we learn? J Vet Intern Med 2010; 24: 457–466.

19.

Jepson

. Current understanding of the pathogenesis of progressive chronic kidney disease in cats. Vet Clin North Am Small Anim Pract 2016; 46: 1015–1048.