Feline USG Predictor v1.6

Estimation of Urine Specific Gravity from Serum Chemistry, CBC, and Patient Age  |  Feline-only clinic  |  April 2026

1. Clinical Rationale

Urine Specific Gravity is a cornerstone of feline renal assessment. IRIS staging of Chronic Kidney Disease incorporates USG alongside creatinine and SDMA to differentiate stages and guide management. Loss of concentrating ability (USG <1.035 in cats) is frequently among the earliest detectable signs of tubular dysfunction, often preceding azotemia.

However, urine collection is not always achievable at the time of presentation. An empty bladder, patient temperament, contraindications to cystocentesis (coagulopathy, abdominal masses), or client constraints may preclude urinalysis.

This model estimates USG from serum chemistry, CBC, and patient age — values already being collected on the same blood draw — providing a screening estimate of concentrating ability at zero additional cost, zero additional procedure time, and zero additional client charge.

2. What Changed in v1.6

Summary of Changes

*v1.5a used row-level cross-validation where the same patient’s multiple visits could appear in both training and evaluation. Its metrics are not directly comparable to v1.6’s patient-independent evaluation.

Why Sensitivity Went Up and Specificity Went Down

With 92% sensitivity on the holdout set, only 9 out of 107 impaired cats were missed. The trade-off is that 57 out of 141 healthy cats are flagged for urinalysis they may not need. For those 57 cats, the consequence is a minimal-cost add-on test that confirms they are healthy. For the 98 impaired cats correctly flagged, the consequence is early detection and intervention.

3. Validation Methodology

v1.6 implements a rigorous validation strategy designed to eliminate optimistic bias and produce publication-grade metrics:

Smart Temporal Split

The holdout set is constructed by taking each patient’s most recent visit and reserving it for final evaluation. This simulates real deployment: the model trains on historical data and is evaluated on the most recent encounter for each cat — exactly the scenario it faces in clinical use.

4. Algorithm & Architecture

Multi-Model Ensemble

v1.6 uses an ensemble of independently-trained machine learning models. Each model is trained on a different cross-validation fold, meaning each sees the development data from a different angle. At prediction time, all models produce score estimates that are averaged for the final prediction.

Required Inputs (5)

Same 5 features as v1.5a — 3 lab values from a standard chemistry + CBC panel, plus patient age. No Amylase, Cholesterol, T4, electrolyte panel, or SDMA required.

Hyperparameter Optimization

Extensive GPU-accelerated hyperparameter search across tree structure, regularization, and class weighting. The optimization objective minimizes clinically-weighted error cost with asymmetric penalties that prioritize sensitivity, enforcing a minimum 90% sensitivity target.

Output

Binary screening classification with uncertainty:

• Classification: Adequate (≥1.035) or Impaired (<1.035)
• Adequacy score on a 0–1 scale (ensemble mean); higher = more likely adequate. Uncalibrated — treat as a ranking signal, not a probability.
• Ensemble uncertainty (standard deviation across models)
• Conformal prediction set: {impaired}, {adequate}, or {impaired, adequate} = uncertain
• Per-feature SHAP explanations showing which values drive the prediction

5. Performance Metrics

Temporal Holdout (n=248, never seen during training)

Each patient’s most recent visit, held out completely from training. This is the primary metric set — it represents expected real-world performance.

Out-of-Fold Cross-Validation (n=1,086)

Each prediction is from the one model (out of 7) that never saw this data point during training. Validates consistency across the full development set.

OOF and holdout metrics are consistent (<2% sensitivity difference), confirming the model generalizes well. AUC-ROC of 0.86–0.87 across both sets demonstrates strong discrimination ability independent of the chosen threshold.

Understanding AUC-ROC: The Threshold-Independent View

6. Threshold & Clinical Trade-offs

The decision threshold was optimized via clinically-weighted search. The optimization objective: minimize missed impaired cats while keeping the false flag rate clinically manageable, with a target sensitivity ≥90%.

This asymmetry is not unique to our tool — it is the foundation of all clinical screening programs. Mammography, PSA testing, and fecal occult blood tests all operate at high-sensitivity/moderate-specificity because the cost of a missed diagnosis vastly exceeds the cost of additional follow-up testing.

The Sensitivity-Specificity Trade-off in Context

Moving from v1.6’s operating point to a “balanced” threshold would reduce unnecessary UAs by ~22, but miss 12 additional sick cats. Those 12 cats may not be diagnosed until their next visit months later, after further nephron loss.

7. Conformal Prediction

New in v1.6: every prediction includes a conformal prediction set with a mathematically guaranteed coverage rate. This is a distribution-free method that tells you not just what the model predicts, but how confident it is.

Coverage Guarantee

The conformal layer guarantees that the true label is contained in the prediction set at least 90% of the time (α = 0.10). This is a distribution-free guarantee — it holds regardless of the underlying data distribution, requiring only the assumption of exchangeable data.

Conformal prediction calibrated on out-of-fold predictions from the development set. Target coverage: 90%.

8. Confusion Matrix

Temporal Holdout (n=248)

Out-of-Fold (Development Set, n=1,086)

OOF results confirm the holdout pattern: 90.5% sensitivity (523/578), 57.3% specificity (291/508). Consistent performance across both evaluation sets.

9. Feature Importance & Physiological Basis

With only 5 features, the model concentrates its predictive power on the strongest renal and hematologic markers. BUN, Age, and Creatinine account for over 70% of the model’s total importance:

Physiological Interpretation

10. Prediction Explanations

Every prediction includes a per-feature explanation breakdown showing how each bloodwork value contributed to the result. In v1.6, explanation values are averaged across all ensemble models for more stable, reliable attributions.

This transparency helps veterinarians understand which bloodwork values are driving the recommendation, rather than treating the model as a black box. For example, a cat might be flagged primarily because of elevated BUN and advanced age, even though its creatinine is still within normal range — the explanation chart makes this reasoning visible.

11. Version History

†v1.5a used row-level cross-validation with patient overlap between splits. v1.6 is the first version with publication-grade patient-independent validation.

12. Limitations

13. Study Population

Age Distribution

Model v1.6  |  April 2026  |  Feline-only clinic  |  For investigational and research use  |  Not validated for clinical deployment