Assessing different cross-validation schemes for predicting novel traits using sensor data: An application to dry...

Area of Science:

• Animal Science
• Dairy Production
• Genetics and Genomics

Background:

• Feed efficiency is crucial for dairy farm profitability, but daily dry matter intake (DMI) recording is costly.
• Mid-infrared (MIR) spectral data from milk offers a potential proxy for DMI prediction.
• Accurate DMI prediction is essential for genetic selection and improving feed efficiency.

Purpose of the Study:

• To evaluate the utility of milk MIR spectral data for predicting proxy phenotypes of DMI.
• To compare predictive models including MIR data, energy sinks, or both, across different cross-validation schemes.
• To assess the performance of MIR-based DMI prediction under cow-independent, experiment-independent, and herd-independent scenarios.

Main Methods:

• Developed three models: MIR data only (M1), energy sinks (body weight, weight change, milk energy) plus cow-level variables (M2), and M2 plus MIR data (M3).
• Utilized weekly and 28-day DMI records from US Holstein cows across multiple experiments and herds.
• Employed 10-fold cow-independent, 10-fold experiment-independent, and 4-fold herd-independent cross-validation schemes.

Main Results:

• Cow-independent cross-validation showed significant improvement with MIR addition (M3 vs. M2), reducing RMSE to 1.59 kg and increasing R² to 0.89.
• Experiment-independent and herd-independent cross-validations revealed no significant benefit of MIR inclusion (M3 vs. M2).
• Broader cross-validation schemes (experiment- and herd-independent) showed increased bias, with herd-independent validation being most critical for real-world application.

Conclusions:

• Milk MIR data shows potential for predicting DMI proxy phenotypes, but its effectiveness is highly dependent on the cross-validation strategy.
• Current MIR-based prediction models lack robustness for herd-independent applications, which are essential for widespread genetic evaluations.
• Future research should focus on developing algorithms suitable for broader cross-validation and improving spectrophotometer calibration for reliable DMI proxy prediction.