Bayesian network � 3 Case examples Advanced Herd Management October 2nd 2009 Tina Birk Jensen Outline 1. The ”Weak Sow Index” model An example of a Bayesian network developed based on collected data 2. The ”PigLame” model An example of an Object+Oriented Bayesian network developed based on expert opinions The ”Weak Sow Index” model An example of a Bayesian network developed based on collected data 1
Involuntary culling of sows � Danish sow herds: ~ 15 % of sows sent to rendering plants � A problem for the animal welfare and the economy Involuntary culling: Sows sent to slaughter due to a poor health status Sows being euthanized Sows experiencing sudden death Involuntary culling of sows � The health status of individual sows is often characterized based on presence or absence of individual clinical signs � There is a need to develop a way to combine information about several diseases The ”Weak Sow Index” model Purpose: Develop a model that characterizes the risk of involuntary culling of individual pregnant sows Weak Sow Index Weak Sow Index (WSI): Probability of a sow to be involuntarily culled 2
Data used for the WSI � 33 sow herds stratified by the feeding system � Each herd visited twice: 49+76 pregnant sows randomly selected each time � Clinical examinations of individual sows The clinical protocol Clinical signs States Lameness �������������� Pressure mark of knee ������ Pressure mark of digit ������ Pressure mark of hock ������ Claw lesion ������ Claw length ��������������� ��������������� ����� Leg position ��������������� ������������� ��������� �������� ����� Reaction ��������������� ����� Shoulder ulcer �� ������������� ������ ������!����� Wounds at rear ���"������ Wounds at head ���"������ Wounds at shoulder ���"������ Vulva bite ������ Filthiness ���#$%&$'��(�����)�� �*&$'��(�����)�� Body condition score ����� ��������+�������+)��� ������� Willingness to stand ������ Data used for the WSI � Farmers recorded all replacements (e.g. euthanasia and sudden death) and reasons for these actions � A total of 2875 sows included in the study � During a 3 month period: 119/2875 (4.1 %) sows involuntarily culled 3
WSI: A Bayesian network model Two steps in the modeling building process Step 1: Structural dependencies between nodes Step 2: Estimate the probabilities for the model Step 1: Structural dependencies PM of PM of Head PM of Shoulder Rear digit knee hock wound wound Reaction wound Lameness Willing to stand Filthiness Vulva Leg bite position Claw BSC length Shoulder Claw ulcer lesion Involuntary culling Step 1: Structural dependencies Using the 16 clinical variables to characterize the underlying correlation structure � We did that using statistical methods (Factor analysis) 4
Step 1: Structural dependencies PM of PM of Head PM of Shoulder Rear knee hock wound digit wound Reaction wound Lameness Willing to stand Filthiness Vulva Leg bite position Claw BSC length Shoulder Claw ulcer lesion Involuntary culling Step 1: Structural dependencies PM of PM of Head PM of Shoulder Rear digit knee hock wound wound Reaction wound Lameness Willing to stand Filthiness Vulva Vulva Leg Leg bite bite position position Pressure Wounds Lameness Claw Claw BSC BSC mark length length Shoulder Shoulder Claw Claw ulcer ulcer lesion lesion Involuntary culling Step 2: Estimate the probabilities for the model Link between the variables and the factors: Factor loadings from the clinical variables 5
Step 2: Estimate the probabilities for the model Link between factors and Involuntary culling: Parameter estimates from logistic regression model Step 2: Estimate the probabilities for the model Factor: Lameness (p=0.01) Weak Sow Index model: Prototype Lameness Willing to stand Lameness Involuntary culling 6
Weak Sow Index model: Prototype Willing Lameness to stand Vulva Vulva bite bite Lameness Now, lets see how it looks Involuntary culling like! Comments! � Bayesian network allows the WSI to be presented even though some of the clinical variables are missing � The WSI model is based solely on collected data � Other variables may be important for involuntary culling � Possible to include expert information in the WSI model The ”PigLame” model An example of an Object+Oriented Bayesian network developed based on expert opinions 7
Outline 1. The case problem 2. Modeling methods in general 3. Qualitative structure of the model 4. Elicitation of probabilities 5. Use of the model (demonstration) Leg disorders in finishers Leg disorders: Any lesion or dysfunction of the leg or claw that might give rise to lameness Lameness: Deterioration in the gait and/or posture The case problem Leg disorders in finishers An economical problem for farmers due to: � Increased work load � Cost of treatments � Reduced productivity � Risk of condemnations � A negative impact on animal welfare The case problem 8
Causes of leg disorders 1. Infectious Mycoplasma hyosynoviae, Erysipelothrix rhusiopathiae, Haemophilus parasuis, Streptococcus suis <2% 2. Physical Fracture, lesion to the claw wall, lesion to the claw sole <1% 50�80% 3. Inherited Osteochondrosis manifesta, osteochondrosis dissecans 70% 30% The case problem Control strategies Control strategies against leg disorders will depend on the cause category • Infectious leg disorders Antibiotics Reconstructing the pen • Physical leg disorders • Inherited leg disorders Boar semen Weight gain The case problem Useful information Herd level • Herd size (number of pigs delivered) • Stocking density (high/low) • Floor type in pens (slatted/concrete) • Supply of straw in pens (deep/sparse/no) • Purchase policy (own piglets/1/>1) • Production type (sectioned/continuous) The case problem 9
Useful information Pig level • Observe pigs from outside the pen Cheap Cheap • Clinical investigation • Bacteriological investigation Expensive Expensive • Pathological investigation The case problem To make a herd diagnosis of leg disorders Challenges • What information to use • How much information to use • How to collect the information The case problem The ”PigLame” model Purpose of the model To estimate probability distributions of different manageable causes of leg disorders in finisher herds The case problem 10
Herd Pig Info Info Leg disorder Strategy Strategy Strategy 1 2 3 The case problem Qualitative structure of the model Characteristics � Based on information from the literature � All nodes are discrete � Each cause�category defined as a risk index I on an arbitrary scale from 0 to 9 Qualitative structure Qualitative structure of the model Characteristics Object oriented structure � Ease the specification of the Bayesian network � Hierarchical structure Two classes: Herd class and pig class Qualitative structure 11
Object�oriented Bayesian network The Pig class Individual pig information Gender Lean meat percentage Diagnostic test results Leg disorder Qualitative structure Object�oriented Bayesian network The Herd class Herd size Production Purchase Pen density Floor type Straw Cause category Pig object Pig object Pig object Feed strat Breed Pen Herd Produc+ Pur+ Floor Straw Gain Den Size tion chase Physical Infectious Inherited Herd class 12
Feed strat Breed Pen Herd Produc+ Pur+ Floor Straw Gain Den Size tion chase Physical Infectious Inherited Herd class Physical Infectious Inherited Gen der LMP Frac Claw Claw Hae Myco Strep Erysi OCM OCD ture Wall Sole mo Pig Lame Obs C3 P3 C4 P4 B1 C5 P5 B2 C6 P6 B3 C7 P7 B4 C8 P8 C9 P9 C1 P1 C2 P2 Lame Pig class Herd class Floor Pen Den Gain …… Physical Infectios Inherited Claw Frac Claw Myco Strep Erysi Haemo OCM OCD ture Wall Sole Pig class Qualitative structure 13
Herd class Floor Pen Den Gain …… Physical Infectios Inherited Frac Claw Claw Myco Strep Erysi Haemo OCM OCD Sole ture Wall Pig class Qualitative structure Herd class Floor Pen Den Gain …… Physical Infectios Inherited Frac Claw Claw Myco Strep Erysi Haemo OCM OCD Sole ture Wall Pig class Qualitative structure Elicitation of probabilities The probabilities in the model are based on 1. Results from published literature Conversion of odds ratios to conditional probabilities <40 conditional probabilities 2. Expert opinions (9 experts) >150 conditional probabilities Not randomly distributed Average of individual elicitations Elicitation of probabilities 14
Recommend
More recommend
