Standardisation of milk MIR spectra, Development of common MIR equations

(1)

Namur, 16/04/2015

Grelet C

1 _{, Fernandez J.A}

1 _{, Dardenne P}

1 _{, Baeten V}

1 _{, Vanlierde A}

1 _{, Soyeurt H}

2 _{, Darimont C}

1 _{& Dehareng F}

1

_{Walloon Agricultural Research Center (CRA-W), Gembloux, Belgique}

2

_{University of Liège, Gembloux Agro-Bio Tech, Gembloux, Belgique}

c.grelet@cra.wallonie.be

Standardisation of milk MIR spectra,

Development of common MIR equations

0 50 100 150 200 250 300 -0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04

(2)

0 100 200 300 400 500 -0.1 0 0.1 0.2 0.3 0.4 0.5

Spectra after slave standardization

wavelength (cm-1) A 0 100 200 300 400 500 -0.1 0 0.1 0.2 0.3 0.4 0.5

wavelength (cm-1)

A

OptiMir

Develop MIR models

predicting cow state:

 Variability is

needed to build

robust models

 67

(3)

-0.015

-0.01

-0.005

0

0.005

0.01

0.015

0.02 -0.02

-0.015

-0.01

-0.005

0

0.005

0.01

0.015 Scores on PC 4 (0.01%)

S

c

o

re

s

o

n

P

C

5 (

0 .0

1 %

)

Samples/Scores Plot of X212

Variability of instruments

PCA on informative wavenumbers , for 5

common samples analysed on 45

instruments from the same brand

(4)

1000

1200 1400

1600 1800

2000 2200

2400

2600 2800

3000

-0.2

-0.1

0

0.1

0.2

0.3

0.4 Spectra after log

wavelength (cm-1)

A

Machine Y

Machine X

Wavenumber (cm-1)

Ab

sor

b

an

ce

Instruments are different

(5)

0

100

200

300

400

500

600

700

800

900

0

100

200

300

400

500

600

700

800 M

e

su

re

d

g

CH

4

/

d

ay

Prediction by MIR (g CH

4

/ day)

Common milks analysed on 2 instruments

from 2 brands

Methane model

(A.Vanlierde, 2013)

(6)

N= 5

y= 590.7677+0.074968x

R square= 0.0181

biais= -766.7733

RMSE= 778.462

-400

-200

0

200

400

600

800

480

500

520

540

560

580

600

620

640

660

680 CH4 (g/day) predictions without standardisation, Master vs BEL0201

Slave predictions

M

a

s

te

r

p

re

d

ic

ti

o

n

s

Common milks analysed on 2 instruments

from 2 brands

Methane model

Creation and use of common models?

Machine X : 670 g CH4/day

Machine Y : -230 g CH4/day

Predictions Instrument Y (g CH4/j)

Pr

e

d

ictio

n

s

ins

tru

men

t

X

(g CH4/

j)

Not possible to use a

common model on 2

(7)

N= 5

y= -5.90+1.84x

RMSE= 200.53

150

200

250

300

350

400

450

500

550

600

300

350

400

450

500

550

600 CH4 Master vs. CH4 Slave

Slave CH4 predictions (g CH4/day)

M

a

s

te

r

C

H

4 p

re

d

ic

ti

o

n

s

(

g

C

H

4 /d

a

y

)

Common milks analysed on 2 instruments

from the same brand

Methane model

Creation and use of common models?

Machine X : 550 g CH4/day

Machine Z : 290 g CH4/day

Predictions Instrument Z (g CH4/j)

Pr

é

d

ictio

n

s ins

tru

men

t

X

(

g CH4/

j)

Not possible to use a common

model on 2 instruments from

(8)

Instrument X

European Master

3.5

3.6

3.7

3.8

3.9

4

4.1

20 1112

12

20

11

12

15

20 1112

21

20 1112

26

20 1112

30

20 1201

04

20 1201

10

20 1201

13

20 1201

20

20 1201

25

20 1201

31

20 1202

10

20 1202

16

20 1202

21

20 1202

24

20 1203

01

20 1203

05

20 1203

07

20 1203

13

20 1203

16

20

12

03

22

20 1203

27

20 1204

02

20 1204

03

20 1204

05

20 1204

10

20 1204

12

20 1204

13

20 1204

19

20 1204

24

20 1204

30

20 1205

04

20 1205

10

20 1205

15

20 1205

22

20 1205

25

20 1206

01

20 1206

06

20 1206

13

20

12

06

18

20 1206

20

20 1206

22

20 1206

26

20 1206

28

20 1207

03

20 1207

06

20 1207

12 Fat

p

re

d

ic

tion

(g/100m

l)

Fat prediction from raw spectra of an instrument analyzing an UHT milk

Lot 1

Lot 2

Lot 3

Lot 4

6 months

(9)

Instrument X

European Master

Instruments also suffer from big perturbations

 maintenance operation, piece replacement

3.45

3.5

3.55

3.6

3.65

3.7

3.75

3.8

3.85

3.9

3.95

4

20 1112

19

20

11

12

21

20 1112

23

20 1112

29

20 1201

02

20 1201

04

20 1201

09

20 1201

13

20 1201

20

20 1201

26

20 1201

30

20

12

02

01

20 1202

06

20 1202

09

20 1202

14

20 1202

17

20 1202

21

20 1203

01

20 1203

05

20 1203

12

20 1203

13

20 1203

14

20 1203

15

20 1203

19

20

12

03

28

20 1204

03

20 1204

18

20 1205

02

20 1205

08

20 1205

15

20 1205

21

20 1205

30

20 1206

04

20 1206

08

20

12

06

13

20 1206

19

20 1206

21

20 1206

22

20 1206

26

20 1206

29

20 1207

06

20 1207

12 Fat

p

re

d

ic

tion

(g/100m

l)

Fat prediction from raw spectra of an instrument analyzing an UHT milk

Lot 1

Lot 2

Lot 3

6 months

(10)

Issues:

• Not possible to create common

tools

• Not possible to transfer a

model on other instruments

• Instruments not stable in time

 Slope/bias correction not possible for models predicting

methane, fertility, ketosis…

 Direct Standardisation of the spectra

 A model can be used on only 1 instrument and within a limited time

(11)

Brand C

Brand B

Brand A

Variability between

instruments

How it works

(12)

Selected instruments for

Master creation

How it works

Brand C

Brand B

Brand A

(13)

Selected instruments for

Master creation

Master

How it works

Brand C

Brand B

Brand A

(14)

Master

18 Selected instruments

for Master creation

How it works

Brand C

Brand B

(15)

Master

18 Selected instruments

for Master creation

How it works

Brand C

Brand B

(16)

Master

18 Selected instruments

for Master creation

How it works

Brand C

Brand B

(17)

Standardized instruments

Master

Standardised

instruments for

creation of equations

Standardised

instruments for use

of equations

18 Selected instruments

for Master creation

How it works

Brand C

Brand B

(18)

Absorbance in an area

r

1 (master)

 correlated to the absorbance within

R

2 (slaves)

« Master »

« Slave »

r

₁

R

₂

PIECE-WISE DIRECT STANDARDIZATION

(PDS)

r

_1j

=

R

_2j

b

_j

+ b

_0j

(19)

0

1

2

3

4

5

0

2

4

6 Pr

o

te

in

s %

Fat %

Ring test samples composition

27 Labs

67 Apparatus

(20)

PDS

0 100 200 300 400 500 -0.1 0 0.1 0.2 0.3 0.4 0.5

Spectra after log

wavelength (cm-1) A 0 100 200 300 400 500 -0.1 0 0.1 0.2 0.3 0.4 0.5

wavelength (cm-1) A

Machine Y

Master

Machine Y

Master

Results

(21)

-0.015

-0.01

-0.005

0

0.005

0.01

0.015

0.02 -0.02

-0.015

-0.01

-0.005

0

0.005

0.01

0.015 Scores on PC 4 (0.01%)

S

c

o

re

s

o

n

P

C

5 (

0 .0

1 %

)

Samples/Scores Plot of X212

Variability of instruments

PCA on informative wavenumbers , for 5

common samples analysed on 45

instruments from the same brand

(22)

Reduce the variability of instruments

-0.01

-0.005

0

0.005

0.01

0.015

0.02

0.025 -0.02

-0.015

-0.01

-0.005

0

0.005

0.01

0.015 Scores on PC 4 (0.01%)

S

c

o

re

s

o

n

P

C

5 (

0 .0

0 %

)

Samples/Scores Plot of X212

NON

S

TD

STD

MASTER

(23)

N= 5

y= 3.15+0.99x

RMSE= 21.1816

300 350 400 450 500 550 600 300 350 400 450 500 550 600 CH4 Master vs. CH4 Slave

Slave CH4 predictions (g CH4/day)

M a s te r C H 4 p re d ic ti o n s ( g C H 4 /d a y )

N= 5

y= -5.90+1.84x

RMSE= 200.53

150 200 250 300 350 400 450 500 550 600 300 350 400 450 500 550 600 CH4 Master vs. CH4 Slave

Slave CH4 predictions (g CH4/day)

M a s te r C H 4 p re d ic ti o n s ( g C H 4 /d a y )

Predictions slave (g/d/cow)

Pr

é

d

ictio

n

s mas

te

r

(g

/d

/c

ow

)

Common milks analysed on the master and on a slave instrument

from the

same brand

Methane model

Creation and use of common models?

Without standardisation

With standardisaton

Predictions slave (g/d/cow)

Pr

é

d

ictio

n

s mas

te

r

(g

/d

/c

ow

)

Prédiction master :

560 g/d/cow

Prédiction slave : 545 g/d/cow

Possible to use a common

model on 2 instruments from

(24)

N= 5

y= 590.7677+0.074968x

R square= 0.0181

biais= -766.7733

RMSE= 778.462

-400 -200 0 200 400 600 800 480 500 520 540 560 580 600 620 640 660 680

CH4 (g/day) predictions without standardisation, Master vs BEL0201

Slave predictions M a s te r p re d ic ti o n s

N= 5

y= -28.4074+1.0493x

R square= 0.89163

biais= 7.8444e-013

RMSE= 23.091

450 500 550 600 650 700 450 500 550 600 650 700

CH4 (g/day) predictions after standardisation, Master vs BEL0201

Slave predictions M a s te r p re d ic ti o n s

Predictions slave, brand B (g/d/cow)

Pr

é

d

ictio

n

s mas

te

r

(g

/d

/c

ow

)

Common milks analysed on the master and on a slave instrument

from

another brand

Methane model

Creation and use of common models?

Without standardisation

With standardisaton

Predictions slave, brand B (g/d/cow)

Pr

é

d

ictio

n

s mas

te

r

(g

/d

/c

ow

)

Prédiction master :

570 g/d/cow

Prédiction slave : 550 g/d/cow

Possible to use a common

model on 2 instruments from

(25)

47 instruments (7 brand A, 1 brand B, 39 brand C)



Methane

model

RMSE between master and slaves predictions, before and after standardisation

Brand A

Brand B

Brand C

Global average

RMSE before PDS

195.54

778.46

132.15

155.34 RMSE after PDS

35.85

23.09

20.48

22.83

0.00

100.00

200.00

300.00

400.00

500.00

600.00

700.00

800.00

900.00 R

M

SE

of

CH4

(g

/d/

co

w

)

RMSE of CH4 predictions between master and slaves

(26)

47 instruments (7 brand A, 1 brand B, 39 brand C)



Poly-Unsaturated Fatty acids

model

RMSE between master and slaves predictions, before and after standardisation

Brand A

Brand B

Brand C

Global average

RMSE before PDS

0.0749

0.1175

0.0206

0.0308

RMSE after PDS

0.0092

0.0053

0.0032

0.0041

0.0000

0.0200

0.0400

0.0600

0.0800

0.1000

0.1200

0.1400

R

M

SE

PU

FA

(g

/10

0ml

)

RMSE of PUFA predictions between master and slaves

(27)

• 1827 milk samples

(standardized since 2012)

• Brand A

• Brand C

• Large variability of breeds (17), seasons,

feeding systems and geographical areas:

 Belgium and Luxembourg:

brand C instruments

 France : Brand A and

C instruments

 Germany: Brand A

and C instruments

 Ireland: Brand C

instruments

 Finland: Brand C

instruments

 UK: Brand C

instruments

Fatty acids models

Description Mean SECV R²cv RPDcv Use

3.912 0.007 1.00 132.99 +++ C4:0 0.104 0.008 0.93 3.67 + C6:0 0.072 0.006 0.91 3.32 + C8:0 0.047 0.004 0.91 3.29 + C10:0 0.111 0.010 0.91 3.37 + C12:0 0.133 0.012 0.92 3.62 + C14:0 0.445 0.031 0.93 3.88 + C14:1 cis 0.038 0.008 0.68 1.78 -C16:0 1.192 0.093 0.94 4.18 + C16:1 cis 0.066 0.013 0.73 1.91 -C17:0 0.027 0.003 0.80 2.24 0 C18:0 0.393 0.057 0.84 2.51 0 Total of C18:1 trans 0.125 0.026 0.79 2.17 0 C18:1 cis9 0.751 0.062 0.95 4.35 + Total of C18:1 cis 0.808 0.064 0.95 4.58 + Total of C18:2 0.096 0.015 0.69 1.79 -C18:2 cis9, cis12 0.061 0.012 0.72 1.91 -C18:3 cis9, cis12, cis 15 0.020 0.004 0.68 1.77 -C18:2 cis 9, Trans 11 0.028 0.010 0.74 1.95 -Saturated FA 2.689 0.072 0.99 10.22 +++ Mono-unsaturated FA 1.077 0.058 0.97 5.83 ++ Poly-unsaturated FA 0.159 0.021 0.77 2.10 0 Unsaturated FA 1.237 0.065 0.97 5.75 ++ Short chain FA 0.347 0.025 0.93 3.88 + Mid chain FA 1.982 0.105 0.97 5.53 ++ Long chain 1.580 0.110 0.95 4.52 + Bbranched FA : iso + anteiso 0.090 0.012 0.75 2.00 0 Omega 3 0.026 0.006 0.66 1.73 -Omega 6 0.103 0.015 0.72 1.89 -Odd FA 0.155 0.016 0.83 2.41 0 Trans FA 0.160 0.030 0.80 2.26 0 C18:1 0.936 0.061 0.96 5.18 ++

Cross validation with 4 subsets R²cv RPDcv Use 1.00 132.99 +++ 0.93 3.67 + 0.91 3.32 + 0.91 3.29 + 0.91 3.37 + 0.92 3.62 + 0.93 3.88 + 0.68 1.78 -0.94 4.18 + 0.73 1.91 -0.80 2.24 0 0.84 2.51 0 0.79 2.17 0 0.95 4.35 + 0.95 4.58 + 0.69 1.79 -0.72 1.91 -0.68 1.77 -0.74 1.95 -0.99 10.22 +++ 0.97 5.83 ++ 0.77 2.10 0 0.97 5.75 ++ 0.93 3.88 + 0.97 5.53 ++ 0.95 4.52 + 0.75 2.00 0 0.66 1.73 -0.72 1.89 -0.83 2.41 0 0.80 2.26 0 0.96 5.18 ++ Cross validation with 4 subsets