Background/Objectives. Intracranial hemorrhages (ICHs) require immediate diagnosis for optimal clinical outcomes. Artificial intelligence (AI) is considered a potential solution for optimizing neuroimaging under conditions of radiologist shortage and increasing workload. This study aimed to directly compare diagnostic effectiveness between standalone AI services and AI-assisted radiologists in detecting ICHs on brain CT. Methods. A prospective, multicenter comparative study was conducted in 67 medical organizations in Moscow over 15+ months (April 2022–December 2024). We analyzed 3409 brain CT studies containing 1101 ICH cases (32.3%). Three commercial AI services with state registration were compared with radiologist conclusions formulated with access to AI results as auxiliary tools. Statistical analysis included McNemar’s test for paired data and Cohen’s h effect size analysis. Results. Radiologists with AI assistance statistically significantly outperformed AI services across all diagnostic metrics (p < 0.001): sensitivity 98.91% vs. 95.91%, specificity 99.83% vs. 87.35%, and accuracy 99.53% vs. 90.11%. The radiologists’ diagnostic odds ratio exceeded that of AI by 323-fold. The critical difference was in false-positive rates: 293 cases for AI vs. 4 for radiologists (73-fold increase). Complete complementarity of ICH misses was observed: all 12 cases undetected by radiologists were identified by AI, while all 45 cases missed by AI were diagnosed by radiologists. Agreement between methods was 89.6% (Cohen’s kappa 0.776). Conclusions. Radiologists maintain their role as the gold standard in ICH diagnosis, significantly outperforming AI services. Error complementarity indicates potential for improvement through systematic integration of AI as a “second reader” rather than a primary diagnostic tool. However, the high false-positive rate of standalone AI requires substantial algorithm refinement. The optimal implementation strategy involves using AI as an auxiliary tool within radiologist workflows rather than as an autonomous diagnostic system, with potential for delayed verification protocols to maximize diagnostic sensitivity while managing the false-positive burden.

Aim: to investigate the capabilities of artificial intelligence systems (AIS) in an inpatient setting, specifically in identifying signs of infiltration in chest radiographs (RG OGK). Materials and methods. The initial data were chest X-rays (CXRs) in a frontal projection of 1,537 inpatients aged 18 to 102 years, obtained from the Unified Radiological Information System of the Unified Medical Information and Analytical Service (ERIS EMIAS). The presence of infiltration on the chest X-ray was determined by a radiologist (Radiologist 1) and an expert radiologist (Radiologist 2) with more than 5 years of experience. Radiologist 2 annotated the studies in two stages: at the first stage, only the chest X-rays were provided; at the second stage, in addition to the radiological data, access to information about the patient’s condition was provided. Results. The accuracy rates for patients with a shifting state were 0.755 for AIS, 0.756 for Radiologist 1, and 0.856 for Radiologist 2, respectively. No statistically significant differences in accuracy were found between AIS and Radiologist 1 (p>0.05). When considering the scenario of using the AIS as a second opinion the accuracy for patients with a shifting state was 0.833, which is statistically significantly higher than the accuracy of Radiologist 1 and the AIS (p<0.05). Conclusion. AIS can be used as a second opinion when detecting signs of infiltration based on RG OGK data in a hospital setting. However, since the diagnostic accuracy of the AIS is still insufficient for practical use, further research is needed to train algorithms using both visual data and electronic medical record data.

Large language models (LLMs) are increasingly being used in radiology-related workflows, but their application to reference, regulatory, and methodological queries remains limited by hallucinations and the static nature of model knowledge. This study aimed to develop and evaluate a retrieval-augmented generation (RAG) system for radiologists designed to provide grounded responses to such queries. A knowledge base was created through a survey of practicing radiologists and expert validation of sources, resulting in a corpus of 1049 documents. The system incorporated structured document parsing, a two-level parent–child vector database, hybrid dense–sparse retrieval, reranking, and a local large language model. Performance was assessed through functional testing, automated LLM-as-a-judge metrics, and multireader expert evaluation by 16 radiologists using 400 technical queries. No hallucinations were detected in the 77-query functional testing set during expert review. On the full technical dataset, automated Contextual Precision, Contextual Recall, and Answer Relevancy were 0.735, 0.881, and 0.890, respectively. Expert evaluation showed high response accuracy (mean, 4.53/5) and high expert-assessed Contextual Precision (0.886). Inter-expert agreement was substantial to excellent for most Likert-scale criteria. These findings suggest that a hierarchical RAG architecture can provide reliable access to radiology-specific reference information, although external validation and automated updating of the knowledge base remain necessary.

Objectives. The article presents a study of heart rate variability using multivariate discriminant analysis. Representing an effective statistical method of classification, discriminant analysis can be used to divide objects into groups based on differences in the parameters characterizing these objects. The effectiveness of multivariate discriminant analysis, which is actively used in medicine to diagnose cardiovascular pathologies, is due to the wide range of analyzed parameters: statistical, spectral, and autocorrelation. The aim of the work is to identify the parameters of variational pulsometry, which provide the best distinction between healthy patients and patients with arrhythmia, by means of discriminant analysis.Methods. The durations of cardiac intervals of patients aged 63–72 years, which had been placed in the open database of biomedical signals PhysioNet.org, were used as initial data. When selecting the arguments of the discriminant function, priority was given to parameters that were weakly correlated with each other, had a normal distribution, and differed between healthy and ill patients. The statistical significance of differences between the parameters of the two groups was tested using Student’s t-test and Mann–Whitney U test.Results. Two discriminant functions were obtained: the first depended on three time-domain parameters, while the second included one spectral and one autocorrelation parameter in addition to time-domain parameters. In both cases, the average values of the discriminant function for healthy and sick patients were calculated. The statistical significance of differences in the average values of the discriminant function in the two groups was investigated using Student’s t-test.Conclusions. The values of the first discriminant function are shown to differ insignificantly between healthy and sick patients, while the inclusion of autocorrelation and spectral parameters in the number of arguments of the discriminant function provides pronounced and statistically significant differences between patients of the two groups. Thus, the high significance of spectral and autocorrelation parameters in arrhythmia diagnosis was demonstrated.