07-10-2014 Detecting oestrus by monitoring sows visits to a boar - - PDF document

07-10-2014 1

Detecting oestrus by monitoring sows’ visits to a boar

• T. Ostersen, C. Cornou, A.R. Kristensen, 2010

Presented by Dan Jensen

Introduction

• Problem:
• 1. Not all inseminations are successful

(5 – 25 %)

• 2. More group housing of pregnant

sows – since 2003

• Goal: Automatic oestrus detection
• Time/money saver!
• Modern standard: Manually applied

pressure-test

• Literature says: visit to boar

Advanced Quantitative Methods in Herd Management – 25/09 2013 Dias 2

Experimental design

Set-up: Sow detection/identification: RFID tags

Advanced Quantitative Methods in Herd Management – 25/09 2013 Dias 3

07-10-2014 2

Experimental design

Learning data:

• 41 sows in controlled environments,

all in post-weaning oestrus:

• 1. 5 sows, 2005
• 2. 12 sows, 2007
• 3. 24 sows, 2008
• 17 were inseminated, 24 were not
• Back pressure test 3 times/day
• Continuous measure of boar visits

Advanced Quantitative Methods in Herd Management – 25/09 2013 Dias 4

Experimental design

Test data:

• Collected: October 2004 – June 2009
• 3886 test periods (>= 14 days)
• 111 cases of gestation oestrus

Advanced Quantitative Methods in Herd Management – 25/09 2013 Dias 5

Quick data structure overview

From this, models must account for:

1. Individual patterns 2. Outliers 3. Diurnal effect

Advanced Quantitative Methods in Herd Management – 25/09 2013 Dias 6

07-10-2014 3

Indicator 1: Visit duration

Dynamic linear model (DLM): Yt = log(seconds/visit) Error terms, vt and wt , are considered independant and normally distributed

Advanced Quantitative Methods in Herd Management – 25/09 2013 Dias 7

Observation equation System equation

Indicator 1: Visit duration

Advanced Quantitative Methods in Herd Management – 25/09 2013 Dias 8

Four states – four sets of DLM: State change determined by:

• Observation
• Markov Prior

probability

• Multi process, class 2

Normal model estimated for each sow

Indicator 1: Visit duration

Advanced Quantitative Methods in Herd Management – 25/09 2013 Dias 9

Factors optimized in Learning data: Method:

• Prior probabilities ( )
• Time threshold (∆)
• Prior point estimate
• f variance (S0)
• Maximize oestrus SE and SP

per 24 hours (-72 h to +48 h)

• Alarm if p(Oestrus) > 0.8

07-10-2014 4

Indicator 1: Visit duration

Advanced Quantitative Methods in Herd Management – 25/09 2013 Dias 10

Optimization results

Per day! SE = 0.96 per pig!

5 minute break

Advanced Quantitative Methods in Herd Management – 25/09 2013 Dias 11

Indicator 2: Visit frequency

Advanced Quantitative Methods in Herd Management – 25/09 2013 Dias 12

High Low Low

07-10-2014 5

Indicator 2: Visit frequency

Advanced Quantitative Methods in Herd Management – 25/09 2013 Dias 13

Data is Poisson distributed:

• Dynamic Generalised Linear Model!

Yt = Nvisits / 6 hrs W = system variance

Probability function Compromise – variability vs. responstime

Indicator 2: Visit frequency

Advanced Quantitative Methods in Herd Management – 25/09 2013 Dias 14

Parameter estimation, using learning data: Method: Minimize squared forecast errors during normal conditions Result:

Indicator 2: Visit frequency

Advanced Quantitative Methods in Herd Management – 25/09 2013 Dias 15

The point: we can make a forecast!

Optimized limit: OI > 3.3

07-10-2014 6

A combined model

Advanced Quantitative Methods in Herd Management – 25/09 2013 Dias 16

The two models are combined using Bayes theorem:

P(oestrus) = result of duration model P(+|oestrus) = sensitivity of duration model P(-|oestrus) = 1-sensitivity of duration model Threshold: 0.95

Results

Advanced Quantitative Methods in Herd Management – 25/09 2013 Dias 17

(LR = Likelihood Ratio)

Conclusions, 1

Advanced Quantitative Methods in Herd Management – 25/09 2013 Dias 18

Best detection model: Model is good (better) for:

• 1. Duration by the boar
• 1. Oestrus detection in gestation station!

07-10-2014 7

Conclusions, 2

Advanced Quantitative Methods in Herd Management – 25/09 2013 Dias 19

Advances over others: Shortcomings(?): Improvement perspectives:

• 1. Better specificity
• 2. Better response time (1 – 6 hours vs. 1 day)
• 1. Lower sensitivity

(e.g. 93 % by Bressers et al.)

• 1. Include more information:
• 1. Feeding rank
• 2. Activity
• 3. Pregnancy test