P-225 Trustworthy AI algorithm for embryo ranking

Abstract Study question Deep-learning algorithms are known to be non-robust: can the variability and inconsistency of AI reduced in embryo selection? Summary answer We (measured on different tasks like rotations brightness changes) by 86% while preserving their quality. What is already methods generally non-robust, i.e., decisions change with even slight modification input data. Current solutions for scoring not robust - example rotating image results a score most market. Despite this fact expressed concerns embryologists, there no other publications focusing problem variance used IVF. Most measure accuracy, sensitivity, specificity, ROC AUC; metrics. design, size, duration The data-set was collected within multiple clinics using various devices. It contains 34,821 embryos (4,510 were transferred pregnancy results), represented time-lapse videos or images. This gives 3,290,481 frames at maturity levels. From 925 randomly selected chosen as test set. modified that supposed algorithm. measured scores given our Participants/materials, setting, have considered seven modifications images should influence scoring: • Rotations (10 angles); Brightness Contrast modifications; Substitutions Frames (from monitoring taken from 2 hours interval); Blur (Generalised Normal filter); Gaussian Noise; Blur; Sharpening. several techniques reduce deep neural network model (architecture commonly selection): Ensemble (of models cross validation); Test time augmentation (TTA); Robust training. Main role chance In order we following method. First, stretched standard uniform distribution. words look which percentile lies. way range normalised thus compared. Second, train EMBROAID augmented data includes all above modifications. Third, compute mean dropped (0.0055 0.0008) across individual drops modifications: Rotations: 77% (0.009 -> 0.002), Contrast: 81% (0.0036 0.0007), Substitution Frames: 76% (0.0076 0.0019), 94% (0.012 0.0008), Noise: 96% (0.0049 0.0002), Blur: 95% (0.0052 0.0003), Sharpening: (0.0015 0.0003). significance tested Wilcoxon Rank Sum giving p-value < 0.01 Finally, stress these obtained without any loss AUC metric. algorithm both original test-set. Both achieved an 0.66 (CI 0.63-0.69) test-sets. Limitations, reasons caution Further work needs done extend set possible augmentations Wider implications findings Increased reliability selection. obtain consistent over wide Trial registration number applicable