Sergei Sergeev

GGRC, Georgia

Title: Advancing Reproductive Medicine Through Integrated AI Solutions

---

Biography

Sergey Sergeev holds a PhD in Embryology and Developmental Biology from Moscow State University and is an ESHRE Senior Clinical Embryologist. Since 2008, he has served as the Embryologist and Laboratory Director at the IVF and Reproductive Genetics Center in Moscow, Russia. In 2023, he expanded his expertise to Georgia, where he became the Embryology Laboratory Director at the Georgia German Reproductive Center (GGRC) in Tbilisi.

Additionally, Sergeev supervises embryology laboratories at LLC “Indigo Invitro” in Tbilisi, Georgia, and LLC “Persona-Clinic” in Almaty, Kazakhstan.

A leading expert in IVF laboratory innovation, he has played a pivotal role in the development of hardware and software systems for laboratory identification and witnessing, automated key performance indicator (KPI) calculations, data processing, internal and external quality control, and risk management. His work also includes expert external audits and data analytics, ensuring laboratory excellence and regulatory compliance.

Abstract

The integration of Kolmogorov-Arnold neural networks (KAN) architectures with transformer models presents a cutting-edge approach for complex decision-making tasks, in in vitro fertilization (IVF) treatment. By incorporate domain-specific rules and key performance indicators (KPIs) alongside the powerful representation learning capabilities of transformers, which utilize attention mechanisms and long-range dependencies, this approach offers an accurate and explainable framework for IVF outcome prediction with conformal calibration to obtain real success probabilities in individual treatment cycle.

We conducted a retrospective analysis of 15,779 IVF protocols. The neural network pipeline was designed in DataSpell 2024.2.1 IDE with Tensorflow 2.15.0 and Keras 2.14.0 libraries with 580000 learning parameters for transformer classifier. An analysis of prediction errors was conducted using area under the receiver operating characteristic curve (AUC), precision-recall curve (PRC), Brier score, Expected calibration error (ECE), and Maximum calibration error (MCE).

Our new predictive pipeline demonstrated AUC = 0.73 ± 0,02, PRC = 0.68, Brier score = 0,20 with ECE = 0,06 and MCE = 0,12 after calibration in patient’s subpopulations from 6 different countries. Its combination with the KPI assessment system not only provides a real probability of pregnancy based on laboratory data but also explains which factors from the patient’s history or laboratory indicators contributed most to the prediction. Our new augmented intelligence shows similar performance to ultra-precise neural networks designed for static image analysis (AUC = 0.74), including next-generation neural networks based on genetic algorithms (AUC = 0.77), and complex CNN+MLP models (AUC = 0.75–0.79) but with lower errors in predictions.