thomas p. oscar, ph.d. usda, ars princess anne, md
DESCRIPTION
General Regression Neural Network Model for Growth of Salmonella Serotypes on Chicken Skin for Use in Risk Assessment. Thomas P. Oscar, Ph.D. USDA, ARS Princess Anne, MD. Risk Assessment Data Gaps. Strain variation. Microbial competition. Initial dose. Food matrix. - PowerPoint PPT PresentationTRANSCRIPT
General Regression Neural Network General Regression Neural Network Model for Growth of Model for Growth of SalmonellaSalmonella
Serotypes on Chicken Skin for Use in Serotypes on Chicken Skin for Use in Risk AssessmentRisk Assessment
Thomas P. Oscar, Ph.D.Thomas P. Oscar, Ph.D.USDA, ARSUSDA, ARS
Princess Anne, MDPrincess Anne, MD
Risk AssessmentRisk AssessmentData GapsData Gaps
• Strain variationStrain variation • Microbial competitionMicrobial competition
• Initial doseInitial dose • Food matrix Food matrix
Strain VariationStrain Variation
• Salmonella entericaSalmonella enterica serotypes (> 2,300) serotypes (> 2,300)
– Top three in chickens are:Top three in chickens are:
Enteritidis Typhimurium Kentucky
Must have been that
chicken!
Strain VariationStrain Variation
Autoclaved chicken meat at 25C
J. Food Safety (2000) 20:225-236.
Microbial CompetitionMicrobial Competition
J. Food Prot. (2003) 66(2):200-207; (2006) 69(2):276-281.
Microbial CompetitionMicrobial Competition
• Natural Antibiotic ResistanceNatural Antibiotic Resistance
– Bad for public healthBad for public health
– Good for predictive microbiologyGood for predictive microbiology
Salmonella Typhimurium DT104
J. Food Prot. (2006) 69(9):2048-2057.
J. Food Prot. (2008) 71(6):1135-1144.
J. Food Prot. (2009) 72(2):304-314.
Initial DoseInitial Dose
Food Microbiol. (2007) 24:640-651.
Secondary Models
PrimaryModel
PrimaryModel
Nmax
Model
Model
PIModel
Model
Observed Predicted
Observed PI Predicted PI
Observed Predicted
Observed Nmax Predicted Nmax
PredictedN(t)
ObservedN(t)
TertiaryModel
PredictedN(t)
Regression ModelingRegression Modeling
J. Food Prot. (2005) 68(12):2606-2613.
Neural Network ModelingNeural Network Modeling
• General Regression Neural Network (GRNN)General Regression Neural Network (GRNN)
– Better performance than regression modelsBetter performance than regression models
– User-friendly commercial softwareUser-friendly commercial software
• Compatible with Monte Carlo simulation softwareCompatible with Monte Carlo simulation software
Jeyamkondan et. al., 2001
ObjectiveObjective
• To develop a GRNN and simulation model for To develop a GRNN and simulation model for growth of growth of SalmonellaSalmonella on chicken skin as a on chicken skin as a function of serotype for use in risk assessment.function of serotype for use in risk assessment.
– Short-term temperature abuse (0 to 8 h)Short-term temperature abuse (0 to 8 h)
Materials and MethodsMaterials and Methods
• Experimental Design (3 x 10 x 5 x 2 x 2)Experimental Design (3 x 10 x 5 x 2 x 2)
– Serotypes (Typhimurium, Kentucky, Hadar)Serotypes (Typhimurium, Kentucky, Hadar)
• 30C in BHIB for 23 h at 150 opm } prehistory30C in BHIB for 23 h at 150 opm } prehistory
– Temperature (5, 10, 15, 20, 25, 30, 35, 40, 45, 50C)Temperature (5, 10, 15, 20, 25, 30, 35, 40, 45, 50C)
– Time (0, 2, 4, 6, 8 h)Time (0, 2, 4, 6, 8 h)
– Trial (1, 2)Trial (1, 2)
– Sample (a, b)Sample (a, b)
7 cfu
5 l
Materials and MethodsMaterials and Methods
D) Typhimurium; 35 C
0 2 4 6 80
2
4
6
8PredictedObserved
Time (h)
Lo
g n
um
ber
MPN
CFU
Materials and MethodsMaterials and Methods
• Plating MediaPlating Media
– XLH-CATS for TyphimuriumXLH-CATS for Typhimurium
– XLH-NATS for KentuckyXLH-NATS for Kentucky
– XLH-TUGS for HadarXLH-TUGS for Hadar
MPN drop plate
XL = xylose lysineH = HEPESC = chloramphenicolA = ampicillinT = tetracycline
S = streptomycinN = novobiocinU = sulfisoxazoleG = gentamicin
Poultry isolates
Ingredients
General Regression Neural NetworkGeneral Regression Neural Network
T t
… …-0.01 7.37
N(x) D(x)
ŷ
Input Layer
Pattern Layer
Summation Layer
Output
S
Specht, 1991 Serotype Temp. time
Distance Function
Predicted Value
Step 1Enter data
Step 2Define the data set
Step 3Set the training parameters
Step 4Train the GRNN
Step 5Review results
5 10 15 20 25 30 35 40 45 50
-2
-1
0
1
2
Hadar
TyphimuriumKentucky
A) Train (n = 464)
0 2 4 6
80 2 4 6
80 2 4 6
80 2 4 6
80 2 4 6
80 2 4 6
80 2 4 6
80 2 4 6
80 2 4 6
80 2 4 6
8
Ch
Independent variables
Res
idual
(log
)
J. Food Prot. (2006) 69(9):2048-2057.
5 10 15 20 25 30 35 40 45 50
-2
-1
0
1
2
Hadar
B) Test (n = 116)
TyphimuriumKentucky
0 2 4 680 2 4 6
80 2 4 6
80 2 4 6
80 2 4 6
80 2 4 6
80 2 4 6
80 2 4 6
80 2 4 6
80 2 4 6
8
Ch
Independent variables
Res
idual
(log
)
J. Food Prot. (2006) 69(9):2048-2057.
5 10 15 20 25 30 35 40 45 500
1
2
3
4
5
6
TyphimuriumKentuckyHadar
B) 5.3 h
Temperature (C)
Salmonella
(log)
0 1 2 3 4 5 6 7 80
1
2
3
4
5
6
TyphimuriumKentuckyHadar
A) 37C
Time (h)
Salmonella
(log)
Step 6Predict
D) Typhimurium; 35 C
0 2 4 6 80
2
4
6
8PredictedObserved
Time (h)
Log
num
ber
E) Kentucky; 35 C
0 2 4 6 80
2
4
6
8PredictedObserved
Time (h)Lo
g nu
mbe
r
F) Hadar; 35 C
0 2 4 6 80
2
4
6
8PredictedObserved
Time (h)
Log
num
ber
G) Typhimurium; 45 C
0 2 4 6 80
2
4
6
8PredictedObserved
Time (h)
Log
num
ber
H) Kentucky; 45 C
0 2 4 6 80
2
4
6
8PredictedObserved
Time (h)
Log
num
ber
I) Hadar; 45 C
0 2 4 6 80
2
4
6
8PredictedObserved
Time (h)Lo
g nu
mbe
r
Step 7Integrate with risk assessment
-1 0 1 2 3 4 50
2
4
6
8
10
Output dataDistribution fit
B) RiskLogistic(0.047094,0.056875)
Log change
Freq
uen
cy
-1 0 1 2 3 4 50
1
2
3
4A) RiskPearson5(4.4594,1.5797,RiskShift(-0.26825))
Output dataDistribution fit
Log change
Freq
uen
cy
Serotype Serotype (%)(%)
Temperature Temperature ((C)C)
Time Time (h)(h)
Log changeLog change
ScenarioScenario T_K_HT_K_H Min_ML_MaxMin_ML_Max Min_ML_MaxMin_ML_Max CorrelationCorrelation Min_50%_MaxMin_50%_Max
AA 31_58_1131_58_11 5_20_505_20_50 0_2_80_2_8 00 -0.21_0.09_4.8-0.21_0.09_4.8
BB 31_58_1131_58_11 5_20_505_20_50 0_2_80_2_8 -1-1 -0.16_0.04_0.5-0.16_0.04_0.5
Conclusion #1Conclusion #1
• Easy to developEasy to develop
• Low cost Low cost
• Flexible predictionsFlexible predictions
• Superior performanceSuperior performance
Neural network modeling outperforms regression modeling in predictive
microbiology applications
Conclusion #2Conclusion #2
• Cocktail of Typhimurium_Kentucky_HadarCocktail of Typhimurium_Kentucky_Hadar
– Overly ‘fail-safe’ predictions for Kentucky.Overly ‘fail-safe’ predictions for Kentucky.
Conclusion #3Conclusion #3
GRNN model was successfully validated for risk GRNN model was successfully validated for risk assessmentassessment
model?
-1 0 1 2 3 4 50
1
2
3
4A) RiskPearson5(4.4594,1.5797,RiskShift(-0.26825))
Output dataDistribution fit
Log change
Fre
quen
cy
Data Gaps
Strain variation
Microbial competition
Initial dose
Food matrix
AcknowledgementsAcknowledgements
• Thank you for your attention!Thank you for your attention!
• Thanks to Thanks to Jaci LudwigJaci Ludwig of ARS and of ARS and Celia Celia WhyteWhyte and and Olabimpe OlojoOlabimpe Olojo of UMES for of UMES for their outstanding technical assistance on their outstanding technical assistance on this project.this project.
I hope it was
Kentucky!