AI-Assisted Kidney Screening: How Machine Learning Is Changing Feline Diagnostics

Artificial intelligence has already demonstrated its value in veterinary diagnostic imaging, from identifying cardiac enlargement on radiographs to flagging abnormalities in ultrasound studies. But the application of AI to laboratory diagnostics — specifically, using machine learning to extract more clinical insight from routine bloodwork — represents a newer and potentially transformative frontier.

In feline medicine, where chronic kidney disease is both highly prevalent and frequently diagnosed late, AI-assisted screening has the potential to close a critical gap between available data and actionable clinical insight.

The Problem AI Can Help Solve

The core challenge in feline CKD screening is not a lack of diagnostic tools. We have creatinine, SDMA, urinalysis, UPC ratios, blood pressure measurement, and renal imaging. The challenge is that:

1. Not all tests are performed on every visit. A wellness panel may include creatinine and BUN but not SDMA, and a urine sample may not be collected.
2. Subtle patterns across multiple values are difficult to detect manually. A creatinine of 1.5, a BUN of 32, a phosphorus trending up over three visits, and a slightly low albumin may individually look unremarkable but together suggest early renal compromise.
3. Veterinary workloads limit the time available for nuanced interpretation. In a busy clinic seeing 25 or more patients per day, there is limited time to trend historical values and identify subtle patterns.

AI is well-suited to address all three of these challenges.

How Machine Learning Approaches Kidney Screening

Machine learning models for kidney screening typically work by analyzing patterns across multiple biomarkers to predict an outcome that would otherwise require an additional test. In the case of RadAnalyzer USG Screening, the model uses routine bloodwork values to predict whether a cat's urine specific gravity is likely to be low — indicating potential renal concentrating deficiency — even when a urine sample was not collected.

The Approach

1. Training data: The model is trained on historical records where both bloodwork and urinalysis results are available, creating a mapping between serum values and urine concentration.
2. Feature selection: The algorithm identifies which bloodwork parameters are most predictive of USG. These typically include creatinine, BUN, phosphorus, and several other values from a standard chemistry panel.
3. Prediction: For a new patient where only bloodwork is available, the model outputs a probability that the cat's USG falls below a clinically significant threshold.
4. Clinical decision support: The output is not a diagnosis but a flag — it identifies patients who would benefit most from follow-up urinalysis and comprehensive renal workup.

What This Is Not

It is important to be clear about the boundaries:

• AI screening does not replace urinalysis — it identifies patients who need it most
• The model does not diagnose CKD — it flags risk
• AI is a clinical decision support tool, not an autonomous diagnostic system
• Predictions are probabilistic and should be interpreted in clinical context

Why Bloodwork-Based Prediction Matters

The most significant practical benefit of this approach is that it works with data that already exists. Every cat that gets a wellness chemistry panel generates the inputs the model needs. No additional sample collection, no additional test cost, and no additional time from the clinician.

This is particularly valuable in scenarios where:

• Urine collection is not feasible during the visit
• Client cost concerns prevent adding urinalysis to the wellness panel
• The cat appears healthy and there is no obvious clinical indication for a full renal workup
• Retrospective screening of existing patient records could identify cats that were missed on previous visits

By lowering the barrier to screening, AI-assisted tools can increase the number of cats that receive meaningful renal assessment.

Validation and Clinical Evidence

Any AI tool used in clinical practice must be validated rigorously. The key metrics for evaluating a kidney screening model include:

• Sensitivity: The ability to correctly identify cats with low USG (true positive rate). High sensitivity minimizes missed cases.
• Specificity: The ability to correctly identify cats with normal USG (true negative rate). High specificity minimizes false alarms.
• Positive predictive value: Given a positive flag, what is the probability that the cat actually has low USG?
• Negative predictive value: Given a negative result, what is the probability that the cat is truly fine?

For a screening tool, sensitivity is prioritized over specificity. It is better to flag a cat that turns out to be fine (leading to a urinalysis that confirms normal renal function) than to miss a cat with early kidney disease.

RadAnalyzer USG Screening is currently being piloted at Memorial Cat Hospital in Houston, Texas, with two additional Texas clinics in the process of onboarding. We believe in transparency about our model's performance and will share validation results as the pilot progresses.

The Broader Vision for AI in Veterinary Lab Diagnostics

Kidney screening is one application, but the same principles extend to other areas of veterinary laboratory medicine:

• Predictive endocrine panels — Using chemistry patterns to flag patients likely to have thyroid or adrenal dysfunction
• Anemia characterization — Using CBC patterns to narrow differential lists before manual review
• Pre-anesthetic risk scoring — Combining bloodwork, age, and breed factors to predict anesthetic complications
• Longitudinal health monitoring — AI that tracks a patient's lab values over their lifetime and alerts clinicians to subtle trends

The common thread is using machine learning to extract more clinical value from data that is already being generated, reducing missed diagnoses without adding cost or workflow burden.

Key Takeaways

• AI-assisted kidney screening uses routine bloodwork to predict renal concentrating ability without requiring a urine sample
• The goal is clinical decision support — flagging at-risk patients for follow-up, not replacing urinalysis
• Bloodwork-based screening works with data already generated during wellness visits, adding no cost or collection burden
• Validation through prospective, multi-clinic pilots is essential before clinical adoption
• The same machine learning principles apply to a wide range of veterinary lab diagnostics
• Transparency about model performance and limitations is critical for responsible clinical integration