Surgical Data Science (SDS) is a discipline aimed at improving the quality of surgical care through the acquisition, organization, analysis, and modeling of surgical data. Progress in SDS is driven by the co-evolution of three components: intraoperative data collection systems, the datasets used to train AI algorithms, and the algorithms themselves. This study proposes an original conceptual timeline of this co-evolution, spanning from the mid-2010s - the time when SDS became institutionalized as a discipline - to the present day. It identifies three stages: the era of proprietary integrated operating rooms (approx. mid-2010s–2019), characterized by closed data collection systems; the era of overcoming the limitations of proprietary systems (approx. 2019–2024); and the era of interoperability and ambient AI (approx. 2024–present). A parallel evolution is observed in the datasets generated at each of these stages—transitioning from monomodal and multichannel to large-scale multimodal datasets. It is shown that each stage was initially driven by data acquisition systems, which set the limits for synchronization quality, modality composition, and the datasets that could be generated, whereas the requirements of AI algorithms guided the architectural evolution of these systems. The literature search was conducted across PubMed, IEEE Xplore, arXiv, the proceedings of the MICCAI and CVPR conferences, manufacturers' official websites, and technical documentation. Out of 127 initially identified publications, 68 were included in the study following screening and full-text analysis, supplemented by 29 sources comprising technical documentation and manufacturers' official websites (totaling 97 sources). Key factors for future progress will include the adoption and implementation of medical device interoperability standards; the development of multimodal datasets, foundational AI models, and multi-agent architectures; and their integration into the operating room environment.

Missing data is one of the key characteristics of Real-World Data and they are an integral part of the Real-World Evidence studies (RWE studies). They arise at all stages of data collection in both prospective and retrospective designs. This review provides a comprehensive analysis of the issue of missing data in the RWE studies at all stages of the research process – the initial conceptualization of the study to the interpretation of results and their presentation in scientific publications.

The first part of the review examines the theoretical foundations: the term "missing data" in the regulatory documents of ICH, FDA, EMA, EAEU and its slightly different definitions, the role of the data source in the occurrence of missing data, the classification of missing data mechanisms according to Rubin (MCAR, MAR, MNAR) and their significance for selecting statistical methods.

The second part of the review focuses on a systematic approach and outlines eight stages of handling missing data. It is emphasized that effective missing data management begins well before statistical analysis — at the study design stage. Effective missing data management is an end-to-end, proactive process integrated into the RWE study lifecycle.

The conclusion of the report substantiates the need for a systematic approach to the issue of missing data as a crucial element in ensuring the reliability of RWE, which is directly relevant to the application of their findings within the healthcare system.

Aim. To evaluate the suitability of large language models (LLMs), deployed in a closed-loop environment, for predicting anxiety and depression scores on the Hospital Anxiety and Depression Scale (HADS) in a sample of Russian students.

Materials and methods. The empirical dataset consisted of de-identified responses from 946 first-year university students (669 female, 277 male) in Russia to 14 HADS items. For 11 models (DeepSeek-R1-32B, Qwen3-32B, T-pro-2.0/2.1, YandexGPT- 8B, GigaChat-20B, and others) deployed locally on GPUs, response probability distributions were extracted from token log probabilities. The experimental design included 5 prompt groups (textual instructions for the model), 2 presentation formats (structured dialogue and unstructured text), and 4 wording variations. Model-generated and empirical distributions were compared using Wasserstein distance (W); expected HADS subscale scores were additionally calculated.

Results. T DeepSeek-R1-32B demonstrated the best accuracy in reproducing distributions (W = 0.293); however, it consistently underestimated expected anxiety scores (5.6 vs. 7.7). The most accurate expected anxiety scores were produced by YandexGPTpretrain (7.2) and T-pro-2.0 (7.1). Model selection explained 47% of metric variance; persona detail and respondent sex had no significant effect (p > 0.05). The depression subscale was reproduced less accurately than the anxiety subscale.

Conclusions. Scores are robust to prompt variations: the level of detail in the persona description and the specific wording do not affect the outcome—a minimal description of the respondent suffices. Models trained in English favor English-language prompts, whereas for Russian-language models, language preferences are determined by the underlying architecture. The hypothesis regarding a systematic advantage for Russian-language models was not confirmed: quality is determined by domain adaptation rather than language. Practical application requires validation through experiments involving human subjects. The models do not replace clinical assessment by a specialist.

Relevance. According to the Ministry of Health of the Russian Federation, more than 6% of the country's population consults an otorhinolaryngologist annually due to ear diseases. Given the high workload of specialists, the accelerated implementation of artificial intelligence (AI) methods to assist in diagnosis is necessary. Objective. To develop a prototype AI-based medical system (AI MS) for diagnosing ear diseases from otoscopic images (OIs), achieving a diagnostic accuracy of at least 95% on the test set through transfer learning techniques applied to pre-trained deep neural networks (NNs).

Materials and methods. A labeled otoscopic image dataset representing five balanced diagnostic classes was selected and prepared from Kaggle.com. To effectively implement transfer learning of neural network models and prevent overfitting, a data preprocessing and augmentation pipeline was established. A custom web interface for NNM integration was developed, and the Grad-CAM method was applied to visualize diagnostic regions within the images.

Results. The CNN models selected for transfer learning were ResNet18, ResNet50, EfficientNet-B0, and EfficientNet-B2. Based on experiments using "soft" augmentation of the training set, the diagnostic accuracy for five ear disease categories—listed in the order mentioned—was 95.3%, 96.1%, 97.8%, and 99.7%, respectively. The ResNet50 model with soft augmentation of the training set for 3 classes showed an overall test data accuracy of 97%. Heatmaps visualizing the diagnostic results using the Grad-CAM method are presented.

Conclusion. Transfer learning, involving the adaptation of selected pre-trained architectures (ResNet18, ResNet50, EfficientNet-B0, and EfficientNet-B2) to the task of multiclass classification of ear diseases using balanced classes, enabled the achievement of the target accuracy for automated diagnosis. Conditions have been established for collaborative efforts with the Mogilev Regional Clinical Hospital (MRCH) to develop an AI-based medical system for diagnosing ear diseases.

A modified deep learning model for binary segmentation of renal pathologies in computed tomography (CT) scans has been developed. The relevance of this study stems from the need to improve the accuracy and efficiency of anomaly detection during primary diagnosis, particularly given the high workload faced by medical institutions.

The aim of the study is to create a universal system for the automatic detection and localization of renal pathologies without prior classification of their morphological type.

Methods. The model is based on a modified U-Net architecture featuring five levels in both the encoder and decoder. Key modifications included replacing max pooling with strided convolutions (stride of two), implementing group normalization, and incorporating group normalization and dropout regularization. The model was trained on a clinical dataset comprising over 400 contrast-enhanced multidetector computed tomography (MDCT) scans. Preprocessing included conversion of intensities into Hounsfield units, range-free normalization, and data augmentation. A hybrid loss function combining the Dice coefficient and binary cross-entropy was used during training.

Results. A weighted average Dice coefficient of 0.85 was achieved on an independent test set. The model demonstrates high segmentation accuracy across various types of pathologies, including cysts (including those classified as complex under the Bosniak system), solid tumors, and mixed structures. The model maintains robustness in the presence of anatomical anomalies (such as horseshoe kidney, renal ectopia) when dealing with small lesions (less than 5 mm in size).

Conclusions. The proposed model offers high versatilityand applicability in routine clinical practice. Its ability to detect a wide range of renal pathologies without prior classification enhances its clinical significance and suitability for screening purposes.

Application Area. The results can be utilized in oncourology for initial CT scan analysis, surgical planning, and training of medical professionals. Integrating the model into medical image archiving and communication systems for the automated preprocessing of imaging studies holds significant promise.

Aim. To compare deep learning architectures and classification schemes for dermoscopic images of skin neoplasms and to assess their generalizability when transferring from open international datasets to independent clinical datasets from Russian practice.

Materials and methods. Four architectures (ViT-B/16, Swin-S, ConvNeXt-S, EfficientNetV2-S) were compared in three schemes: binary (malignant / benign), single-stage four-class, and a two-stage cascade (binary triage with a tunable threshold – MEL / SCC / BCC differentiation). All models were initialized with ImageNet weights, trained on a combined dataset of publicly available ISIC Archive data, and evaluated on an internal hod-out sample and two clinical datasets (Melanoscope AI; Sechenov University). The significance of differences was assessed using the DeLong test, McNemar's test, and stratified bootstrapping.

Results. On the internal dataset the binary-stage ROC-AUC was 0.952–0.966; when applied to the most challenging external clinical dataset, this figure dropped to 0.797–0.893 and sensitivity fell to 0.53–0.67, while the calibration error (ECE) increased from 0.02 to 0.27–0.39, accompanied by an underestimation of malignancy. A performance deficit for ViTB/ 16 was statistically confirmed only at the binary selection stage; at the nosological differentiation stage, no architectural advantages were identified given the available number of malignant cases. The performance gain provided by the cascade approach was significant only for ViT-B/16.

Conclusion. Cascade decomposition with an adjustable threshold enables sensitivity control that is unattainable in singlestage classification. The persistent generalization gap highlights the need for external clinical validation and recalibration prior to implementation.

Medical institutions and clinics accumulate a significant amounts of data that can and should be used in planning clinical and operational activities. There are intuitive connections between influencing factors and target KPIs that cannot be described by functional dependencies. Such conditions give rise to a wide range of tasks that can be addressed using machine learning and artificial intelligence methods. This article examines the task of predicting blood loss during the surgical treatment of endocrine system disorders. Machine learning methods based on decision trees are used to estimate blood loss. The data for the study were provided by one of the world’s leading clinics of endocrine surgery from St. Petersburg.

Background. Analyzing cognitive states using electroencephalography (EEG) is a key task for brain-computer interfaces. Despite the success of deep learning methods, classical convolutional neural networks (CNNs) face significant limitations when processing non-stationary EEG signals: they struggle to adapt to local time-frequency features, and their decisions are difficult to interpret. Integrating the Kolmogorov–Arnold Network (KAN) architecture with wavelet transformations opens up new possibilities for improving the accuracy and interpretability of EEG analysis models by replacing linear weights with learnable nonlinear functions.

Aim. To study the impact of replacing conventional convolutional layers with layers based on the Kolmogorov–Arnold architecture utilizing various wavelet activation functions on the quality of identifying cognitive states of varying complexity.

Methods. Data from the open-access COG-BCI dataset were used, comprising recordings from 29 participants performing cognitive tasks N-Back (working memory) and MATB-II (multitasking). The original FFCLNetEEG architecture was modified by replacing CNN layers with KAN and WavKAN convolutional layers using various wavelets (Gabor, DOG, Mexican Hat, Morlet, Meyer, Shannon). Evaluation was performed using Accuracy, Precision, Recall, F1-score, and Specificity metrics with pairwise Wilcoxon tests with Bonferroni correction (p < 0,05).

Results. In the N-Back task, the FFCLNetEEG-WavKAN-DOG model showed the best performance achieving a mean accuracy of 0.790±0.059; this significantly outperformed FFCLNetEEG-KAN, FFCLNetEEG-WavKAN-Meyer and FFCLNetEEG-WavKAN-Morlet (p < 0,001 after correction). In the MATB-II task, the best performance was achieved by the FFCLNetEEGWavKAN-Gabor model performed best (mean accuracy 0.892±0.005). Analysis of three independent sessions showed high reproducibility of the results with correlation coefficients between sessions up to 0.653 for the best-performing models.

Conclusion. Integrating the Kolmogorov–Arnold architecture with embedded wavelet functions enables a reliable improvement in the classification quality of cognitive states based on EEG data. The choice of the optimal wavelet depends on the type of cognitive task: DOG for working memory (N-Back) and Gabor for multitasking (MATB-II). The findings have practical implications for the development of neuroadaptive interfaces and systems for cognitive monitoring and rehabilitation.

Introduction: to prospectively validate a previously developed machine learning model for predicting sepsis onset in this patient population.

Methods: Prospective validation was conducted in accordance with TRIPOD+AI recommendations using data from the Russian Intensive Care Dataset (RICD v3.0). The XGBoost-based model estimated the risk of sepsis onset on an hourly basis with a 6-hour prediction window, using demographic, clinical, laboratory, and vital parameters aggregated into dynamic time windows. Model performance was evaluated using the area under the receiver operating characteristic curve (AUROC) with a 95% confidence interval (CI), as well as diagnostic metrics at the optimal cutoff point.

Results: Validation was performed on a cohort of patients admitted to the ICU of the Federal Research and Clinical Center of Intensive Care Medicine and Rehabilitology between October 2024 and September 2025 (n = 103; 53,561 patient-hours). The median ICU length of stay was 32 (27; 38) days. The XGBoost model achieved an AUROC of 0.737, consistent with the internal testing results (AUROC 0.752; 95% CI 0.730; 0.767). The optimal cutoff (≥0.5651) yielded a sensitivity of 60.3%, specificity of 81.5%, and a high negative predictive value (97.4%). Although the model showed a high rate of false positives, it demonstrated strong clinical utility for ruling out sepsis within the next 6 hours. Moreover, a multilevel interpretation approach allows risk stratification across low, intermediate, and high probability zones of sepsis development.

Conclusion: Prospective validation confirmed the stability and reproducibility of the machine learning model for sepsis prediction in patients with prolonged ICU stays. The results support the feasibility of integrating the model into an electronic health system as a component of a decision support framework for subsequent real-world evaluation.

The article presents a situational algorithm for analyzing event sequences to predict postoperative time parameters – such as the duration of hospital and intensive care unit stays. It describes the construction of an event model using specialized neural language models. Algorithmic processing was performed using anonymized data from an electronic medical record archive and patient discharge summaries following cardiac surgery. Such approach enables the identification of operation clusters and the prediction of postoperative temporal parameters. The article presents an algorithm for constructing event sequences, describes the analysis method, and reports the results of the method's experimental validation.