Development and evaluation of Nile tilapia (Oreochromis niloticus) body composition models

Mathematical models can be used as an alternative to conventional analytical methods (AOAC), allowing to indirectly estimate the body composition of fish without, in some cases, being necessary to sacrifice animals. To develop models with high predictive capacity, in addition of having access to representative data, it is important to use calibration methods that minimize the estimation of the generalization error. In this work, Nile tilapia whole-body composition data were collected from 138 scientific publications, covering fish from 0.01 g to 1470 g. Predictive models were obtained for each body component using different combinations of models and calibration methods. The different combinations were evaluated through cross-validation approaches in order to select the models with the best predictive capacity. Such models were compared against other published Nile tilapia body composition models, using an independent validation dataset. The results show that model predictions are greatly affected by the type of model, calibration method and amount of calibration data available. Models calibrated under the assumption of multiplicative error had better predictive capacity than those calibrated assuming additive error, which indicates that, in this particular case, performing regression on log-transformed data, even for isometric models, is advantageous. From the models tested, the ones with the best predictive capacity are the allometric model (al_mu; calibrated assuming multiplicative error) and a robust hybrid model (hyb_rob; ensemble of isometric and allometric models, calibrated assuming multiplicative error using a robust regression method), with both having good prediction capacity when compared with models published by other authors. Although the results obtained support the hypothesis that Nile tilapia body composition is essentially allometric, isometric models can also potentially be used without much performance loss, if they are calibrated assuming multiplicative error (i.e., using log-transformed data).