Calibrated two-threshold rejection option classification of shape-index extremes in quail eggs using weight and shell color

Hark B. D., ALAŞAHAN S., ATEŞ C. T., ŞAMLI M. Ş.

MEDYCYNA WETERYNARYJNA-VETERINARY MEDICINE-SCIENCE AND PRACTICE, cilt.82, sa.7, ss.335-341, 2026 (SCI-Expanded, Scopus)

Özet

The shape index (SI) is an indicator that quantifies egg shape and is associated with quality/hatching outcomes. However, direct measurement of SI in the field requires time and labor. The aim of this study is to develop and validate a probability-calibrated, two-threshold rejection option policy that can distinguish SI outliers (LOW/HIGH) in quail eggs using only weight and shell color, without measurement. In a dataset of 1,031 eggs (color: brown 61.9%, grayish white 38.1%), SI extremes were defined empirically (Q(0.25) approximate to 67.20, Q(0.72) approximate to 76.20). A classifier based on Hist Gradient Boosting was built using weight + color inputs. The probabilities obtained in the out-of-fold (OOF) scheme were corrected by sigmoid calibration. Two probability thresholds p(lo), p(hi) were selected on the OOF distribution through a coverage-targeted search, and the final model was refined in two phases by "snapping" to the model's discrete probability support points. The final policy produced approximately 61.4% coverage within the extreme tails with p(lo) approximate to 0.527 and p(hi) approximate to 0.761, while the accepted subset provided balanced accuracy (BA) approximate to 0.72 (95% CI [0.67 - 0.77]). The proportion of outliers in the entire data was 53.1%, and the expected overall automation (acceptance & times; outlier rate) was calculated to be approximate to 32.6%. The reject region made the risk-coverage balance controllable by directing uncertain samples to measurement. Calibrated probabilities, OOF threshold setting, and a two-threshold rejection policy can reliably identify SI outliers based solely on two observable phenotypes. This approach provides a scalable pre-screening layer that reduces measurement burden in incubation and quality control lines. It can be further enhanced in the future with microstructure and line/age-stratified designs.