P-205 A video-based system for extracting osmotic behavior and morphological dynamics: a foundation for predictive models in oocyte cryopreservation survival

Abstract

Study question Can video-derived osmotic and morphological data during vitrification protocols serve as a foundation for future predictive models in oocyte cryopreservation survival? Summary answer Dynamic parameters such as dehydration velocity, circularity, and area changes provide a foundation for future predictive models in oocyte cryopreservation survival. What is known already Artificial intelligence (AI) is increasingly used in assisted reproduction laboratories, primarily through time-lapse microscopy of embryo culture or static image analysis. These AI models achieve high predictive accuracy by training on large datasets but function as “black boxes,” providing results without biological explanations. Unlike other medical AI applications, those in reproductive medicine lack mechanistic validation and rely on correlations rather than causal insights. Furthermore, European law (Article 22, GDPR) prohibits decisions affecting human lives from being based solely on AI predictions. This highlights the urgent need for biologically interpretable, validated models that integrate dynamic physiological data into AI-driven reproductive assessments. Study design, size, duration A prospective study (October 2023–January 2024) analyzed 37 oocytes from 10 donors. Immature oocytes discarded 4 hours post-retrieval were cultured for 24 hours to promote maturation. Oocytes were exposed to equilibration solution of the vitrification protocol (1/2 concentration for 3 min, then 2/3 for 3 min). Video recordings captured osmotic responses using an inverted microscope and micromanipulator. A holding pipette stabilized each oocyte, while a covering pipette facilitated media transitions for precise exposure. Participants/materials, setting, methods A custom software pipeline analyzed videos to extract area change, circularity, and roundness. Oocyte dehydration and deplasmolysis velocities were quantified, along with circularity and roundness differentials. Descriptive statistics assessed data distribution, Pearson correlations identified variable relationships, and stepwise regression determined key osmotic response predictors. The dataset was structured for machine learning applications, allowing AI-driven models to refine oocyte selection criteria and optimize cryopreservation strategies in assisted reproduction. Main results and the role of chance The analysis of osmotic and morphological parameters extracted from video recordings revealed significant variability in oocyte responses during vitrification. Descriptive statistics showed that the minimum area averaged 0.797±0.041%, while the difference in area was 20.26±4.12%, indicating substantial volume reduction during dehydration. The minimum circularity was 0.847±0.033, suggesting shape alterations under osmotic stress. Correlational analyses identified key relationships between parameters. Dehydration velocity during phase 1 showed strong positive correlations with phase 2 dehydration velocity (r = 0.79, p < 0.01) and moderate associations with deplasmolysis velocity in phase 1 (r = 0.69, p < 0.05), indicating consistency in osmotic responsiveness. Differences in circularity correlated with changes in area (r = 0.52, p < 0.05), highlighting shape recovery dynamics. Exploratory regression suggested that dehydration velocity, area change, and circularity alterations contribute significantly to osmotic adaptation, making them promising candidates for machine learning models. In contrast, deplasmolysis velocity in phase 2 and roundness variations showed weak or inconsistent correlations, suggesting limited predictive value. These findings support the hypothesis that dynamic osmotic behavior provides meaningful features for oocyte viability prediction. Further validation with survival outcomes and AI-based modeling is needed to confirm the predictive power of these variables. Limitations, reasons for caution Small sample size and pilot-level analyses limit the generalizability of findings. Additionally, data extraction using an inverted microscope with a micromanipulator is a tedious process. To improve usability, we are adapting the system for stereomicroscope use, ensuring seamless integration into routine vitrification protocols currently followed in IVF laboratories. Wider implications of the findings This study provides a biologically meaningful dataset for AI training, moving beyond static images to assess oocyte behavior. By integrating dynamic osmotic responses, AI models gain mechanistic insights, improving transparency and regulatory compliance. This aligns with European law (GDPR, Article 22), ensuring AI-assisted reproductive decisions remain interpretable and evidence-based. Trial registration number Yes