Supplemental Tables for milk FTIR spectra vs health papers

This repository contains the supplemental tables associated with two related studies:

"Predicting postpartum diseases in Holstein cows using milk spectra and machine learning – retrospective assessment from diagnosis date", and

"Prediction and classification of metritis and mastitis in Holstein cows using transition milk spectra and machine learning – prospective assessment following parturition"

Both studies aimed to evaluate the feasibility of using milk Fourier-transform infrared (FTIR) spectroscopy for the early detection of postpartum diseases in Holstein dairy cows. By leveraging machine learning, deep learning, and regression-based models, the studies assessed milk spectral patterns collected within the first 30 or 14 days in milk (DIM), incorporating both spectral data and cow-level features.

In the retrospective study, we examined predictive performance over eight distinct time points prior to clinical diagnosis. In the prospective study, we focused on differences in milk spectral profiles between cows that developed metritis or mastitis and healthy counterparts, identifying temporal patterns in fat, protein, and lactose absorption.

The supplemental tables provided here include detailed results from all statistical analyses, model performance metrics (e.g., AUROC, accuracy, sensitivity, specificity), permutation testing results, and comparisons across different feature sets.

Supplemental Table for study 1:

001-Supplemental Table 1. The comparision of milk yield and SCC between healthy group and disease group at different time periods prior to disease diagnosed date.  We presented both original and adjusted P-values (derived from Multiple Mann-Whitney U test with BH correction) here.

001-Supplemental Table 2. The comparision of absorbance at six wavenumber (2923, 2854, 1747, 1157, 1542 and 1041 cm-1) between healthy group and disease group at different time periods prior to disease diagnosed date.  We presented both original and adjusted P-values (derived from Multiple Mann-Whitney U test with BH correction) here.

001-Supplemental Table 3. Model performance (AUROCs, Accuracy, Sensitivity, and Specificity; mean with 95%CI with 1,000 permutation testing results, from both outer validation and inner training and validation) using spectra and their first derivative by PLS-DA.

001-Supplemental Table 4. Model performance (AUROCs, Accuracy, Sensitivity, and Specificity; mean with 95%CI with 1,000 permutation testing results, from both outer validation and inner training and validation) using spectra coupled with some cow-level features by PLS-DA, ridge regression, random forest and LSTM.

Supplemental Table for study 2:

002-Supplemental Table 1. The comparision of milk yield and SCC between healthy group and two disease groups at each DIM.  We presented both original and adjusted P-values (derived from Multiple Mann-Whitney U test with BH correction) here.

002-Supplemental Table 2. The comparision of absorbance at six wavenumber (2923, 2854, 1747, 1157, 1542 and 1041 cm-1) between healthy group and two disease groups at each DIM.  We presented both original and adjusted P-values (derived from Multiple Mann-Whitney U test with BH correction) here.

002-Supplemental Table 3. Model performance for metritis prediction (AUROCs, Accuracy, Sensitivity, and Specificity; mean with 95%CI with 1,000 permutation testing results, from both outer validation and inner training and validation) using spectra coupled with some cow-level features by PLS-DA, random forest and LSTM. 

002-Supplemental Table 4. Model performance for mastitis prediction (AUROCs, Accuracy, Sensitivity, and Specificity; mean with 95%CI with 1,000 permutation testing results, from both outer validation and inner training and validation) using spectra coupled with some cow-level features by PLS-DA, random forest and LSTM.

002-Supplemental Table 5. Model performance for metritis and mastitis classification (AUROCs, Accuracy, Sensitivity, and Specificity; mean with 95%CI with 1,000 permutation testing results, from both outer validation and inner training and validation) using spectra coupled with some cow-level features by PLS-DA, random forest and LSTM.

The supplemental figures provided here include detailed linear relationship between the predictive milk components and the absorption peaks of these components (fat, true protein and lactose) and predictive performance (accuracy, sensitivity and specificity).

Supplemental Figure for study 1:

001-Supplemental Figure 1. Line plots with 95% confident interval showing the linear relationship between the predictive milk components and the absorption peaks of these components (fat, true protein and lactose) at eight separate time periods prior to disease diagnosis (>10 d, 10 – 8 d, 7 – 6 d, 5 – 4 d, 3 d, 2 d, 1 d, and 0 d). Coefficient of determination(R2) which represents the percentage of the variance in the absorption peaks that the predictive milk components explain collectively, and P-value which indicates overall significance of the regression model were present.

001-Supplemental Figure 2. Predictive performance (Accuracy, Sensitivity, and Specificity) of PLS-DA using milk FTIR spectra (R1+R2+R3 and R1+R2+R3+R4) and the first derivative.

Supplemental Figure for study 2:

002-Supplemental Figure 1. Line plots with 95% confident interval showing the linear relationship between the predictive milk components and the absorption peaks of these components (fat, true protein and lactose) at different DIM. Coefficient of determination (R2) which represents the percentage of the variance in the absorption peaks that the predictive milk components explain collectively, and P-value which indicates overall significance of the regression model were present.

002-Supplemental Figure 2. Model performance for metritis/mastitis prediction and classification (Accuracy, Sensitivity, and Specificity) of various models (PLS-DA, random forest and LSTM) using milk FTIR spectra (R1+R2+R3). NA indicates that the time points were not included in modeling due to fewer than 10 samples per group. RF: random forest.