Skip to content

An early prediction model for canine chronic kidney disease based on routine clinical laboratory tests

In human healthcare, longitudinal patient data and deep learning approaches have recently been applied to build models to diagnose CKD.25predict future risk of acute kidney injury26 and to differentiate between CKD patients who are more likely to progress27. We recently described the development and validation of a recurrent neural network (RNN) that took advantage of four EMR features (creatinine, blood urea nitrogen, urine specific gravity, and age) to predict cats at risk of developing CKD before clinical signs of the disease. they are apparentfifteen. Here we apply a similar computational modeling approach to a large EHR ensemble from a network of primary care hospitals to derive a model to predict the risk of CKD in canines at a given time based on current and past EHR data. These data, consisting primarily of clinicopathologic findings, were evaluated and refined, resulting in an RNN model consisting of six features: serum creatinine, BUN, urine protein and USG, and patient age and weight. We then evaluated the performance of this model in predicting the risk of dogs developing CKD in the future.

Characteristics of the model and performance of the model.

Of the dogs that developed CKD, 68.8% were correctly predicted 1 year prior to diagnosis, 44.8% 2 years prior to diagnosis, and 23.2% up to 3.5 years prior to clinical diagnosis. Model performance improved with more frequent data collection points, reflected in a sensitivity of 78.6% with data from two visits, which increased to 86.4% when longitudinal data from four visits were available. There are two possible explanations for this observation. First of all, simply having more data available allows the algorithm to make a better prediction. It’s also possible that the vet’s suspicion of an underlying condition leads to more regular testing, which in turn leads to more data and therefore better prediction. Specificity increased to greater than 95% with additional data. The model also performed well across different age categories of dogs, where specificity was consistently above 90% in dogs at the adult (1.5 to 6.5 years), mature (6, 5 to 9.75 years) and senior (9.75 to 13.0 years). As the dogs got older, the specificity dropped to 70% for dogs in the geriatric life stage (over 13 years). In contrast, the sensitivity of the prediction improved in older dogs, with a sensitivity of 99% in dogs 13 years and older. In older pets, it is more common to have multiple diagnoses and comorbid conditions. Here we observed that dogs with nuclear sclerosis and osteoarthritis were more likely to have a false-positive result, perhaps due to the strong association of age in the prevalence of these conditions. Therefore, it is a benefit that this model sacrifices some specificity (which slightly increases the false-positive rate) for better sensitivity (which reduces the false-negative rate) in older pets. In practice, this will help ensure that the doctor does not miss the presence of CKD in an older pet.

clinical utility

The prevalence of disease in the population is particularly important when considering the clinical utility of any diagnostic tool. While sensitivity and specificity are not influenced by disease prevalence, positive (PPV) and negative (NPV) predictive values ​​are, becoming particularly relevant in explaining differences within the test population. Compared to cats, dogs tend to show a low prevalence of CKD (Table 2) and it is important to recognize the impact this will have on the true diagnostic accuracy of the algorithm. Compared to applying a similar modeling approach in cats, where the PPV remained above 90% in all conditions testedfifteen, here the PPV was highest in predicting future risk for dogs 6.5 years and younger (23%), but dropped to 15% for older dogs included in the test data set. PPV is useful to the clinician as it indicates the probability of disease in a patient when the test result is positive, but it is important not to discount the value of a tool that shows lower PPV values.28. Interestingly, the model showed consistent high specificity (97-99%) and NPV (>99%), indicating that a negative test result will accurately predict CKD misdiagnosis in dogs up to 3.5 years in the future, with a very low false positive Speed. With this in mind, the present approach could be used effectively to support proactive wellness initiatives, where the goal is to provide confidence that the pet will remain healthy. In addition, it is important to point out that the population prevalence of canine CKD continues to be a latent variable in our context. We have estimated the prevalence using the EHR data we have available, and this aligns with other previously reported estimates.1 but it may still over- or under-explain the true prevalence of canine CKD within the total population or within particular subpopulations of dogs at risk.

In using EHR to enable the development of an algorithm to predict future disease, confirming the accuracy of the diagnosis was an important first step. The data used to build and validate this model came from a large number of established clinics and a wide range of veterinarians collected over a period of more than 20 years. Dogs with a formal diagnosis of CKD had blood parameters and urine patterns consistent with currently accepted guidelines, and the majority of dogs in the CKD group had creatinine values ​​consistent with IRIS stages 2 and 3 (Fig. 3). ); this provides confidence in using this data to develop the model. Defining the health status of the complementary set of dogs without a formal diagnosis of CKD was more problematic. A subset of these, those classified as ‘probable CKD’, had clear indications of CKD in blood or urine test results or references in medical notes suggestive of CKD. This group of dogs includes those in which the veterinarian was uncertain of the diagnosis (probably due to conflicting information) or because the dog was in an early stage of the disease, or in which, for unknown reasons, they could not be diagnosed. While case-control studies generally exclude these somewhat ambiguous patients, thereby creating a wider gap between groups and improving the statistical significance of the findings, we decided that inclusion of these during the trial phase was important to ensure that the predictive capacity of the algorithm was tested in a real-world scenario where there are ambiguous cases. However, we did not include this group when calculating sensitivity and we are aware that this could bias our estimates as it might contain the most difficult cases to predict. For the other dogs without a formal CKD diagnosis, we imposed a 2-year window with observations to be certain of their “CKD-free” status. This might have lowered our estimates of specificity, as some might have had very early stage CKD that was diagnosed more than 2 years later.

Although serum creatinine is widely cited as a diagnostic marker of CKD and a surrogate indicator of changes in glomerular filtration rate (GFR), it has limitations and should be applied with caution, especially during times of muscle wasting. lean. It is possible that some of the application of creatinine deficiency in CKD may be ameliorated by integrating body weight and age within the algorithm. In human medicine, elimination methods to measure glomerular filtration rate (GFR) are routinely performed in the diagnosis of CKD, and repeated measurements of GFR within the same individual are typically used to determine the effects of treatment. and the prognosis of the disease. Furthermore, statistical approaches to estimate GFR (eGFR) have been validated and implemented in clinical settings to track the progression of CKD patients. Determining GFR in a primary care veterinary setting is not feasible and, to our knowledge, no validated eGFR has been successfully developed in companion animals. A variety of prediction models have been developed to support the diagnosis and progression of kidney disease in humans.25,26,27,29. Predicting patients at risk of CKD progression can help guide individualized treatment plans, and this has been shown to be particularly effective in improving outcomes for those at risk of kidney failure requiring dialysis or transplantation in human patients.29. The present model is not capable of distinguishing between patients in whom the disease will progress and those in whom it will not, and further work to improve the capabilities of the predictive algorithm in this way would be beneficial.

In conclusion, CKD in dogs remains a challenging clinical condition with a shorter survival time compared to cats. When diagnosed, the disease is often advanced (IRIS stage 3 or 4) and, unlike cats receiving appropriate treatment, it usually progresses more rapidly with a poor clinical outcome. The development of tools that allow veterinarians to identify dogs at risk of CKD before the presence of clinical signs opens up opportunities to take a significant step in the management of this condition. The main objective of the present study was to develop a method that allows a prediction for a future diagnosis of CKD in dogs, but the obvious question to ask next is how best to support patients with a positive prediction. In this sense, we believe that it is essential to closely monitor the development of proteinuria and evaluate blood pressure, since both renal proteinuria and systemic hypertension are risk factors in the progression of CKD in dogs.3.4. It would also be prudent to recommend other management tactics, including appropriate nutritional interventions.6, and avoidance or extreme caution with the use of drugs that are potentially nephrotoxic, but more studies are required to validate the effectiveness of such interventions to improve clinical outcome. Although the sensitivity and specificity of our model were similar to the performance of an algorithm designed to predict future CKD risk in cats, the modest PPV means that its utility as an independent screening tool at present may be limited. However, this algorithm showed high specificity and NPV, which makes it particularly effective in identifying patients who will not develop CKD. Taken together, we believe that this approach has the potential to complement existing clinical suspicion and diagnostic investigations of patients, thereby supporting the physician’s decision-making process.

Leave a Reply

Your email address will not be published.