Growth charts for small sample sizes using unsupervised clustering: Application to canine early growth

Area of Science:

• Veterinary Medicine
• Animal Science
• Biostatistics

Background:

• Breed-specific growth curves (GCs) are essential for neonatal puppies.
• Obtaining sufficient breed-specific data for GCs can be challenging.

Purpose of the Study:

• To investigate an unsupervised clustering methodology for modeling neonatal puppy GCs.
• To augment limited breed-specific data with data from breeds exhibiting similar growth patterns.

Main Methods:

• Hierarchical clustering on principal components was used to group puppy breeds by median growth profiles (birth to Day 20).
• Generalized Additive Models for Location, Shape and Scale (GAMLSS) were employed to model cluster GCs.
• Cluster-scale breed GCs were generated by recentering cluster models to individual breed profiles.

Main Results:

• Cluster-scale breed GCs maintained high accuracy even with very small sample sizes (down to three puppies).
• This hybrid approach slightly overestimated breed variability but produced smooth, consistent centile curves.
• Breed-specific GCs showed a notable decline in quality with sample sizes of 20 or fewer.

Conclusions:

• A breed-cluster hybrid methodology offers a more satisfactory approach to generating neonatal puppy GCs compared to purely breed-level models when data is scarce.
• This method enhances the reliability of GCs for breeds with limited available data.