1. Predicting survivors of Neonatal calf diarrhea (NCD) using Logistic Regression or Gradient Boosting
Stefano Biffani*, Cesare Lubiano1, Davide Pravettoni1 and Antonio Boccardo1
*IBBA-CNR (UOS-Lodi),
1 Clinic for ruminants, swine and management - Large Animal Veterinary Teaching Hospital (UNIMI-LODI)

2. Background
NCD
Neonatal calf diarrhea (NCD), is a common disease affecting the newborn calf.

3. Affected by Neonatal calf diarrhea (NCD) within 1st 2 weeks
Background
Affected by Neonatal calf diarrhea (NCD) within 1st 2 weeks
25-35 %

4. Background
Clinical problems
NCD
60-80 %

5. Consequences over $250 million /year Death Reduced growth
Veterinary costs
Reduced growth
Death

6. The idea NCD Clinical data Ancillary tests
LODI - Large Animal Veterinary Teaching Hospital
NCD
Clinical data
Ancillary tests

7. The idea
Emogas Analysis (pH, Ht, NA+,K+, Cl-,Anion Gap, pCO2, bicarbonati, base excess)
Total Protein
Dehydration and vitality Score
Rectal temperature
Age, hospitalization month, Sex
Features

8. Will it survive ? The idea Features Statistical Learning:
Tools for understanding data
Will it survive ?

9. Data 131 Holstein Calves (males=16, females=115) affected by NCD
hospitalized from January 2006 until August 2014
Proportion deceased = 34% (45/131)

10. Methods Logistic Regression (LR): Parametric method
Models the probability that the response variable (survive YES/NO) belongs to a particular category

11. Methods Gradient Boosting Machine (GBM) Non parametric method
the learning procedure consecutively fits new models (trees) to provide a more accurate estimate of the response variable

12. Methods

13. Methods Building the predictive model: Split the data in
Training dataset (75% , n=92, incidence = 34%)
Testing dataset (25%, n=39, incidence = 36 %)
Use training data to build the predictive model using a 5-fold cross-validation scheme

14. Methods
Finally:
Use the obtained model to predict the probability of the disease outcome (dead/survived) in the testing data

15. Statistics
Accuracy (AC) : total number of predictions that were correct
Sensitivity or true positive rate (TPR):
Correctly predicted as survived
Specificity or True Negative Rate (TNR):
Correctly predicted as deceased

16. Results- Training Statistics LR GBM Accuracy 0.84 1.00 TPR 0.83 1.00
TNR
0.65
1.00 it
How
many
times
worked
well
How
may
really
survived
calves it
predicted
How
many
really
deceased
calves it
predicted

17. Results - Testing Statistics LR GBM Accuracy 0.70 0.54 TPR 0.74 0.63
TNR
0.58
0.33 it
How
many
times
worked
well
How
may
really
survived
calves it
predicted
How
many
really
deceased
calves it
predicted

18. Results - Consideration
What happened ????
Overfitting

19. Results - Consideration
GBM did a fantastic job on training data but failed on testing data – why ???
Size of training data
Reduce complexity (# trees)

20. Results – Testing reducing complexity (# trees)
Statistics
GBM
GBMn
Accuracy
0.54
0.62
TPR
0.63
0.69
TNR
0.33
0.46 it
How
many
times
worked
well
How
may
really
survived
calves it
predicted
How
many
really
deceased
calves it
predicted

21. Results - Consideration
Feature importance

22. Conclusions Learning from data !!!
It’s possible to build predictive model for NCD. Accuracy depending on training sample size & method
Total protein and Anion Gap important features

23. Thanks