Validation unit and factor encoding determine apparent prediction quality in small antimicrobial fermentation designs
Year: 2026
Type: article
Abstract: Small fermentation datasets can yield high apparent model fit without supporting prediction at experimental conditions that were never run. We audited four related central composite datasets comprising 78 numerical inhibition-zone records for Lactococcus lactis MT186647 and Limosilactobacillus fermentum MT186598, with 14 or 15 distinct conditions and six center records per dataset. 10 model configurations were evaluated using leave-one-record-out and leave-one-condition-out cross-validation, under a specified-design encoding and an encoding taken from the levels printed in the source tables. Condition-holdout predictions were scored both by individual records and by equally weighted condition means. All evaluation was internal to one experimental programme rather than external validation. The simulated-annealing fuzzy ensemble tuned against a conditional PRESS-derived root-mean-square error reached a mean apparent R 2 of 0.842, but its mean record-wise Q 2 was −0.050 and its condition-holdout Q 2 was −0.464 under record weighting and −0.686 under condition-mean weighting. The two tree ensembles achieved the highest average held-out scores, although performance varied by dataset and does not establish the superiority of any algorithm family. Withholding the center records increased the condition number of the quadratic model matrix from 3.50–3.59 to 78.67–88.43 in designs complete as specified as well as incomplete ones and produced center predictions of 38.99–49.64 mm against observed center means of 11.29–12.85 mm. For physical L. lactis , taking the printed coordinates literally reduced that prediction from 49.64 to 20.63 mm without resolving which levels were used. These results argue for matching the validation unit to the intended prediction target, documenting factor coordinates, and reporting every condition-holdout fold. Predicted inhibition-zone diameters do not establish antimicrobial concentration, new fermentation optima, or preservation efficacy in food.
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: gold
Estimated APC paid: $2,353
OpenAlex ID: W7225614127