Evaluation of High Pressure Processing Kinetic Models for Microbial Inactivation Using Standard Statistical Tools and Information Theory Criteria, and the Development of Generic Time-Pressure Functions for Process Design Serment-Moreno, V., Fuentes, C., Barbosa-Cánovas, G., Torres, J. A., & Welti-Chanes, J. (2015). Evaluation of High Pressure Processing Kinetic Models for Microbial Inactivation Using Standard Statistical Tools and Information Theory Criteria, and the Development of Generic Time-Pressure Functions for Process Design. Food and Bioprocess Technology, 8(6), 1244-1257. doi:10.1007/s11947-015-1488-x 10.1007/s11947-015-1488-x Springer Accepted Manuscript http://cdss.library.oregonstate.edu/sa-termsofuse 1 Evaluation of high pressure processing kinetic models for microbial 2 inactivation using standard statistical tools and information theory 3 criteria, and the development of generic time-pressure functions for 4 process design 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 Author: Vinicio Serment-Moreno Affiliation: Escuela de Ingenierías y Ciencias, Tecnológico de Monterrey, Av. Eugenio Garza Sada 2501 Sur, Col. Tecnológico, 64849, Monterrey, NL, México E-mail address: vsermentm@gmail.com Author: Claudio Fuentes Affiliation: Department of Statistics, Oregon State University, 54 Kidder Hall, Corvallis, OR 97331, USA Phone: 1-541-737-3587 Fax: 1-541-737-3489 E-mail address: fuentesc@stat.oregonstate.edu Author: Gustavo Barbosa-Cánovas Affiliation: Center for Nonthermal Processing of Food, Washington State University, Pullman, WA 99164-6120, U. S. A. Phone: 1-509-335-6188 Fax: 1-509-335-2722 E-mail address: barbosa@wsu.edu Corresponding Author: José Antonio Torres Affiliation: Food Process Engineering Group, Department of Food Science and Technology, Oregon State University, 100 Wiegand Hall, Corvallis, OR 97331, U. S. A. Phone: 1-541-737-4757 Fax: 1-541-737-1877 E-mail address: j_antonio.torres@oregonstate.edu Corresponding Author: Jorge Welti-Chanes Affiliation: Escuela de Ingenierías y Ciencias, Tecnológico de Monterrey, Av. Eugenio Garza Sada 2501 Sur, Col. Tecnológico, 64849, Monterrey, NL, México Phone: +52-81-8358-2000 ext. 4923 Fax: +52-81-8358-4293 E-mail address: jwelti@itesm.mx 1 48 49 Abstract 50 Generic kinetic models for microbial inactivation by high pressure processing (HPP) would accelerate the 51 development of commercial applications. The aim of this work was to develop a generic model obtained 52 by fitting peer-reviewed microbial inactivation data (124 kinetic curves) to first order kinetics (LKM), 53 Weibull (WBLL), and Gompertz (GMPZ) primary and secondary models. Standard statistics (coefficient of 54 determination (R ), variance, residuals plots, experimental vs predicted plots) and information theory 55 criteria (Akaike Information Criteria, AIC; Akaike differences, ∆AICi, Bayesian Information Criteria) 56 determined their goodness of fit. Standard statistics showed no differences between WBLL and GMPZ, 57 whereas information theory criteria identified WBLL as the best model (lowest AICi value, 61.3% of 58 cases). LKM performed poorly according to all statistics (e.g., ∆AICi > 10, 58.1% of cases). The 59 dispersion of model parameters prevented the derivation of a secondary model for the whole dataset, but 60 clear trends and sufficient data (56 kinetic curves) were found to develop one for milk. A secondary WBLL 61 model (b’ = 0.056-2.230, n = 0.758-0.403; 150-600 MPa) was the best alternative (AICi = 183.8). A GMPZ 62 model yielded similar predictions, but registered ∆AICi = 19.3 reflecting its larger number of parameters (p 63 = 8). Selecting datasets with pressure holding times of commercial interest (t ≤ 10 min) yielded different 64 parameter estimates for the generic WBLL model (b’ = 0.079-1.859, n = 1.340-0.557; 300-600 MPa). In 65 conclusion, information theory criteria complemented standard statistics, and the simpler WBLL 66 secondary model (p = 4) provided a product-specific time-pressure function of industrial relevance. 2 67 68 Keywords: High pressure processing (HPP); microbial inactivation kinetics; first order kinetics model; 69 Weibull model; Gompertz model; Akaike Information Criteria (AIC); information theory criteria 70 2 71 1. Introduction 72 1.1 High pressure processing microbial inactivation kinetic models 73 High Pressure Processing (HPP) can deliver safe food products while preserving most of their nutritional 74 and sensorial properties (Mújica-Paz et al. 2011; Heinz and Buckow 2009). Consumer acceptance of 75 HPP food products keeps raising and by 2018 they are expected to represent a US$12,000 million 76 market, while sales of HPP equipment will reach US$600 million (Spinner 2014). Still, the availability of 77 microbial and enzymatic kinetic information remains well behind the standardized databases and 78 calculation methodologies available for traditional food technologies such as thermal processing 79 (Serment-Moreno et al. 2014). A mathematical model predicting microbial inactivation as a function of 80 time (primary model) is an essential tool when designing HPP experiments and industrial processes. 81 Furthermore, the application of a primary model can be extended if a secondary model describing the 82 pressure dependence of the primary model parameters is available too. 83 84 Santillana Farakos and Zwietering (2011) presented a methodology for the prediction of microbial 85 inactivation rates as a function of pressure. Decimal reduction times (D) and z-values were estimated by 86 assuming a first order inactivation rate (Eq. 1) for microbial inactivation by HPP treatments. Under this 87 assumption, plots of microbial survival fraction (log10 N/N0) versus time should generate straight lines with 88 a slope proportional to the inactivation rate constant (k). However, numerous authors report that linearity 89 is rarely observed and that non-linear equations should be employed to model HPP microbial inactivation 90 data (Bermúdez-Aguirre and Barbosa-Cánovas 2011; Campanella and Peleg 2001; Corradini et al. 2009; 91 Wilson et al. 2008; Serment-Moreno et al. 2014). 92 log10 N = −k ⋅ t N0 (1) 93 94 The Weibull distribution used to predict the failure time of mechanical and electrical systems has been 95 applied to describe microbial inactivation. In this application, the cumulative probability of the Weibull 96 distribution determines the time at which microbial cells fail to survive (Eq. 2). The cumulative distribution 97 function of the Weibull distribution has been widely used to describe non-linear microbial inactivation 98 kinetics due to its simplicity, flexibility and semi-empirical nature. The parameter b’ is equivalent to the 99 inactivation rate constant to the -nth power. The exponent n defines the concavity of the curve and the 100 type of microbial population, classified as resistant (downward concavity; n > 1), low-resistant (upward 101 concavity; n < 1), or homogeneous (straight line; n = 1) (Serment-Moreno et al. 2014; van Boekel 2002; 102 see Chapters 2 and 6, Peleg 2006). n log10 N N 1 1 t = −b'⋅t n =− ⋅ ; b' = ; log10 n N0 N0 2.303 b 2.303 ⋅ b (2) 3 103 104 Secondary models for the Weibull parameters include a logistic-exponential expression for parameter b’, 105 indicating that the inactivation rate constant will be close to zero until a critical pressure level is reached 106 (Pc; Eq. 3). If the pressure level is increased beyond Pc, b’ will increase at a rate wP. An exponential 107 decay model has been used to predict pressure effects on the Weibull parameter n (Eq. 4), where the 108 dimensionless parameter d0 is the value of n at pressures P approaching 0, and d1 is the rate at which n 109 exponentially decays (Doona and Feeherry 2007; Peleg et al. 2002; see Chapter 6, Peleg 2006; Doona et 110 al. 2007). Pilavtepe-Çelik et al. (2009) proposed a Bigelow-type model to describe the pressure 111 dependence of parameter b’ (Eq. 5), where Pref is the reference pressure, b’ref is the Weibull rate constant 112 at Pref, and zP is the amount of pressure needed to increase b’ by 10. Alternative Weibull secondary 113 models mainly consist of empirical equations aiming to linearize the effects of pressure on the parameters 114 b’ and n (Carreño et al. 2011; Chen and Hoover 2003b; Pilavtepe-Çelik et al. 2009; Serment-Moreno et 115 al. 2014). 116 b' (P ) = ln{1 + exp[wP ⋅ (P − PC )]} (3) n(P ) = d 0 ⋅ exp(− d1 ⋅ P ) (4) b' (P ) = bref ⋅ 10 P − Pref − zP (5) 117 118 Other alternatives for non-linear HPP microbial inactivation curves include those capable of describing 119 concavity changes such as the log-logistic (Eq. 6) and the modified Gompertz models (Eq. 7) (Chen and 120 Hoover 2003b; Guan et al. 2005; Saucedo-Reyes et al. 2009). However, the log-logistic model has an 121 unfortunate mistake when applying the chain rule in its theoretical derivation that leads into the 122 misinterpretation of its parameters (Serment-Moreno et al. 2014; Serment-Moreno et al. 2015). Another 123 inconvenience of the log-logistic model (Eq. 6) is the need to approximate log10 (t = 0) with an arbitrary 124 value which were chosen as -1 (Cole et al. 1993) or -6 (Chen and Hoover 2003b). More recently, 125 Serment-Moreno et al. (2015) tested log10 (t = 0) from -1 to -9 to evaluate its effect on the non-linear 126 regression, and reported that the log-logistic model parameter estimates varied for different 127 approximation values for log10 (t = 0). 128 log 10 N = N0 A 4Ω ⋅ (t − log 10 t ) 1 + exp A log 10 t 0 ≈ −1 or − 6 − A 4Ω ⋅ (t − log 10 t 0 ) 1 + exp A (6) 129 4 130 The modified Gompertz model has been rarely used to model the microbial inactivation kinetics by HPP 131 treatments even though biological meaning can be assigned to its parameters. Parameter A represents 132 the difference between the lower and upper asymptotes of microbial survival curves (log10 N/N0 vs time), 133 µmax is the maximum inactivation rate, and parameter λ is the inflection point, or the time at which the 134 linear portion of the curve starts. An extensive and recent literature review showed that no secondary 135 model is available to describe the relationship of the log-logistic or Gompertz model parameters with 136 pressure (Serment-Moreno et al. 2014). 137 log10 N m ⋅ exp(1) = A ⋅ exp− exp max ⋅ (l − t ) + 1 N0 A (7) 138 139 1.2 Information theory criteria for model selection 140 Statistical quantities based on squared differences such as the coefficient of determination R , the mean 141 squared error (MSE), and variance estimates ( σˆ ) are widely used for the evaluation of mathematical 142 models. The coefficient of determination R alone is useful to obtain a quick assessment of a model 143 performance but it is an insufficient criterion to discriminate between different models (Breiman 2001; 144 Johnson and Omland 2004; van Boekel 2008). Mathematical functions that yield the highest R , and the 145 2 lowest MSE or σˆ are usually chosen as the model best describing the experimental data (Chen and 146 Hoover 2003a; Guan et al. 2006). However, Spiess and Neusmeyer (2010) reported that R alone is a 147 poor discriminator stating that similar non-linear models displayed differences only after the third or fourth 148 decimal place of the computed R values. The authors also suggested that R is better suited for linear 149 regression, and discourages R to determine the goodness of fit of nonlinear models (Spiess and 150 Neumeyer 2010). Additionally, the sole use of statistical quantities based on the squared differences may 151 be misleading if the model consistently over- or under- estimates the observed values, for which residuals 152 plots and experimental vs plots are recommended to detect consistent model deviations (Ross 1996). 2 2 2 2 2 2 2 2 153 154 On the other hand, information theory criteria measuring the information lost when an empirical probability 155 distribution given by a proposed model is used to approximate the underlying “true” distribution of a given 156 dataset, provide a powerful alternative for model selection (Burnham and Anderson 2002; Stoica and 157 Selén 2004). Despite its widespread application in environmental science and industrial engineering, 158 information theory criteria have been rarely used for model comparisons in food science (Albert and 159 Mafart 2005; Koseki et al. 2009; López et al. 2004; van Boekel 2008; Reyns et al. 2000; Coroller et al. 160 2006). The Akaike Information Criteria (AIC) favors the accuracy of the model through the maximum log- 161 likelihood estimate (ℓ; Eq. 8) while penalizing the indiscriminate use of extra parameters (p; Eq. 8) when 162 assessing goodness of fit among nested or non-nested models. If the underlying distribution of the 163 sample data is a normal distribution, ℓ can be calculated from the data sample size (n) and the variance 5 164 estimate ( σˆ ). The smaller the computed AICi value, the better that model (i = 1, 2, … M models) can 165 describe a particular dataset. 2 166 AICi = −2 + 2 p; = − n ln σˆ 2 2 (8) 167 168 The parameter AICi is likely to change with differences in the sample size and variability of dataset even 169 when the models compared remain unchanged. Therefore, several authors recommend using instead the 170 model AICi differences (∆AICi) calculated using the minimum AICi as a reference (Eq. 9). Furthermore, 171 ∆AICi values can be normalized via the log-likelihood term of AICi (Eq. 8-11) to obtain Akaike weights 172 (wAICi) generating a 0-1 relative scale to help in visualizing Akaike analysis differences. As a general rule 173 of thumb (Table 1), ∆AICi values in the 0-2 range yield exp(-½ ∆AICi) values around 1.0-0.3 (Eq. 10), 174 suggesting that the experimental data supports the model fit. In the case of larger ∆AICi values, Eq. 10 175 generates lower values indicating that the current experimental data does not support the model, with 176 ∆AICi ≥ 10 having little or no weight on the normalization. For example, assuming that the hypothetical 177 model 1 has the minimum AICi value, then ∆AIC1 = 0 (Eq. 9) and exp(-½ ∆AIC1) = 1 (Eq. 10; Table 1). For 178 model 2, if it is assumed that ∆AIC2 = AIC2 - AIC1 = 2 then exp(-½ ∆AIC2) = 0.368. The denominator of Eq. 179 11 (exp[-½ ∆AIC2] + exp[-½ ∆AIC2]) would yield 1.000 + 0.368 = 1.368, and the corresponding Akaike 180 weights as wAIC1 = 0.731 and wAIC2 = 0.269, for which both models are considered to adequately describe 181 the current experimental dataset. As the value of ∆AIC2 increases, wAIC2 tends to zero and the likelihood 182 that the current dataset supports model 2 is more questionable. Another way to interpret wAICi is to 183 consider wAICi as conditional probabilities according to a Bayesian framework briefly mentioned in the next 184 paragraph. In this setting, the lower the value of wAICi, the less probable that the current dataset supports 185 model i (Burnham and Anderson 2002; Johnson and Omland 2004; Wagenmakers and Farrell 2004; 186 Burnham and Anderson 2004). 187 ∆AICi = AICi − AIC min 1 = n ln σ 2 ∝ exp − ∆AIC i 2 w AICi = (9) (10) exp(− 12 ∆AIC i ) M ∑ exp(− i =1 1 2 ∆AIC i ) (11) 188 189 INSERT TABLE 1 190 6 191 The Bayesian Information Criteria (BICi) is another alternative for model selection (Eq. 12). As in the case 192 of ΔAICi values, models yielding the lowest ΔBICi value (Eq. 13) describe best the experimental data 193 collected. The Bayesian Information Criteria is not derived from information theory criteria such as the 194 AICi, but from a Bayesian approach that assumes that one of the tested models is true. Akaike weights 195 can be related to the BICi probabilities as shown in Eq. (14). 196 BICi = −2 ln + p ln n (12) ∆BICi = BICi − BIC min (13) pi = Pr[g i | data ] = exp(− 12 ∆BICi ) M ∑ exp(− i =1 1 2 ∆BICi ) = exp(− 12 ∆AICi ) M ∑ exp(− i =1 1 2 ∆AICi ) = w AICi (14) 197 198 In this work, the possibility of developing a generic pressure-time mathematical expression capable of 199 describing HPP microbial inactivation kinetics was explored, expecting that the resultant model will 200 provide a starting point towards the standardization of HPP kinetic data. The first order kinetics, Weibull 201 and Gompertz models were fit to HPP microbial inactivation data reported in peer-reviewed sources. The 202 goodness of fit of kinetic models was evaluated using standard statistics (R , mean squared error values 203 (MSE), residuals plots, experimental vs. model prediction value plots) and information theory criteria 204 (maximum log-likelihood estimate ℓ, Akaike Information Criteria AICi, Bayesian Information Criteria BICi). 205 Furthermore, secondary models describing the pressure effect on the primary model parameters were 206 analyzed and proposed as an attempt to establish a much needed model predicting pressure effects on 207 microbial inactivation rates. 2 208 209 2. Methodology 210 2.1 High pressure processing microbial inactivation data and model fit 211 An exhaustive review of references reporting microbial, chemical and enzymatic HPP kinetics was 212 performed to generate the kinetic database of this study, by selecting kinetic curves with ≥4 data points 213 (including t = 0). The most extensive database (Table 2) was found for HPP kinetic data of aerobic 214 microorganisms at ambient temperature (20-25°C), and low acidity media (pH > 4.5), totaling 124 215 microbial survival curves (974 pressure-time microbial counts sets). All microbial inactivation data was 216 manually digitized using Engauge (V4.1, www.freecode.com/projects/engauge), a freeware tool that has 217 been used in other studies for the digitalization and analysis of previously published scientific data 218 (Treseder 2008; Beauchemin et al. 2007). Data fit to the first order kinetics (LKM, Eq. 1), Weibull (WBLL, 219 Eq. 2) and modified Gompertz (GMPZ, Eq. 7) models were performed with the MATLAB “fit” function 220 (Version R2012b, www.mathworks.com/ products/matlab) using the Levenberg-Marquardt algorithm. 221 7 222 INSERT TABLE 2 223 224 2.2 Statistical analysis 225 Standard statistical analysis included residuals analysis, that is the inspection of the differences between 226 experimental (yobs) and predicted (yfit) values (Eq. 15), coefficient of determination (R ; Eq. 16), mean 227 squared error values (MSE; Eq. 17), and variance estimates ( σ ; Eq. 18). The theoretical information 228 criteria that complemented this statistical analysis of goodness of fit required the following approach. 229 Under standard assumptions and if the model is correctly specified, the residuals (Eq. 15) should be 230 distributed with a population mean µ equal 0, and a variance estimate that depends on the proposed 231 model parameters obtained through non-linear regression (Eq. 18). Therefore, the maximum log- 232 likelihood estimate ℓ is calculated with Eq. 19, AICi values are determined with Eq. 20, and ∆AICi values 233 with Eq. 9 (Burnham and Anderson 2002; Stoica and Selén 2004). It should be noted that the variance 234 estimate is also considered a parameter in the calculation of AICi values, thus the number of parameters 235 (p) for each model is: (a) LKM, p = 2 (k, 236 ). 2 2 σ 2 ); (b) WBLL, p = 3 (b’, n, σ 2 ); (c) GMPZ, p = 4 (A, µmax, λ, σ 2 237 residuals = y obs − y fit (15) SST SSE ∑ ( y − yˆ ) R = − = 1− 2 SST SST ∑ (y − y) 2 2 ∑ (y MSE = − y fit ) 2 obs n− p sˆ 2 = =− ∑ (y (16) − y fit ) (17) 2 obs (18) n n ln σˆ 2 2 AICi = −2 + 2 p = n ln σˆ 2 + 2 p (19) (20) 238 239 2.3 Development of a secondary model for HPP microbial inactivation kinetics in milk 240 First, the datasets used in this study describing the kinetics for the microbial inactivation by HPP were 241 grouped according to the type of media (milk, juice, buffer, other) and microbial population (Listeria, 242 Escherichia, mesophiles) to explore possible common microbial inactivation patterns. Next, the following 243 equations were evaluated as secondary model alternatives: (a) Eyring model (Eq. 21) for the LKM model 244 parameter k (Katsaros et al. 2010); (b) logistic-exponential (Eq. 3) and exponential decay (Eq. 4) 245 equations for the WBLL model parameters b’ and n, respectively (Peleg 2006); (c) Bigelow type model 246 (Eq. 5) and exponential decay (Eq. 4) equations for the WBLL model parameters b’ and n; (d) an 8 247 exponential sigmoidal equation (Eq. 22) for parameter A, an exponential decay equation (Eq. 23) for the 248 parameter µmax; and an exponential decay equation (Eq. 24) for the parameter λ in the GMPZ model 249 proposed in this study. Finally, the parameters obtained for the secondary model equations were used as 250 initial values to perform a one-step non-linear regression analysis in which primary and secondary models 251 were integrated into one equation. 252 ∆V ≠ ⋅ (P − Pref ) k (P ) = k 0 ⋅ exp − R ⋅T (21) A(P ) = q1 ⋅ exp[− exp(q 2 − P ⋅ q 3 )] (22) m max (P ) = f1 ⋅ exp(− f 2 ⋅ P ) (23) λ (P ) = g1 ⋅ exp(− g 2 ⋅ P ) (24) 253 254 3. Results and Discussion 255 3.1 Goodness of fit analysis of primary models with standard statistical criteria 256 The WBLL and GMPZ models gave the best data fit yielding R ≥ 0.9 on 90% of the 124 kinetic curves 257 tested in this study. The comparison of R and 258 apparent differences as summarized in Table 3. In the case of MSE, the WBLL registered slightly higher 259 and more disperse values (0.477±0.876) than the GMPZ model (0.369±0.548; Table 3), although it is 260 hard to tell if these differences are relevant due to the lack of a scale or defined interval for MSE values 261 as is the case of* the [0, 1] interval available for R . As expected, the LKM model greatly underperformed 262 consistently yielding high MSE values, low R < 0.9, and consequently high variance estimates, resulting 263 in substantial differences when compared to the WBLL and GMPZ models (Table 3). 2 2 σ2 values of the WBLL and GMPZ equations showed no 2 2 264 265 INSERT TABLE 3 266 2 267 LKM was the only model to register R < 0 (-0.382, -1.189), which occurred after fitting LKM to E. coli 268 inactivation data (200 and 250 MPa) reported in Yamamoto et al. (2005). Nevertheless, the coefficient of 269 determination should never give negative values. In this case, R < 0 occurs because the deviation of 270 model prediction is higher than the variability measured by the total sum of squares (Eq. 16), reflecting a 271 very deficient model choice and the inadequacy of R as a statistical criteria for model selection (Spiess 272 and Neumeyer 2010). The microbial survival curves of E. coli yielding R < 0 presented a very steep drop 273 of microbial counts (3-5 log reductions) in very short times (t ~ 1.0-1.5 min), and slower inactivation rates 274 afterwards that resulted in clear nonlinear curves with upward concavities (data not shown). These 275 microbial survival curve shapes caused extremely high LKM prediction deviations since the intercept (a0) 2 2 2 9 276 was fixed as a0 = 0 to force LKM fit through the origin (0, 0). Strictly, LKM prediction should be forced 277 through the origin because the integration constant (N0) is already included when microbial counts are 278 normalized as log10 (N/N0) (Serment-Moreno et al. 2014). If a0 is obtained with the regression, a0 ≠ 0 279 increases the goodness of fit as expected, yielding R = 0.615 (200 MPa) and 0.479 (250 MPa). However, 280 the equation obtained with a0 ≠ 0 is no longer a first order kinetic model. Furthermore, the resultant 281 mathematical expression becomes an empirical first order polynomial model with two parameters that 282 describes a behavior rarely observed in HPP microbial inactivation kinetics. 2 283 284 The analysis of plots of predicted vs. experimental values also showed non-significant differences 285 between the WBLL and GMPZ models (Figure 1b,c). Although both yielded R > 0.98 with slope m 286 ranging 0.991-0.993, the WBLL model showed a straighter and less disperse line (Figure 1b). Even LKM 287 values were acceptable yielding R = 0.881 and m = 0.934 (Figure 1a); however, the predictions clearly 288 deviate from the linear trend due to high overestimations observed for x-axis values in the -7 to -4 range 289 which is more evident in residuals plot (Figure 1d). The residuals for the other two models showed similar 290 90% confidence intervals (-0.45 to 0.47, WBLL; -0.54 to 0.58, GMPZ), and inaccurate predictions with 291 high positive residual values for the experimental microbial log reductions near zero (Figure 1e,f). 292 Moreover, the GMPZ residuals plot displayed an important deviation for values lower than -7 in the x-axis 293 in addition to an oscillating trend (Figure 1f). 2 2 294 295 INSERT FIGURE 1 296 297 Most GMPZ model sub-estimations occurred for microbial survival curves with two or more concavity 298 changes (DAT07, DAT08, DAT09, DAT11, and DAT17). In these cases, the asymptotic GMPZ model 299 parameter A, and the weight of microbial log count measurements prior to a last concavity change near 300 the end of the microbial survival data, do not allow predictions to reach lower values. All standard 301 statistical criteria confirmed the need to discard the use of the LKM approach, but only the subjective 302 residuals plot analysis suggested a slight superiority of the WBLL over the GMPZ model as the most 303 reliable alternative to describe the 28 microbial inactivation datasets used in this study. The GMPZ model 304 has been rarely considered as a model alternative to describe data on microbial inactivation by HPP 305 treatments. The widely preferred log-logistic model was not included in the comparisons presented in this 306 study for the reasons previously stated in section 1.1 (Serment-Moreno et al. 201X). The WBLL model 307 has been reported previously as yielding better estimates than the GMPZ model. However, in these - 308 reports, MSE and R values were the preferred statistical criteria for model comparisons (Chen and 309 Hoover 2003a, b; Guan et al. 2005; Guan et al. 2006; Koseki et al. 2009; Yamamoto et al. 2005). 2 310 311 3.2 Goodness of fit of primary models assessed with information theory criteria 10 312 When used to assess HPP microbial inactivation kinetics, AICi values varied greatly from -46.9 to 313 25.5, -54.1 to 9.1, and -178.9 to 8.9 for the LKM, WBLL, and GMPZ model, respectively. As described in 314 section 1.2, this behavior was expected since the first term of the AICi factor, the maximum log-likelihood 315 estimate (ℓ) calculated through the dataset variance, depends largely on the population size and 316 dispersion. The WBLL model yielded the lowest ∆AICi, i.e., values of 0 and ≤ 2 for 76 (61.3%) and 98 317 (79%), respectively, of the 124 kinetic curves analyzed (Table 4). The GMPZ model varied widely 318 registering ∆AICi values of 0 and ≤ 2 in 27.4 and 37.9% of the curves analyzed, while for LKM more than 319 half (50.8%) of the ∆AICi values exceeded 4. As expected the LKM model performed poorly and 320 information theory criteria just confirmed the results observed with standard statistics. While yielding the 321 lowest ∆AICi values on 11.3% of the curves analyzed, the LKM reached values > 10 in 58.1% of them 322 (Table 4), suggesting that this particular model may not be so reliable to predict HPP microbial 323 inactivation kinetics. The conclusions obtained using the BICi criteria agreed with the AICi for all tested 324 models, thus the results are not reported. 325 326 INSERT TABLE 4 327 328 3.3 HPP microbial inactivation kinetic model in milk 329 Primary model parameters showed wide scattering when plots as a function of pressure were prepared 330 for the complete microbial inactivation data available for this study (data not shown). Therefore, microbial 331 inactivation data was classified according to the media and microorganism involved in each dataset. 332 Kinetic data for microorganisms in milk (see Table 2) provided the largest amount of information with 56 333 kinetic curves, and showed identifiable trends of pressure effects on primary model parameters despite 334 wide differences in the type of microorganisms involved (data not shown). The Eyring, logistic- 335 exponential, Bigelow-exponential, sigmoidal and exponential secondary models listed in section 2.3 336 showed good agreement when correlating the corresponding primary kinetic model parameters with the 337 pressure level of the microbial inactivation treatment used. However, when the secondary model 338 parameters were used to predict the log microbial counts for the corresponding pressure-time 339 combinations (n = 429), the LKM model severely underperformed by scoring the largest variance estimate 340 ( σ = 3.37), and ∆AICi = 345.8 (Table 5). 2 341 342 These results differ with previous work by Santillana Farakos and Zwietering (2011) who found good 343 agreement when using the Bigelow model to describe pressure and temperature effects on LKM 344 parameters. These authors fitted the Bigelow model to decimal reduction time (D) values reported in 345 peer-reviewed sources, and generated a different dataset from the one used in this study. However, when 346 D values were not included in the published reports used in their work, the LKM regression included only 347 the linear portion of the microbial survival curve even when concave trends were observed (Santillana 348 Farakos and Zwietering 2011). 11 349 350 INSERT TABLE 5 351 352 Standard statistical criteria showed a similar goodness of fit when describing microbial inactivation using 353 WBLL and GMPZ models parameters as a function of pressure (Table 5). However, a one-step non-linear 354 regression could not be performed when incorporating Eq. 22-24 into the primary GMPZ model (Eq. 7) 355 due to the complexity of the resulting mathematical expression. The secondary models for pressure 356 effects on GMPZ parameters registered a variance estimate 357 observed values with slope m = 0.930, considering the 429 pressure-time kinetic data combinations 358 available for milk. The 90% confidence residuals interval was (-2.13, 2.26; Table 5) but values were not 359 evenly distributed as only positive residuals occurred for x-axis values between [-1, 0], and only negative 360 residuals were observed for x-axis values [-9, -7] (data not shown). The maximum inactivation rate and 361 the inflection point tended to follow exponential trends when plotted against pressure. σ 2 = 1.55 (Table 5) and a fit of predicted vs. 362 363 The asymptote difference parameter (A) showed the greatest dispersion of all three parameters. The 364 divergence might be attributed to microbial log reductions achieved during the come up time (CUT). Even 365 when reports used in this study presented CUT values (see Table 2; DAT 01-02, DAT10, DAT12-16, 366 DAT18-19, DAT26-28), traditional normalization of microbial inactivation data (log10 N/N0 or log10 N/NCUT) 367 to linearize the kinetic data does not take into account the CUT reduction effect as exemplified with data 368 reported by Ahn et al. (2007). Normalizing the spore inactivation data as log10 N/NCUT does not consider 369 the 2-3 log reductions during CUT at 700 MPa/105°C, or the 4-6 log reductions at 700 MPa/121°C (see 370 Figures 1b and 1d presented by Ahn et al. 2007). Conversely, if the microbial inactivation data is 371 normalized as log10 N/N0 the spore survival curves will present a significantly steeper slope, theoretically 372 resulting in a higher b’ value for the WBLL model (Eq. 2), or a larger A parameter in the case of the 373 GMPZ model (Eq. 7). Still, 69.2% of the milk HPP microbial inactivation data fell within the interval defined 374 by the GMPZ model prediction ± standard deviation (yfit ± σ ; Figure 2). The 300-400 MPa pressure range 375 registered 14.2% of the cases in which the experimental data remained outside the yfit ± σ interval, and a 376 similar percentage occurred when the pressure level exceeded 400 MPa (14.7%). These deviations may 377 be attributed to microbial physiological factors or biochemical changes occurring in milk at pressure 378 exceeding 200 MPa (Considine et al. 2007; Huppertz et al. 2006). 379 380 This work represents the first time that secondary models are developed for sigmoidal kinetics for 381 microbial inactivation by HPP. Erkmen (2009) used Arrhenius and square root models as secondary 382 models, but only evaluated the temperature effect for Salmonella typhimurium inactivation in tryptic soy 383 broth (200-350 MPa; 15-45°C). The use of the expressions here proposed could be extended to the 384 analysis of sigmoidal curves observed in other microbial inactivation alternatives including conventional 385 thermal food processing. Regarding other sigmoidal secondary models, Serment-Moreno et al. (201X) 12 386 analyzed the possibility of developing a secondary model for the log-logistic model (Eq. 6), but the 387 dependence of the log-logistic model parameters on pressure could not be determined. 388 389 INSERT FIGURE 2 390 391 The logistic-exponential WBLL model obtained after a one-step non-linear regression yielded higher than 392 expected predicted values caused by microbial inactivation data following sigmoidal curves. Predicted 393 survival counts were comparable to those obtained when using the secondary GMPZ model (Figure 3), 394 although the WBLL model clearly registered the highest goodness of fit described by the AICi and ∆AICi 395 values reported in Table 5. The comparison of the variance estimate using a log scale via the log- 396 likelihood estimate and the parsimony term of the AICi value (Eq. 20) played an important role when 397 addressing WBLL and GMPZ model differences, as variance estimates differed just by 0.05 units (Table 398 5). 399 400 INSERT FIGURE 3 401 402 The WBLL model displayed upward concavity at all tested pressure levels as denoted by predicted n 403 values <1 and decreasing exponentially from n = 0.758 at 150 MPa to n = 0.403 at 600 MPa. The value of 404 n under isothermal conditions remained relatively unchanged (|n(P1) – n(P2)| < 0.1) for narrow pressure 405 intervals (~100 MPa) as reported in other studies (Table 6). Nevertheless, precautions must be taken with 406 this assumption, since currently there is no evidence that non-apparent differences among n values will 407 result in non-significant differences of the predicted microbial counts which are more transcendent. The 408 WBLL inactivation rate constant increased from b’ = 0.056 to b’ = 2.230 min 409 range. Both parameters (b’, n) displayed good agreement with the microbial pressure resistance 410 described by the Weibull secondary model (Eq. 2-4, van Boekel 2002; see Chapter 6, Peleg 2006; 411 Serment-Moreno et al. 2014). -n for the same pressure 412 413 INSERT TABLE 6 414 415 The Bigelow model predicting the pressure dependence of b’ (Eq. 5) yielded zP = 290 MPa (Table 5), 416 which is similar to zP = 298 MPa for Listeria spp. Reported by Santillana Farakos and Zwietering (2011). 417 Nevertheless these results are not equivalent, since zP obtained by these authors corresponds to the first 418 order kinetic model fit of Listeria spp. data in various media such as beef, pork, milk or buffers. The 419 Bigelow model yielded b’ref = 0.0481 at Pref = 117.4 MPa (Table 5), which is close to b’ = 0.056 (100 MPa) 420 predicted with the logistic-exponential model. The parameters describing the pressure dependence of n 421 (d0, d1; Eq. 4), and the estimated variances also remained close when comparing the logistic-exponential 422 and Bigelow-exponential WBLL models (Table 5). Despite this apparent similarities, the logistic- 13 423 exponential model clearly resulted a better option registering ℓ = -86.9, while the Bigelow-exponential 424 model yielded ℓ = -111.9 and a wider, non-symmetric residuals interval (Table 5). The Bigelow- 425 exponential model also uses one more parameter than the logistic-exponential model, resulting in a large 426 ∆AICi = 51.9 (Table 5). 427 428 Doona et al. (2008) used the logistic-exponential Weibull secondary model (Eq. 3-4) to analyze E. coli 429 inactivation in a whey protein model (214-439 MPa; 30-50°C). The low MSE values (0.003-0.932) and 430 visual inspection of the model fit showed a good agreement between experimental values and model 431 predictions. The concavity of the curves changed when the pressure level applied increased, showing 432 upward concavity at the higher pressure levels with the exponent decreasing from n = 3.37 (214 MPa) to 433 n = 0.70 (439 MPa) at 30°C. The Weibull rate constant increased from b’ = 2.70x10 to b’ = 0.482 min 434 for the same pressure-temperature conditions (214-439 MPa; 30°C). The parameter estimates obtained 435 by Doona et al. (2008) diverged from values obtained in this study using the generic WBLL model for 436 HPP microbial inactivation data in milk at 25°C. Different experimental factors can affect non-linear 437 regression results including the pressure-temperature range, media treated, and microbial flora type likely 438 being the most influential ones. -8 -n 439 440 Most importantly, the results in the current study imply that a relatively simple equation such as the 441 logistic-exponential WBLL four parameter secondary model might be sufficient to deliver an effective 442 mathematical tool for the experimental and industrial design of HPP treatments when inactivation kinetic 443 data of similar microorganism types in a common media is grouped together. In spite of the regular use of 444 the WBLL model in thermal food processing (Halder et al. 2007; Peleg 2002; van Boekel 2002; Peleg 445 2006), except for the work by Doona et al. (2008), no other secondary model applications to model HPP 446 inactivation kinetics were found in published work. Although alternative Weibull secondary models have 447 been proposed, these mathematical expressions have been tested only with the datasets used to derive 448 them (Chen and Hoover 2003b; Buzrul et al. 2008; Pilavtepe-Çelik et al. 2009; Serment-Moreno et al. 449 2014). 450 451 3.3 HPP microbial inactivation kinetic model in milk with short processing times (<10 min) 452 Microbial inactivation data for extremely prolonged HPP treatments is of questionable relevance to 453 industrial applications of this technology. In addition, longer times increased the prevalence of sigmoidal 454 curves in the milk data used in this study. Therefore, the one-step non-linear regression with the 455 secondary logistic-exponential Weibull model was repeated by analyzing only microbial kinetic curves 456 with n ≥ 3 data points and treatment times below 10 min. The pressure range for the remaining 12 kinetic 457 curves (n = 73 pressure-time microbial data combinations) covered the 300-600 MPa range (DAT08-09, 458 DAT11, DAT13-14, and DAT24-25). Although this dataset is very limited, the results showed important 459 differences when compared to the complete milk kinetic data used in Section 3.3. The parameters b’ and 14 -n 460 n ranged from 0.079 to 1.859 min and 1.340 to 0.557, respectively, for the 300-600 MPa range whereas 461 the complete milk dataset yielded b’ = 0.056 to 2.230 min and n = 0.758 to 0.403 for the 150-600 MPa 462 region. The secondary model parameters wP = 0.0139 MPa and Pc = 478.7 MPa remained close to the 463 values reported in Table 5, whereas the parameter d0, representing the value of the Weibull exponent (n; 464 Eq. 2) at low pressure levels, dramatically increased from d0 = 0.9358 (Table 5) to d0 = 3.224. 465 Consequently, the concavity of the survival curves changed from downward to upward concavity as the 466 pressure level increased, displaying linear trends in the 400-450 MPa region, while the Weibull model 467 obtained for the complete milk dataset only showed upward concavity curves. Despite the limited amount 468 of kinetic data, these results suggest that an inappropriate choice of the HPP experiment duration can 469 result in estimates of kinetic parameters that do not describe well HPP treatment conditions of commercial 470 interest. -n -1 471 472 4. Conclusions 473 This study provided a much needed comparison for three of the most widely used microbial inactivation 474 kinetic models. It showed also that the sole use of standard statistical tools (R , MSE, residuals plots, and 475 fit of experimental vs. model prediction plots) to select a microbial inactivation model may be insufficient. 476 Standard statistics could not distinguish whether the WBLL or the GMPZ model were the most adequate 477 primary HPP microbial inactivation model. Further analysis with statistical parameters based on 478 information theory (maximum log-likelihood estimate, ℓ; Akaike Information Criteria, AICi; Akaike 479 differences, ∆AICi) indicated that the WBLL outperformed the LKM and the GMPZ primary models on 76 480 of the 124 kinetic curves analyzed. In spite of its historical importance in food processing, LKM is not a 481 reliable alternative to model HPP microbial inactivation kinetics since it performed poorly during primary 482 and secondary modeling. Akaike differences were ≥10 in 58.1% of the kinetic curves analyzed for primary 483 modeling, whereas the secondary Eyring model registered ∆AICi = 345.8 when compared to the logistic- 484 exponential Weibull secondary model. Most importantly, this study showed that information theory criteria 485 adequately complemented the use of standard statistics, and highlighted the importance of considering 486 several statistical parameters to evaluate the goodness of fit of mathematical models. 2 487 488 Although a global model for HPP inactivation could not be developed, this study suggests that generic 489 models are possible when limited to the HPP inactivation of similar types of microorganisms and media, 490 which would be useful for HPP design. The logistic-exponential WBLL and empirical GMPZ secondary 491 models derived for microbial inactivation kinetics in milk had equal goodness of fit according to standard 492 statistics. However, the logistic-exponential WBLL secondary model registered the lowest AICi value 493 (181.8), and a high difference when compared to the large AICi value of the GMPZ model (∆AICi = 19.3) 494 or the Bigelow-exponential WBLL model which performed poorly (∆AICi = 51.9). The simplicity of the 495 logistic-exponential Weibull model was an advantage, since only 4 parameters plus the variance estimate 496 are required by the flexible, easy to use logistic-exponential and exponential decrease functions used in 15 497 the secondary WBLL model. The empirical equations defining the secondary GMPZ model (8 498 parameters) also predicted accurately microbial log counts at different pressure-time combinations, but 499 they were deemed to be impractical because their complexity prevented a one-step non-linear regression. 500 These empirical equations are the only ones available for HPP microbial inactivation data following a 501 sigmoidal behavior and their application could be extended to analyze the kinetics of microbial 502 inactivation for other food processing technologies. 503 504 Finally, an important concern raised by this study is that much of the research on microbial inactivation by 505 HPP treatment does not match commercial processing conditions. When a microorganism survives a low 506 pressure level, researchers consistently extend treatment times far beyond values with commercial 507 viability. These long treatment times often result in microbial inactivation curves characterized by multiple 508 concavity changes that may interfere with the estimation of “practical” kinetic parameter for more realistic 509 HPP treatment conditions. A non-linear regression analysis considering only microbial data for t ≤ 10 min 510 (12 kinetic curves; n = 73 pressure-time microbial data combinations) supports the recommendation of 511 focusing on short pressure holding times. The estimated WBLL model parameters and the shape of the 512 microbial survival curves clearly differed from those obtained when long treatment times were included. 513 All resources are limited and they should be employed to generate microbial inactivation data at an 514 appropriate pressure level to deliver significant microbial inactivation in a reasonable time (preferably less 515 than 10 min) that complies with the needs of industrial HPP food processors. 516 517 Acknowledgments 518 Authors Vinicio Serment-Moreno and Jorge Welti-Chanes acknowledge the support from the Tecnológico 519 de Monterrey (Research chair funds CAT 200 and CDB081), and México’s CONACYT Scholarship 520 Program. This project was also supported by Formula Grants no. 2011-31200-06041 and 2012-31200- 521 06041 from the USDA National Institute of Food and Agriculture. 522 16 523 Nomenclature A AICi AICmin b’ b'ref d0 d1 fi; i=1,2 gi; i=1,2 HPP k k0 ℓ MSE N N0 n P Pc Pref p qi; i=1,2,3 2 R SSE SST T t wAICi wP yfit yobs zP -1 Asymptote difference, Gompertz model, log-logistic model; log10 cfu ml Akaike Information Criterion for model i Minimum AICi value among the models tested -n Slope parameter, Weibull model; min -n Slope parameter at a reference pressure; Bigelow type Weibull secondary model; min Parameter n at atmospheric pressure, Weibull secondary model; dimensionless -1 Exponential pressure dependence of parameter n, Weibull secondary model; MPa Parameters of empirical secondary Gompertz model Parameters of empirical secondary Gompertz model High Pressure Processing -1 Inactivation rate constant, first order kinetics model; min -1 Inactivation rate constant at a reference pressure, Eyring model; min Maximum log-likelihood estimate Mean squared error -1 -1 Microbial population; cfu g , cfu ml -1 -1 Microbial population prior to the high pressure treatment; cfu g , cfu ml Number of observed data points Exponent, Weibull model Pressure; MPa Critical pressure parameter, Weibull log-logistic secondary model; MPa Reference pressure, Eyring model, Bigelow type Weibull secondary model ; MPa Number of estimated parameters, Akaike Information Criterion Parameters of empirical secondary Gompertz model Coefficient of determination Error sum of squares; coefficient of determination Total sum of squares; coefficient of determination Temperature; °C Time; min Akaike weight factor -1 Exponential pressure dependence of parameter b’, Weibull secondary model; MPa Microbial counts with primary or secondary HPP inactivation models Experimental microbial counts for literature reported HPP treatments Pressure resistance factor; Bigelow type Weibull secondary model; MPa 524 525 Greek symbols α ΔAICi ≠ ∆V λ Ω δ μmax σˆ σ̂ 2 τ Significance levels for statistical analysis AICi differences among the tested models by taking AICmin as reference 3 -1 Activation volume change, Eyring model; cm mol Inflection or starting point of the linear portion of sigmoidal inactivation curves, Gompertz and log-logistic models; min -1 Maximum inactivation rate, log-logistic model; log cfu min -1 Slope parameter, log-logistic model; decimal microbial reductions (log time units) -1 -1 Maximum inactivation rate, Gompertz model; log10 (cfu ml ) min Standard deviation Variance estimate Log time for maximum inactivation rate in “vitalistic” log-logistic model; log10 min 526 17 527 Table 1. Exponentiation of Akaike differences (∆AICi) ∆AICi 0 1 2 3 4 5 7 10 exp(-½ ∆AICi) 1.000 0.607 0.368 0.223 0.135 0.082 0.030 0.007 528 18 529 Table 2. References reporting microbial inactivation by high pressure processing (HPP) of low acid foods at ambient temperature (2025°C) Label T(°C) P(MPa) Microorganism Media pH Number of microbial survival curves DAT01 DAT02 DAT03 20 20 25 150-350 150-350 200-300 Milk Milk Ewe milk NR NR 6.7 4 4 3 DAT04 25 200-300 Ewe milk 6.7 3 DAT05 DAT06 DAT07 DAT08 DAT09 DAT10 DAT11 DAT12 DAT13 DAT14 DAT15 DAT16 DAT17 DAT18 DAT19 DAT20 25 25 22 22 25 21 25 25 25 25 25 25 25 20 20 25 200-500 200-500 300-450 350-500 400-500 250-400 400-600 200-700 300-600 300-600 300-600 300-600 350-600 200-500 300-600 100-300 Listeria monocytogenes Natural microflora Escherichia coli Pseudomonas fluorescens Lactobacillus helveticus Staphylococcus aureus Yersinia enterocolitica Yersinia enterocolitica Listeria monocytogenes Mesophiles Listeria monocytogenes Listeria monocytogenes Mesophiles Listeria monocytogenes Mesophiles Listeria monocytogenes Salmonella typhimurium Escherichia coli Escherichia coli Escherichia coli Ewe milk Ewe milk Phosphate buffer Milk Milk Milk Milk BHIB Milk Milk Peach juice Peach juice Milk HEPES-KOH buffer Carrot juice Phosphate buffer 6.7 6.7 7.0 6.7 6.7 NR 6.7 6.7 6.7 6.7 5.2 5.2 NR 7.0 6.6 6.8 5 5 4 4 2 4 3 6 3 3 3 3 6 7 7 6 DAT21 20 200-400 Escherichia coli Peptone water NR 5 DAT22 DAT23 DAT24 DAT25 DAT26 DAT27 DAT28 23 23 22 22 25 25 25 350-550 450-600 400-600 400-600 250-400 250-400 200-350 Staphylococcus aureus Staphylococcus aureus Escherichia coli Listeria innocua Escherichia coli Listeria monocytogenes Salmonella typhimurium Phosphate buffer Ham Whole milk Whole milk Fish slurry Fish slurry Tryptone soy broth 6.7 6.7 6.6 6.6 NR NR 7.0 6 6 5 5 4 4 4 Reference (Mussa et al. 1999) (Mussa et al. 1999) (Gervilla et al. 1999a) (Gervilla et al. 1999a) (Gervilla et al. 1999b) (Gervilla et al. 1999b) (Chen and Hoover 2003b) (Chen and Hoover 2003b) (Chen & Hoover, 2003a) (Pandey et al. 2002) (Chen and Hoover 2004) (Dogan and Erkmen 2004) (Dogan and Erkmen 2004) (Dogan and Erkmen 2004) (Dogan and Erkmen 2004) (Dogan and Erkmen 2004) (Guan et al. 2005) (Van Opstal et al. 2005) (Van Opstal et al. 2005) (Yamamoto et al. 2005) (Koseki and Yamamoto 2007) (Tassou et al. 2007) (Tassou et al. 2007) (Buzrul et al. 2008) (Buzrul et al. 2008) (Ramaswamy et al. 2008) (Ramaswamy et al. 2008) (Erkmen 2009) 530 531 19 2 Table 3. Coefficient of determination (R ), mean squared error (MSE) and variance estimate ( σˆ ) for the first order kinetics (LKM), Weibull (WBLL) and Gompertz (GMPZ) model describing HPP microbial inactivation at ambient temperatures (20-25°C) 2 2 σˆ 2 LKM mean sd min max R 0.789 0.294 -1.189 0.997 MSE 7.969 13.652 0.005 97.618 0.925 1.644 0.001 10.846 WBLL mean sd min max 0.965 0.050 0.708 0.999 0.477 0.876 -5 0.987x10 5.616 0.110 0.232 -5 4.894x10 1.871 GMPZ mean sd min max 0.965 0.037 0.822 1.000 0.369 0.548 0.000 2.709 0.110 0.165 0.000 1.156 532 20 533 Table 4. Akaike difference (∆AICi) distribution for the first order kinetics (LKM), Weibull (WBLL) and Gompertz (GMPZ) primary inactivation models describing HPP microbial inactivation kinetics in milk at ambient temperature (20-25°C) Criteria LKM WBLL GMPZ ∆AICi = 0 11.3% 61.3% 27.4% ∆AICi ≤ 2 19.4% 79.0% 37.9% 2 < ∆AICi ≤ 4 4.8% 5.7% 11.3% 4 < ∆AICi ≤ 10 17.7% 9.7% 25.8% ∆AICi > 10 58.1% 5.7% 25.0% 21 534 535 536 Table 5. Statistical summary of secondary HPP microbial inactivation models in milk at ambient temperature (20-25°C). (1) (1) (1) LKM WBLL WBLL k(P) Parameter k0 0.0255 ≠ -∆V Pref Residuals (d) range b'(P) -19.65 wP Pc (a) n(P) 0.0111 d0 d1 408.7 0.9358 0.0014 b'(P) bref Pref zP 240.3 (b) 0.0481 117.4 290 n(P) d0 d1 0.9022 0.0010 GMPZ A(P) p1 p2 p3 Λ(P) µmax(P) f1 -7.795 f2 2.111 -0.017 -3 6.174x10 g1 429.8 g2 -0.0174 -3 8.996x10 -3.30 to 2.00 -1.92 to 1.96 -2.16 to 1.78 -2.13 to 2.26 3.37 1.50 1.68 1.55 AICi 529.6 183.8 235.7 203.1 ∆AICi 345.8 0 51.9 19.3 (1) Results obtained from a one-step non-linear regression incorporating the secondary model into the primary model as one equation; (a) logisticexponential model; (b) Bigelow type-exponential model; (d) 90% confidence interval of residuals 22 Table 6. Parameter n average values for different pressure ranges. Microorganism Various Pressure range Pressure (MPa) levels 0.661±0.038 200-300 3 0.574±0.033 300-400 3 Generic Weibull 0.499±0.029 400-500 3 secondary milk model of 0.434±0.025 500-600 3 current work 0.467±0.046 400-600 5 0.477 300-400 3 Milk 0.583 400-500 3 Milk 0.483 400-600 5 Chen and Hoover (2004) 400-600 5 Buzrul et al. (2008) Media Milk Phosphate Y. eneterocolitica L. monocytogenes buffer E. coli L. innocua Milk n 0.780 0.790 Reference Chen and Hoover (2003b) 537 538 23 539 5. References 540 Ahn J, Balasubramaniam VM & Yousef AE (2007) Inactivation kinetics of selected aerobic and anaerobic 541 bacterial spores by pressure-assisted thermal processing. International Journal of Food 542 Microbiology. 113(3), 321-329. 543 544 545 546 547 Albert I & Mafart P (2005) A modified Weibull model for bacterial inactivation. International Journal of Food Microbioloogy. 100(1-3), 197-211. Beauchemin C, Dixit NM & Perelson AS (2007) Characterizing T cell movement within lymph nodes in the absence of antigen. The Journal of Immunology. 178, 5505–5512. Bermúdez-Aguirre D & Barbosa-Cánovas GV (2011) An update on high hydrostatic pressure, from the 548 laboratory to industrial applications. Food Engineering Reviews. 3(1), 44-61. 549 Breiman L (2001) Statistical modeling: the two cultures. Statistical Science. 16(3), 199–231. 550 Burnham KP & Anderson DR (2002) Information and likelihood theory: a basis for model selection and 551 inference. In: 552 Approach. 2nd edn. p^pp 49-97. Springer-Verlag, New York, NY. 553 554 Model Selection and Multimodel Inference: A Practical Information-Theoretic Burnham KP & Anderson DR (2004) Multimodel Inference: understanding AIC and BIC in model selection. Sociological Methods & Research. 33(2), 261-304. 555 Buzrul S, Alpas H, Largeteau A & Demazeau G (2008) Modeling high pressure inactivation of Escherichia 556 coli and Listeria innocua in whole milk. European Food Research and Technology. 227(2), 443- 557 448. 558 Campanella OH & Peleg M (2001) Theoretical comparison of a new and the traditional method to 559 calculate Clostridium botulinum survival during thermal inactivation. Journal of the Science of 560 Food and Agriculture. 81(11), 1069-1076. 561 Carreño J, Gurrea M, Sampedro F & Carbonell J (2011) Effect of high hydrostatic pressure and high- 562 pressure homogenisation on Lactobacillus plantarum inactivation kinetics and quality parameters 563 of mandarin juice. European Food Research and Technology. 232(2), 265-274. 564 Chen H & Hoover DG (2003a) Modeling the combined effect of high hydrostatic pressure and mild heat 565 on the inactivation kinetics of Listeria monocytogenes Scott A in whole milk. Innovative Food 566 Science and Emerging Technologies. 4(1), 25-34. 24 567 568 Chen H & Hoover DG (2003b) Pressure inactivation kinetics of Yersinia enterocolitica ATCC 35669. International Journal of Food Microbiology. 87(1-2), 161-171. 569 Chen H & Hoover DG (2004) Use of Weibull model to describe and predict pressure inactivation of 570 Listeria monocytogenes Scott A in whole milk. Innovative Food Science and Emerging 571 Technologies. 5(3), 269-276. 572 Cole MB, Davies KW, Munro G, Holyoak CD & Kilsby DC (1993) A vitalistic model to describe the thermal 573 inactivation of Listeria monocytogenes. Journal of Industrial Microbiology & Biotechnology. 12(3), 574 232-239. 575 Considine T, Patel HA, Anema SG, Singh H & Creamer LK (2007) Interactions of milk proteins during 576 heat and high hydrostatic pressure treatments — A Review. Innovative Food Science and 577 Emerging Technologies. 8(1), 1-23. 578 Coroller L, Leguerinel I, Mettler E, Savy N & Mafart P (2006) General Model, Based on Two Mixed 579 Weibull Distributions of Bacterial Resistance, for Describing Various Shapes of Inactivation 580 Curves. Applied and Environmental Microbiology. 72(10), 6493-6502. 581 Corradini MG, Normand MD, Newcomer C, Schaffner DW & Peleg M (2009) Extracting Survival 582 Parameters from Isothermal, Isobaric, and “Iso-concentration” Inactivation Experiments by the “3 583 End Points Method”. Journal of Food Science. 74(1), R1-R11. 584 585 586 587 Dogan C & Erkmen O (2004) High pressure inactivation kinetics of Listeria monocytogenes inactivation in broth, milk, and peach and orange juices. Journal of Food Engineering. 62(1), 47-52. Doona CJ & Feeherry FE (2007) High pressure processing of foods. In. p^pp 265. Blackwell Publishing, Ames, IA. 588 Doona CJ, Feeherry FE, Ross EW, Corradini M & Peleg M (2008) The quasi-chemical and Weibull 589 distribution models of nonlinear inactivation kinetics of Escherichia coli ATCC 11229 by high 590 pressure processing. In: 591 Publishing Ltd. High Pressure Processing of Foods. p^pp 115-144. Blackwell 592 Doona CJ, Feeherry FE, Ross EW, Corradini MG & Peleg M (2007) The quasi-chemical and Weibull 593 distribution models of nonlinear inactivation kinetics of Escherichia coli ATCC 11229 by high 594 pressure processing. In: Doona CJ & Feeherry FE (eds) High Pressure Processing of Foods. 595 p^pp 115-114. IFT Press Blackwell Publishing, Ames, IA. 25 596 597 Erkmen O (2009) Mathematical modeling of Salmonella typhimurium inactivation under high hydrostatic pressure at different temperatures. Food and Bioproducts Processing. 87(1), 68-73. 598 Gervilla R, Mor-Mur M, Ferragut V & Guamis B (1999a) Kinetics of destruction of Escherichia coli and 599 Pseudomonas fluorescens inoculated in ewe’s milk by high hydrostatic pressure. Food 600 Microbiology. 16(2), 173-184. 601 Gervilla R, Sendra E, Ferragut V & Guamis B (1999b) Sensitivity of Staphylococcus aureus and 602 Lactobacillus helveticus in ovine milk subjected to high hydrostatic pressure. Journal of Dairy 603 Science. 82(6), 1099-1107. 604 Guan D, Chen H & Hoover DG (2005) Inactivation of Salmonella typhimurium DT 104 in UHT whole milk 605 by high hydrostatic pressure. International Journal of Food Microbiology. 104(2), 145-153. 606 Guan D, Chen H, Ting EY & Hoover DG (2006) Inactivation of Staphylococcus aureus and Escherichia 607 coli O157:H7 under isothermal-endpoint pressure conditions. Journal of Food Engineering. 77(3), 608 620-627. 609 Halder A, Datta AK & Geedipalli SSR (2007) Uncertainty in thermal process calculations due to variability 610 in first-order and Weibull kinetic parameters. Journal of Food Science. 72(4), E155-E167. 611 Heinz V & Buckow R (2009) Food preservation by high pressure. Journal für Verbraucherschutz und 612 613 614 615 616 Lebensmittelsicherheit. 5(1), 73-81. Huppertz T, Smiddy MA, Upadhyay VK & Kelly AL (2006) High pressure‐induced changes in bovine milk proteins: a review. International Journal of Dairy Technology. 59(2), 58-66. Johnson JB & Omland KS (2004) Model selection in ecology and evolution. Trends in Ecology and Evolution. 19(2), 101-108. 617 Katsaros GI, Tsevdou M, Panagiotou T & Taoukis PS (2010) Kinetic study of high pressure microbial and 618 enzyme inactivation and selection of pasteurisation conditions for Valencia Orange Juice. 619 International Journal of Food Science & Technology. 45(6), 1119-1129. 620 Koseki S, Matsubara M & Yamamoto K (2009) Prediction of a required log reduction with probability for 621 Enterobacter sakazakii during High-Pressure Processing, using a survival/death interface model. 622 Applied and Environmental Microbiology. 75(7), 1885-1891. 623 624 Koseki S & Yamamoto K (2007) A novel approach to predicting microbial inactivation kinetics during high pressure processing. International Journal of Food Microbiology. 116(2), 275-282. 26 625 López S, Prieto M, Dijkstra J, Dhanoa MS & France J (2004) Statistical evaluation of mathematical 626 models for microbial growth. International Journal of Food Microbioloogy. 96(3), 289-300. 627 Mújica-Paz H, Valdez-Fragoso A, Tonello Samson C, Welti-Chanes J & Torres JA (2011) High-pressure 628 processing technologies for the pasteurization and sterilization of foods. Food and Bioprocess 629 Technology. 4(6), 969-985. 630 631 Mussa DM, Ramaswamy HS & Smith JP (1999) High pressure (HP) destruction kinetics of Listeria monocytogenes Scott A in raw milk. Food Research International. 31(5), 343-350. 632 Pandey PK, Ramaswamy HS & Idziak E (2002) High pressure destruction kinetics of indigenous 633 microflora and Escherichia coli in raw milk at two temperatures. Journal of Food Process 634 Engineering. 26(3), 265-283. 635 636 637 638 639 640 Peleg M (2002) Simulation of E. Coli inactivation by carbon dioxide under pressure. Journal of Food Science. 67(3), 896-901. Peleg M (2006) Advanced Quantitative Microbiology for Foods and Biosystems. Contemporary Food Science. CRC Taylor & Francis, Peleg M, Engel R, Gonzalez-Martinez C & Corradini MG (2002) Non-Arrhenius and non-WLF kinetics in food systems. Journal of the Science of Food and Agriculture. 82(12), 1346-1355. 641 Pilavtepe-Çelik M, Buzrul S, Alpas H & Bozoğlu F (2009) Development of a new mathematical model for 642 inactivation of Escherichia coli O157:H7 and Staphylococcus aureus by high hydrostatic pressure 643 in carrot juice and peptone water. Journal of Food Engineering. 90(3), 388-394. 644 Ramaswamy HS, Zaman SU & Smith JP (2008) High pressure destruction kinetics of Escherichia coli 645 (O157:H7) and Listeria monocytogenes (Scott A) in a fish slurry. Journal of Food Engineering. 646 87(1), 99-106. 647 Reyns KMFA, Soontjens CCF, Cornelis K, Weemaes CA, Hendrickx ME & Michiels CW (2000) Kinetic 648 analysis and modelling of combined high-pressure-temperature inactivation of the yeast 649 Zygosaccharomyces bailii. International Journal of Food Microbiology. 56(2/3), 199-210. 650 651 Ross T (1996) Indices for performance evaluation of predictive models in food microbiology. Journal of Applied Bacteriology. 81, 501-508. 652 Santillana Farakos SM & Zwietering MH (2011) Data analysis of the inactivation of foodborne 653 microorganisms under high hydrostatic pressure to establish global kinetic parameters and 654 influencing factors. Journal of Food Protection. 74(12), 2097-2106. 27 655 Saucedo-Reyes D, Marco-Celdrán A, Pina-Pérez MC, Rodrigo D & Martínez-López A (2009) Modeling 656 survival of high hydrostatic pressure treated stationary- and exponential-phase Listeria innocua 657 cells. Innovative Food Science and Emerging Technologies. 10, 135-141. 658 659 Serment-Moreno V, Barbosa-Cánovas G, Torres JA & Welti-Chanes J (2014) High pressure processing: Kinetic Models for Microbial and Enzyme Inactivation. Food Engineering Reviews. 6, 56-88. 660 Serment-Moreno V, Torres JA, Welti-Chanes J & Fuentes C (201X) Information theory analysis of log- 661 logistic model forms describing microbial inactivation by high pressure processing (HPP) following 662 sigmoidal trends. In review. 663 664 2 Spiess A-N & Neumeyer N (2010) An evaluation of R as an inadequate measure for nonlinear models in pharmacological and biochemical research: a Monte Carlo approach. BMC Pharmacology. 10(6). 665 Spinner J (2014) Hiperbaric “can’t complain” about growth in HPP market. Food Production Daily, 666 Retrieved May 12, 2014, from http://www.foodproductiondaily.com/Processing/Hiperbaric-can-t- 667 complain-about-growth-in-HPP-market. 668 669 Stoica P & Selén Y (2004) Model‐order selection: a review of information criterion rules. Signal Processing Magazine, IEEE. 21(4), 36-47. 670 Tassou CC, Galiatsatou P, Samaras FJ & Mallidis CG (2007) Inactivation kinetics of a piezotolerant 671 Staphylococcus aureus isolated from high-pressure-treated sliced ham by high pressure in buffer 672 and in a ham model system: Evaluation in selective and non-selective medium. Innovative Food 673 Science and Emerging Technologies. 8(4), 478-484. 674 675 676 677 678 679 Treseder KK (2008) Nitrogen additions and microbial biomass: a meta-analysis of ecosystem studies. Ecology Letters. 11, 1111–1120. van Boekel MAJS (2002) On the use of the Weibull model to describe thermal inactivation of microbial vegetative cells. International Journal of Food Microbiology. 74, 139-159. van Boekel MAJS (2008) Kinetic modeling of food quality: a critical review. Comprehensive Reviews in Food Science and Food Safety. 7(1), 144-158. 680 Van Opstal I, Vanmuysen SCM, Wuytack EY, Masschalck B & Michiels CW (2005) Inactivation of 681 Escherichia coli by high hydrostatic pressure at different temperatures in buffer and carrot juice. 682 International Journal of Food Microbiology. 98(2), 179-191. 683 684 Wagenmakers E-J & Farrell S (2004) AIC model selection using Akaike weights. Psychonomic Bulletin & Review. 11(1), 192-196. 28 685 Wilson DR, Dabrowski L, Stringer S, Moezelaar R & Brocklehurst TF (2008) High pressure in combination 686 with elevated temperature as a method for the sterilisation of food. Trends in Food Science & 687 Technology. 19(6), 289-299. 688 Yamamoto K, Matsubara M, Kawasaki S, Bari M & Kawamoto S (2005) Modeling the pressure 689 inactivation dynamics of Escherichia coli. Brazilian journal of medical and biological research. 690 38(8), 1253-1257. 691 692 693 29 694 6. List of Tables 695 696 Table 1. 697 Exponentiation of Akaike differences (∆AICi). 698 699 Table 2. 700 References reporting microbial inactivation by high pressure processing (HPP) at ambient temperature 701 (20-25°C) and low acidity foods. 702 703 Table 3. 704 Coefficient of determination (R ), mean squared error (MSE) and variance estimate ( σˆ 2 ) for the first order 705 kinetics (LKM), Weibull (WBLL) and Gompertz (GMPZ) models describing HPP microbial inactivation at 706 ambient temperatures (20-25°C). 2 707 708 Table 4. 709 Akaike difference (∆AICi) distribution of the first order kinetics (LKM), Weibull (WBLL) and Gompertz 710 (GMPZ) primary inactivation models describing milk HPP microbial inactivation kinetics at ambient 711 temperature (20-25°C). 712 713 Table 5. 714 Statistical summary of secondary HPP microbial inactivation models in milk at ambient temperature (20- 715 25°C). 716 717 30 718 7. List of Figures 719 720 Figure 1. 721 Model vs. reported data plot and residuals plots for HPP primary inactivation models for microbial 722 inactivation at ambient temperature (20-25°C) and low acidity foods. 723 724 Figure 2. 725 Empirical secondary Gompertz model fit (solid line), secondary model standard deviation (dotted line), 726 and reported microbial inactivation data (color symbols) at different pressure-time combinations for milk 727 HPP treatments at ambient temperature (20-25°C). 728 729 Figure 3. 730 Logistic-exponential secondary Weibull model fit (solid line), secondary model standard deviation (dotted 731 line), and reported microbial inactivation data (color symbols) at different pressure-time combinations for 732 milk HPP treatments at ambient temperature (20-25°C). 733 31 734 735 32 736 737 33 738 739 34 740 er 35