**Food adulteration is at the top of the list when it comes to food safety concerns, especially following recent incidents, such as the 2008 Chinese powdered milk scandal. That scandal involved milk and infant formula along with other food materials and components being adulterated with melamine. There were an estimated 300,000 victims in total, including six infant deaths, resulting from kidney stones and other kidney damage.**

Protein content is a key parameter that measures the quality of food products such as milk powder and infant formula. Protein concentration is traditionally measured by Kjeldahl and Dumas methods through measuring nitrogen content; however, neither method can distinguish non-protein nitrogen from naturally occurring nitrogen in protein. This gap in detectability allowed the protein content in food to be falsified by adding nitrogen-rich chemicals, such as melamine. Melamine adulteration not only causes protein deficiency, but also kidney stones and renal failure when it reacts with cyanuric acid inside the body.

In food products other than infant formula, the FDA concludes that levels of melamine and melamine-related compounds below 2.5 parts per million (ppm) do not raise concerns. They also conclude that melamine or cyanuric acid alone, at or below 1 part per million in infant formula do not raise public health concerns in babies^{1}. To detect melamine at such low concentrations, highly sensitive techniques such as LC/MS have been developed; however, it is very time- consuming and does not accommodate rapid screening. Near infrared (NIR) technique, on the other hand, is a non- destructive, fast screening method used for illicit drug screening, raw material inspection, etc.; however, the conventional NIR method is not very sensitive for concentrations less than 1%. To increase the sensitivity of the NIR method, a patent-pending algorithm has been developed for targeted screening at concentrations as low as 0.01%. In this study, the Advanced-ID™ algorithm with an FT-NIR spectrometer will be evaluated for low concentration melamine screening in milk powders.

## Materials and Methods

### Materials

Four milk powder samples were purchased from local supermarkets: Market Basket instant nonfat dry milk powder, Carnation instant nonfat dry milk powder, Enfamil infant step 1 powder, and Enfamil newborn powder. Melamine (purity = 99%) was purchased from Aldrich (St. Louis, MO) and added to the milk powder samples in various low concentrations as listed in Table 1.

**Table 1. **Melamine-Adulterated sample concentrations

Sample ID |
100% Melamine % |

Carnation Mix 1 |
0.099 |

Carnation Mix 2 |
0.191 |

Carnation Mix 3 |
0.495 |

Carnation Mix 4 |
0.716 |

Carnation Mix 5 |
1.020 |

Carnation Mix 6 |
1.479 |

Carnation Mix 7 |
1.964 |

Carnation Mix 8 |
0.355 |

MB Mix 1 |
0.139 |

MB Mix 2 |
0.267 |

MB Mix 3 |
0.407 |

MB Mix 4 |
0.600 |

MB Mix 5 |
0.919 |

MB Mix 6 |
1.457 |

MB Mix 7 |
5.384 |

Enfamil Infant Mix 1 |
0.197 |

Enfamil Infant Mix 2 |
0.077 |

Enfamil Infant Mix 3 |
0.436 |

Enfamil Infant Mix 4 |
1.333 |

Enfamil NB Mix 1 |
0.197 |

Enfamil NB Mix 2 |
1.463 |

Enfamil NB Mix 3 |
0.231 |

Enfamil NB Mix 4 |
0.782 |

### Sample Measurement

NIR spectra were collected using a QuasIR™ 3000 (Galaxy Scientific, Nashua, NH, USA ) equipped with a sample spinner accessory. Samples were placed into a 98 -mm cup with low OH quartz window and then loaded onto the sample spinner, which is mounted off-center on the 23-mm sample window of the integrating sphere. This increases the quantity of sample measured and is suitable for inhomogeneous samples. Each sample was measured five times with 4 cm^{-1} resolution and 50 scans. Samples were reloaded between measurements.

### Data Processing:

Spectral Sage™ software was used for data collection and the CLS-based Advanced-ID™ algorithm and software were used for analysis .

## Result and discussion

For a sample comprising n components, its spectrum S can be modeled as the sum of the spectra of n components K_{1}...K_{n}, assuming the Beer–Lambert law is obeyed.

Where K is the matrix of reference spectra of the sample components, c_{1}...c_{n} are known coefficients and R is a residual, or error. The least squares solution to this equation for the coefficients can be found by standard matrix algebra, otherwise known as Classical Least Squares (CLS), or K-matrix regression.

If each spectrum contains m data points, then we can write this in matrix notation as :

S = K * c + R

where S and R are mxl matrices, K is a mxn matrix of reference spectra, and c is an nxl matrix of coefficients. Often, all of the components represented in K are known to be present and the objective of the regression is to find the coefficients c that can then be used to calculate their relative concentrations. In certain cases, however, one of the components may be an unknown that needs to be identified, or a suspected component whose presence in the mixture needs to be confirmed. If we designate this component as a target component and the spectrum of this component as T (target spectrum), then for convenience we can rewrite the equation as:

S = T * c_{0} + K’*c’ + R

where S, T, and R are mxl matrices, K’ is an mx(n-1) matrix of reference spectra of known components that does not contain the spectrum in T, c’ is an (n-1)xl matrix of coefficients, and c_{0} is a scalar coefficient.

Various methods can be used to judge the quality of the model, which includes the common practice of examining the size of the residuals, R. However, if the contribution of the target components to the spectrum S is very small, then the comparison metrics given above are very poor indicators of the presence of the target components. This is because the regression of only the spectra of the known components K’ will result in a very good fit to the sample spectrum S, resulting in a very small residual (close to zero) and a high correlation coefficient (close to one).

The patent-pending Advanced-ID™ algorithm finds a new approach to resolve this issue. It first calculates an approximation to the target spectrum by performing a regression that includes the target and known spectra (S = T * c_{0} + K’*c’ + R), and then calculates a residual with the coefficient for the target spectrum c0 set to 0, thus defining the extracted spectrum E:

E = S - K’*c’

This can be compared with the expression for the residual R:

R = S - Tc_{0} - K’*c’

The residual R will be small if either the target component is not present and K’*c’ is a good approximation to S, or if the target component is present and T*c_{0} + K’c’ is a good approximation to S. As noted above, this is not a good indicator of the presence of the target component. The extracted spectrum E will also be small and will resemble R if the target component is not present and K’*c’ is a good approximation to S. However, if the target component is present in the sample at any significant concentration and T*c_{0} + K’c’ is a good approximation to S, then the extracted spectrum will resemble the spectrum of the target component. Additionally, if the target component is not present and K’*c’ is not a good approximation to S because another component is present that was not included in the regression, then the extracted spectrum will not resemble either R or the target spectrum.

Comparison of the extracted and target spectrum, typically scaled by the regression coefficient c_{0}, can, therefore, be a reliable indicator of the target’s presence. Furthermore, the comparison could be visual by overlaying the two spectra on the computer screen.

As long as the spectra of all components present are included in the regression, the method described above will also work if the sample contains more than one suspected component that needs to be confirmed. In this case, one of the target spectra is T, all the other target spectra are included in K’ and the extracted spectrum is calculated and compared with the target spectrum. This is then repeated for each of the other target spectra. The method described above may also work with more than one unknown component, especially if the principal spectral features of the unknowns are in different spectral regions. In this fashion, individual components in a mixture may be identified.

In this study, the target spectrum is the 99% melamine sample. The extracted spectrum is expressed as

E = S_{AMP} –K_{MK}’*c’

where S_{AMP} is the spectrum of adulterated milk powder, K_{MK} are the spectra of milk powder components. The Advanced-ID package was used to solve the unknown coefficients and calculate the correlation coefficient between the extracted spectrum and melamine reference spectrum.

The average NIR spectra of pure milk powder and infant formula used in this study and the spectrum of 99% melamine are shown in Figure 1. Different than milk powders, melamine has very sharp peaks in the region of 6500–7000 cm^{- 1}; this difference was used to develop the Advanced-ID method algorithm.

**Figure 1. **Spectra of pure components

As for Enfamil Infant step 1 formula, melamine was added in four concentrations: 0.077%, 0.197%, 0.436%, and 1.333%. By visually checking the NIR spectra of adulterated samples an elevated peak can be noticed around 6800 cm^{-1} (Figure 2) only when melamine concentration is higher than around 0.5%.

**Figure 2. **Original spectra of Enfamil Infant 1 Mixture

Figure 3 presents the extracted spectra of the mixtures using the Advanced-ID™ method. As can be seen, extracted spectra of the four mixtures have apparent absorbance peaks around 6800 cm^{-1} where the Enfamil Infant step 1 reference sample shows none.

**Figure 3. **Extract spectra of Enfamil Infant 1 Mixture

If we overlay the extract spectrum of Enfamil Infant step 1 mixture 2, which had the lowest melamine concentration, with its original spectra and that for pure melamine, no melamine is detected in the original spectra, whereas in the extract spectrum we see the absorbance matching the melamine sample in the designated region (Figure 4).

**Figure 4. **Overlay of melamine spectrum, original and extract spectra of Enfamil infant mixture 2

Table 2 lists the correlation coefficients of extract spectra to the melamine reference spectrum at respective concentrations. The extract spectrum of pure Enfamil infant step 1 powder shows no correlation with the melamine spectrum, whereas extract spectra of adulterated samples, with melamine levels as low as about 0.1%, have a high correlation with the melamine spectrum, with a correlation coefficient higher than 0.97. The reported NIR melamine detection limit is 2–5%; therefore, the Advanced-ID™ method increases the NIR method’s sensitivity.

**Table 2. **Advanced-ID result of Enfamil infant mixture

Sample ID |
Melamine ( %) |
Ave. Corr. Cofficient |

Enfamil Infant Mix 1 |
0.197 |
0.9961 |

Enfamil Infant Mix 2 |
0.077 |
0.9726 |

Enfamil Infant Mix 3 |
0.436 |
0.9968 |

Enfamil Infant Mix 4 |
1.333 |
0.9979 |

Enfamil Infant Pure |
0.000 |
-0.1392 |

Similar results were obtained with the rest of the mixtures as shown in Table 3. All extract spectra of melamine- adulterated samples (down to 0.1%) had a higher than 0.97 correlation coefficient with the pure melamine reference spectrum (with the exception of Market Basket instant nonfat dry milk powder mixture 1, which had a melamine concentration of 0.139% and correlation coefficient value of 0.881). This could be due to the fact that the milk powder itself is freeze-dried, which leaves a porous structure and various particle sizes that are more inhomogeneous compared with the rest. Grinding the sample before measurement could possibly increase the accuracy of the result.

**Table 3. **Advanced-ID™ result of milk powder and Enfamil New Born mixtures

Sample ID |
Melamine ( %) |
Ave. Corr. Cofficient |

MB Mix 1 |
0.139 |
0.881 |

MB Mix 2 |
0.267 |
0.9698 |

MB Mix 3 |
0.407 |
0.9883 |

MB Mix 4 |
0.600 |
0.9932 |

MB Mix 5 |
0.919 |
0.9952 |

MB Mix 6 |
1.457 |
0.9969 |

MB Mix 7 |
5.384 |
0.9983 |

MB powder pure |
0.000 |
-0.1454 |

Melamine Mix 1 |
0.099 |
0.9755 |

Melamine Mix 2 |
0.191 |
0.9834 |

Melamine Mix 3 |
0.495 |
0.9976 |

Melamine Mix 4 |
0.716 |
0.9978 |

Melamine Mix 5 |
1.020 |
0.9987 |

Melamine Mix 6 |
1.479 |
0.9988 |

Melamine Mix 7 |
1.964 |
0.9981 |

Melamine Mix 8 |
0.355 |
0.9964 |

Carnation Pure |
0.000 |
-0.1448 |

Enfamil NB Mix 1 |
0.197 |
0.9909 |

Enfamil NB Mix 2 |
1.463 |
0.9977 |

Enfamil NB Mix 3 |
0.231 |
0.9824 |

Enfamil NB Mix 4 |
0.782 |
0.997 |

Enfamil NB Pure |
0.000 |
-0.1512 |

If we know other possible chemicals that can be added to milk powder to increase the nitrogen reading, then we can measure the spectra of those pure chemicals and use them as targets and identify them in a similar way. We can also use the Advanced-ID™ program to study the residual spectrum at the same time and, if there is any unknown adulteration present, then at least we can see a high residual value and be alerted.

## Conclusion

Using our CLS-based Advanced-ID™ algorithm, we can detect melamine adulteration at concentrations down to 0.1%. Even though this is still not at the 2.5 ppm level, it eliminates the economic motivation to adulterate milk powder with such low concentrations since 0.25% melamine added to milk powder only increases protein content by 1%. Therefore, with a detection limit of 0.1% melamine, NIR can be used as a fast screening tool for milk powder melamine adulteration.

## References

1 "Limits set on melamine levels". People's Daily. 9 October 2008

This information has been sourced, reviewed and adapted from materials provided by Galaxy Scientific Inc.

For more information on this source, please visit Galaxy Scientific Inc.