High-risk medical devices approved through the US Food and Drug Administration (FDA) premarket approval (PMA) pathway may be supported by limited premarket clinical evidence, leaving uncertainties about safety. Whether device, premarket clinical evidence, or regulatory characteristics are associated with subsequent device recalls is unknown. To examine the association between device, premarket clinical evidence, and regulatory characteristics and serious recalls among high-risk therapeutic devices. This cross-sectional study reviewed 193 high-risk therapeutic medical devices approved through the FDA's PMA pathway between January 1, 2014, and December 31, 2023. Data were analyzed between August 2025 and December 2025. Device characteristics, premarket pivotal clinical study characteristics, and regulatory characteristics. The main outcome was occurrence of serious recall (Class I or II) and Class I (highest severity) recall. From 2014 to 2023, the FDA approved 193 original high-risk therapeutic medical devices through the PMA pathway. As of July 31, 2025, a total of 68 (35.2%) were subject to at least 1 serious recall, with a median (IQR) of 2.6 (1.5-4.6) years from approval to first recall; 20 (10.4%) devices were subject to at least 1 Class I recall. Life-supporting or life-sustaining devices more often had serious recalls (28 of 60 [46.7%] vs 40 of 133 [30.1%]; P = .02) and Class I recalls (14 of 60 [23.3%] vs 6 of 133 [4.5%]; P < .001) than non-life-supporting and life-sustaining devices. Cardiovascular devices more often had Class I recalls than noncardiovascular devices (16 of 99 [16.2%] vs 4 of 94 [4.3%]; P = .004). Class I recalls differed by pivotal study design: single-arm with no controls, 3 of 9 (33.3%); single-arm with nonconcurrent controls, 11 of 89 (12.4%); and multiple-arm with concurrent controls, 6 of 95 (6.3%) (P = .03). Devices with required postapproval studies had more serious recalls (56 of 140 [40.0%] vs 12 of 53 [22.6%]; P = .04) than those without. Other characteristics were not associated with recall likelihood. In this cross-sectional study, more than one-third of high-risk therapeutic medical devices were subject to serious recalls. Because device, premarket clinical evidence, and regulatory characteristics were not consistently associated with serious recalls, robust postmarket surveillance is needed to ensure patient safety.
Today's external manufacturing landscape demands more than retrospective audits and subjective traffic-light risk scores. With many CMO and API partners and fragmented oversight processes, quality organizations often discover problems only after they have escalated into supply disruptions or regulatory actions.This session will explore how predictive intelligence can modernize supplier oversight. By unifying structured and unstructured data sources - inspections, deviations, recalls, enforcement documents, audits, batch records, changes - into a single system of intelligence, we can surface early signals of risk. Hazard models and AI-based risk scoring now make it possible to quantify the likelihood of recalls, enforcement actions, or supplier disruptions weeks to months in advance.Real-world use cases will illustrate how large pharma organizations are reducing surprises, better targeting audits, and improving supplier selection decisions. The presentation will highlight the shift from static dashboards to near real-time monitoring, automated alerts, and data-driven next-best-action recommendations.
The impacts of agricultural commercialization on nutrition are highly contested. Agriculture-nutrition linkages are complex and context-dependent and may result in trade-offs between different dietary outcomes. This study compared dietary patterns between children in oil palm and non-oil palm-adopting Indigenous households in two oil palm contexts in Indonesia: "plasma" smallholder oil palm in West Kalimantan and oil palm plantation labor in Papua. Using 24-h dietary recall data from children aged between 12 and 59 mo in West Kalimantan (696 children; 963 recalls across two seasons) and Papua (345 children), we compared quantities of foods consumed at the food group level between children in oil palm and non-oil palm Indigenous households. We also compared the probabilities of consuming ultraprocessed foods. Children living in oil palm-adopting households in West Kalimantan consumed lower quantities of cereals and dark green leafy vegetables (DGLVs), more dairy and eggs, and were more likely to have consumed processed meat and sugar-sweetened beverages. In Papua, children living in households engaged in oil palm plantation labor consumed lower quantities of tubers and fish, but more cereals, DGLVs, and legumes. They were also more likely to have consumed ultraprocessed refined grains and bakery products. Associations between agricultural commercialization and children's diets are context-dependent and may reinforce local dietary transitions. These may include increased consumption of some healthy foods in some contexts, but also be associated with diets linked to negative health outcomes, including reduced consumption of some nutrient-rich foods and increased consumption of ultraprocessed foods.
Gestational diabetes mellitus (GDM) presents significant risks to maternal and offspring health. Modifiable factors like dietary quality are important for prevention. The Dietary Obesity-Prevention Score (DOS) is a validated tool for assessing obesity-related dietary quality. Nevertheless, its association with GDM remain unclear. This study aimed to examine the relationship between DOS and GDM risk. This prospective cohort included 1,894 pregnant women across nine medical centers in China. Dietary intake was assessed by 3-day 24-h recalls. The DOS was derived from 6 food groups. Multivariable binary logistic regression, component analysis, subgroup analysis, and restricted cubic splines (RCS) were used to evaluate associations. In the overall population, DOS showed no significant association with GDM (OR = 0.944, 95% CI: 0.879-1.015). The interaction between DOS and energy intake status was not significant (p for interaction = 0.193). Exploratory subgroup analyses showed that among women with inadequate energy intake, higher DOS was associated with reduced GDM risk (OR = 0.867, 95% CI: 0.773-0.973). Likelihood ratio tests showed that excluding fruit from the full model worsened the model fit (Δ-2LL = 16.091, df = 1, p < 0.001). Component analysis suggested a possible inverse association between fruit intake and GDM risk (OR per 50 g/1,000 kcal = 0.905, 95% CI: 0.817-1.001, p = 0.053). Fruit intake was associated with lower GDM risk in the middle DOS tertile (OR = 0.783, 95% CI: 0.620-0.990, p = 0.041) and showed a borderline significant association in the high DOS tertile (OR = 0.828, 95% CI: 0.686-1.000, p = 0.049). In the inadequate energy intake subgroup, RCS analysis revealed a non-linear association between fruit intake and GDM risk (p for non-linearity = 0.048), with risk decreasing to a nadir at approximately 400 g/day, after which it stabilized. The 95% CI widened beyond 600 g/day due to sparse data, and these inflection points should be interpreted as exploratory descriptive features rather than precise thresholds. In pregnant women with inadequate energy intake, higher dietary quality was associated with reduced GDM risk. Fruits may play a role, although these findings are exploratory and require confirmation in independent cohorts.
Early and accurate prediction of rheumatoid arthritis (RA) is critical for improving patient prognosis; however, existing approaches rely excessively on single autoantibody markers, neglect the systematic predictive value of routine hematological parameters, and lack mechanistic interpretability. In this study, 500 patients attending a rheumatology outpatient clinic were enrolled, and 29 routine laboratory features were collected. Anti-CCP positivity and early RA onset within 12 months were defined as dual binary prediction targets. Five traditional machine learning models (logistic regression, random forest, gradient boosting, SVM, and KNN) and five deep learning models (MLP, ResNet, Transformer, AE-Classifier, and TCN) were systematically compared using six evaluation metrics: accuracy, AUC-ROC, F1-score, precision, recall, and Matthews correlation coefficient (MCC). A four-dimensional SHAP explainability analysis was subsequently applied to the best-performing deep learning model. Logistic regression achieved the best overall performance (Accuracy = 0.848, AUC = 0.857, F1 = 0.910, MCC = 0.441). Among deep learning models, the Transformer performed relatively well (AUC = 0.812), whereas ResNet and TCN exhibited severe class collapse (MCC ≈ 0). SHAP analysis identified ESR (Mean|SHAP| = 0.097) and CRP (0.090) as the most important positive predictive drivers, and albumin (ALB, 0.062) as the key protective factor, together forming a core biomarker triad for early RA risk prediction. Dependence plots further revealed the synergistic interaction between ESR and CRP, as well as a non-linear protective threshold effect of ALB. Individual waterfall plots confirmed close alignment between model decisions and clinical pathological mechanisms. The proposed machine learning pipeline based on routine laboratory parameters can effectively predict early RA risk, and the SHAP explainability analysis transforms the model "black box" into clinically readable decision rationale, providing evidence-based support for optimizing early screening strategies in rheumatology.
Heart disease prediction utilizes machine learning methods to estimate the probability of a person developing heart disease, considering various medical and lifestyle factors. These factors commonly include age, gender, cholesterol levels, blood pressure, smoking habits, diabetes status, family history, and physical activity. In real-world medical datasets, patient information may be missing or incomplete, which can hinder accurate predictions. Heart disease datasets often have more negative samples (people without heart disease) than positive ones. This imbalance can lead to biased models that predict the majority class (no heart disease) more accurately. Hence, this paper presents the Stacked Wrapper Attribute Machine Learning Model (SWA-ML) for the prediction and classification of heart disease. The proposed SWA-ML model incorporates the multi-class estimation of features in the heart disease dataset. The SWA-ML utilizes the stacked wrapper model for the estimation of features and classification. The machine learning model integrates the estimated features for classification. The model integrates three distinct datasets, resulting in a combined dataset of 1800 instances, encompassing key attributes such as age, sex, cholesterol levels, and blood pressure. Wrapper-based feature selection and stacking ensemble architecture are used in the SWA-ML framework to forecast heart disease utilising heterogeneous tabular clinical datasets. After attribute alignment, duplication removal, missing-value treatment, categorical encoding, feature scaling, and class balancing, three public heart disease datasets were integrated. Before stacked ensemble learning, wrapper-based feature selection approaches such RFE, Forward Selection, Backward Elimination, and Genetic Algorithm optimisation were compared. For robust model assessment, stratified train-test partitioning, k-fold cross-validation, and hyperparameter optimisation were used. Classification accuracy, precision, recall, F1-score, and AUC were consistently higher for the stacked wrapper configuration than for individual feature-selection procedures. Ablation experiments, statistical significance testing, calibration evaluation, and comparing with existing methodologies confirm the framework's efficacy. SWA-ML appears to be an effective and interpretable decision-support mechanism for structured clinical data-based heart disease risk prediction.
Craniopharyngioma is a rare pediatric brain tumor that is often treated with surgical resection and radiation therapy, all of which can have significant impacts on cognitive abilities. The objectives of the current study were to characterize the neurocognitive functioning of pediatric patients that have been treated for craniopharyngioma and examine the impact of medical, demographic, and treatment variables on neurocognitive outcomes. We completed retrospective chart reviews to collect medical, demographic, and neurocognitive data for 27 patients that were treated for craniopharyngioma and underwent follow-up neuropsychological evaluation. Bivariate correlations revealed that preoperative hydrocephalus was associated with worse performance in IQ, working memory, verbal fluency, learning memory, recall memory, and memory recognition. Patients with hydrocephalus did significantly worse than those without hydrocephalus in almost all cognitive domains. Those who underwent surgical resection via the endoscopic endonasal approach (vs. craniotomy) did significantly worse on recall memory and processing speed. No significant correlations or differences were observed between neurocognitive outcomes and age at treatment, extent of surgical resection, time elapsed since surgery, new hypothalamic injuries, new endocrinopathies, or radiation treatment history. These data suggest hydrocephalus is a key predictor of neurocognitive outcomes among patients treated for craniopharyngioma, with surgery factors also contributing to outcomes. Larger studies are needed to understand the individual risk factors for cognitive impacts within this patient population.
Trigger point injection (TPI) therapy is widely used for masseter myofascial pain syndrome (MPS), yet outcomes vary substantially. Individualized prediction tools are lacking, often leading to trial-and-error treatment selection. To develop, externally validate, and generalize ensemble machine learning models for predicting composite treatment success following masseter TPI, and to deploy a web-based clinical decision support system (CDSS). This multicenter study included 1,181 patients with DC/TMD‑diagnosed masseter MPS treated with one of six injectable modalities. Baseline variables included pain (VAS), maximum mouth opening (MMO), and oral health‑related quality of life (OHIP-14). Composite treatment success was defined as simultaneous clinically meaningful improvement at 3 months: VAS reduction ≥ 2 points, MMO increase ≥ 5 mm, and OHIP-14 reduction ≥ 5 points. Random Forest (RF) and XGBoost models were trained on an internal cohort and externally validated on a geographically independent cohort. Performance was assessed using ROC‑AUC, precision‑recall AUC (PR‑AUC), calibration, decision curve analysis (DCA), and SHAP interpretability. Overall composite success rate was 43.1%. Internal validation ROC‑AUC was 0.914 (RF) and 0.888 (XGBoost); external validation ROC‑AUC was 0.771 and 0.787, respectively. External PR‑AUC values were 0.759 (RF) and 0.761 (XGBoost). Both models showed good calibration and positive net benefit on DCA across clinically relevant thresholds. SHAP analysis identified baseline MMO, OHIP-14, age, pain intensity, and injectable modality as the most influential predictors, with consistent rankings across models and cohorts. The validated XGBoost model was deployed as a web‑based CDSS. Machine learning models demonstrated good ability to predict multidimensional treatment success following masseter TPI. Baseline MMO, OHIP-14, age, pain intensity, and injectable modality were the strongest outcome determinants. External validation, SHAP interpretability, and DCA support model robustness and potential clinical utility for personalized treatment planning in MPS. This tool may support treatment selection, reduce ineffective interventions, and improve patient outcomes in myofascial pain management.
Against the backdrop of increasingly diversified and concealed forms of stock market manipulation, approaches based solely on trading data or financial indicators face growing limitations in complex and information-intensive market environments. To assess stock market manipulation risk, this study constructs a firm-month level multi-source panel dataset by retrospectively labeling violation periods at the monthly frequency based on manipulation cases sanctioned by the CSRC (China Securities Regulatory Commission). The dataset integrates corporate disclosures, investor sentiment derived from online public opinion, and market trading characteristics. A supervised learning framework that fuses textual representations and numerical features is then employed to generate manipulation risk probabilities, supporting risk ranking and tiered screening in regulatory applications. Empirical results show that the fusion model consistently outperforms single-source baselines, achieving an AUC of 0.8811 and a PR-AUC of 0.6943, along with substantial improvements in Recall@10% and Recall@20% for high-risk screening. These findings indicate that multi-source information exhibits complementary effects in manipulation risk assessment and enables effective characterization of joint anomalies along the "information disclosure-sentiment reaction-trading behavior" chain. Theoretically, this study highlights the complementary role of heterogeneous information sources, including disclosure, sentiment, and trading-related signals, in characterizing manipulation risk. In practice, it provides a feasible data-driven pathway for risk monitoring and tiered regulatory screening.
Discovering hazardous pollutants currently relies on the tedious and often inaccurate structural identification step, required for further toxicity and exposure studies. Here, we propose and validate a workflow for prioritizing the detected features prior to structural elucidation. The proposed approach relies on the priority score (environmental concentration divided by lethal concentration) serving as a proxy of the risk and is deduced solely from mass spectrometric data. The workflow integrates two machine learning approaches, MS2Tox and MS2Quant, that predict toxicity and ionizability of unidentified molecular features, respectively, and the relative risk-based ranking of detected features does not require any additional standards to be measured in the same run, allowing the application to digitally frozen data. Validation by using priority score for classifying 23 chemicals with available risk quotient values into high and low risk categories yielded recall value of 0.33 to 0.81, precision of 0.08 to 0.50, and accuracy of 0.52 to 0.81, depending on the acquisition mode and fish species. Applying the developed workflow to wastewater effluent prioritized 13-19% of features with predicted fingerprints as "precautionary risk features" with a risk quotient ≥1 based on the lower limit of the 95% prediction interval. All prioritized features were subject to spectral library matching, with 12% of the features yielding level 2 identification. Features categorized as high-risk were further subject to structural annotation using in silico identification tools SIRIUS+CSI:FingerID and MetFrag. While plausible candidates were suggested, the in silico tools disagreed in the top 10 suggested structures, highlighting structural assignment as a bottleneck in risk estimation. This work investigates the usefulness of applying data-driven feature prioritization prior to identification.
Long-read sequencing enables improved genome inference but remains challenged by high error rates that lead to excessive false-positive (FP) variant calls, particularly for small INDELs in complex genomic regions. We present VCboost, a deep learning-based post-calling framework designed to reduce FP in single-nucleotide polymorphism (SNP) and indel detection from long-read sequencing data. VCboost extracts discriminative features from pileup reads and consensus sequences and employs a dedicated filtering model integrating convolutional and recurrent neural networks with residual connections and multi-head attention. Evaluated on multiple human long-read datasets, VCboost consistently improved variant calling performance over Clair3, achieving a 5%-8% increase in precision and a 2%-5% gain in F1-score for INDELs, with minimal recall loss. For SNPs, VCboost improved precision by 4%-8% and F1-score by 2%-5%, with negligible impact on recall. Substantial performance gains were observed in difficult-to-map regions, including low-mappability and segmental-duplication regions, where SNP precision increased by up to 17.6%. Overall, VCboost effectively enhances the accuracy of long-read variant calling while preserving high sensitivity, offering a robust solution for variant detection in challenging genomic regions.
This research aims to develop an efficient model for Traffic Sign Recognition (TSR) with uncertainty quantification. The process involves pre-processing, sign localization, TSR, and criticality detection. After pre-processing with a Gaussian filter, sign localization is carried out using Segmentation U-Net (SegU-Net). Following TSR, a Deep Q Network (DQN) is trained using Gradient Descent-Teamwork Optimization Algorithm-based Deep Q Network (GD-TOA based DQN), which incorporates Gradient Descent model into Teamwork Optimization Algorithm (TOA), is used to classify traffic signs into three categories: mandatory, cautionary. and informatory. After classification, the system revalidates critical signs using Mahalanobis distance to compute confidence levels. A significant drop triggers a flag for human review if it has a high uncertainty. Empirical findings demonstrate that presented methodology achieved an accuracy rate of 97.9%, a precision rate of 98.2%, a recall rate of 98.3%, F-measure of 98.2%, specificity of 97.5%, and balanced accuracy of 97.9%. The results have proven that proposed model is a very promising technique for criticality detection.
Accurate and efficient three-dimensional visualization of cerebral vasculature is essential for clinical evaluation; however, manual vessel extraction from time-of-flight (TOF) magnetic resonance angiography angiography (MRA) is time-consuming and operator-dependent. This study aimed to develop a deep learning-based cerebrovascular segmentation model and an automated vessel extraction method, and to evaluate their accuracy, volumetric reliability, and impact on volume rendering (VR) workflow efficiency. A 3D U-Net-based vessel segmentation model was trained using TOF-MRA images. Automated vessel extraction was performed by dilating predicted vessel regions by one voxel. Forty-eight intracranial aneurysm cases were analyzed. Segmentation performance was evaluated using the Dice similarity coefficient (DSC), normalized surface Dice (NSD); tolerance = 1 mm), and centerline distance (CLD). Inter-rater reliability was assessed using DSC between independently generated vessel masks in a subset of the dataset. Aneurysm volumes from original and vessel-extracted images were compared using equivalence testing with a 1% margin and two one-sided tests (TOST). VR image creation time was measured by 12 radiological technologists. The DSC between independently generated vessel masks was 0.916. The DSC, recall, and precision of dilated vessel masks were significantly higher than those of non-dilated masks (p < 0.0001). The NSD was 0.982 ± 0.015, and the CLD was 0.196 ± 0.182 mm. Aneurysm volumes showed strong correlation (r = 0.999) with a small mean absolute error (MAE) (0.0915 mm³), and equivalence by TOST (p < 0.001). VR image creation time was significantly reduced (p = 0.0130). The proposed method enables accurate, reproducible, and time-efficient generation of cerebral vascular VR images, suggesting its potential utility in clinical practice.
Immune memory responses are rapid and qualitatively distinct from primary responses. They typically develop in the presence of antigen-experienced memory T and B cells and preexisting antibodies. Although the contribution of T and B cells to recall responses is well defined, the contribution of antibody "memory" and the mechanisms by which preexisting antibodies modulate the development of germinal center (GC) and plasma cell responses is not precisely understood. Here, we report on mechanisms that mediate antibody enhancement of GC and plasmablast (PB) compartments, and the parallel process by which antibodies change the affinity threshold for B cell recruitment into immune responses. The data indicate that antibody-mediated enhancement of GC and PB responses is Fc gamma receptor (FcγR) dependent and largely complement receptor 1 and 2 (CR1/2) independent. In contrast, the reduction in the affinity threshold for GC entry is independent of both FcγRs and CR1/2.
Systematic literature reviews (SLRs) are central to evidence generation in health economics and outcomes research, but traditional SLR workflows are labor-intensive. Artificial intelligence (AI) is increasingly being explored to support evidence synthesis tasks. To describe methods used to assess AI performance in screening and data extraction in SLRs, and to provide recommendations on how to ensure appropriate use of AI in literature reviews. We conducted an AI-assisted SLR that mirrored a traditional SLR performed by humans only and assessed the performance of the AI tools employed. AI was used to screen titles/abstracts, screen full texts, and conduct data extraction. Using the traditional SLR as the benchmark, AI performance for title/abstract screening was evaluated in terms of accuracy, recall, and precision at predefined screening milestones, followed by the accuracy assessment of full-text screening by AI. Data items extracted by AI were verified by humans and categorized as correct, incomplete, missing, incorrect, or requiring human checks. The average accuracy rate calculated on the basis of correct data items extracted was used as an indicator of AI performance in data extraction. During title/abstract screening, accuracy and recall remained high but precision was low, increasing false positives. Full-text screening identified a minority of studies included in the traditional SLR. Mean extraction accuracy was 72.93% (range, 57.69%-88.46%). No data were hallucinated by the AI model used.
Clinical coding is vital yet complex, often hindered by imperfect training data. This study addresses the overlooked issue of undercoding and coding errors in standard datasets and investigates their impact on automated coding algorithms. Furthermore, it explores the potential of AI-assisted tools to facilitate the challenging process of clinical coding audit. We developed a novel and effective interpretable coding pipeline that integrates Large Language Models (LLMs) for evidence extraction and code verification with a multiclass classifier trained on a large-scale silver-standard evidence dataset for code prediction. Using this pipeline as an audit tool, three professional coders systematically identified, categorised, and corrected errors in two widely used datasets containing human-annotated evidence (MDACE and CodiEsp). The audit uncovered significant data quality issues, including an estimated 76.3% undercoding rate in MDACE and a 29.7% error rate in CodiEsp. Re-evaluating existing models on the corrected datasets improved performance, consistently across all metrics on MDACE, and in recall on CodiEsp. Additionally, the proposed pipeline achieved superior or comparable performance to state-of-the-art LLM-based methods, demonstrating its viability as a high-potential dual-use coding framework. Models and results that are publishable in compliance with the data privacy requirements of the datasets have been made publicly accessible in the acocmpanied Github repository: https://github.com/Supriya090/LLM-Clinical-Coding-Audit. Our findings demonstrate that substantial errors in benchmark datasets significantly impact model evaluation. The results underscore a critical need to shift the research focus from purely model-centric approaches to data-centric solutions in clinical AI, highlighting the effectiveness of AI-assisted audits in improving data integrity.
The tomato plant is considered one of the most important crops in the world, yet it is vulnerable to various diseases that affect crop quality and agricultural productivity. These challenges have driven the need for an efficient and intelligent plant disease detection system. With the development of computer vision and artificial intelligence, this proposed methodology based on deep learning for tomato leaf diseases has been presented. Two public datasets: Taiwan DS with nine classes and Tomato Leaf Diseases Detection Computer Vision Dataset (TLDDCV DS) with seven classes have been used to test this system. This system begins with plant image processing, which includes gamma correction and bilateral filtering, to enhance image quality and clarity while preserving key disease features. Then, a genetic metaheuristic algorithm was used to automatically select the most significant hyperparameters, further optimizing both processing time and accuracy. After that, the tomato leaf disease detection applies the You Only Look Once version 11 Nano (YOLOv11n) model. The YOLOv11n backbone is edited through a Data-efficient Image Transformer (DeiT) to improve the system's capacity for learning global contextual information and long-range dependencies. Experimental results demonstrate that the proposed system outperforms existing methods. It achieved an average mAP@50 of 97.8%, mAP@50-95 of 93.4%, precision of 97.3%, recall of 93.8%, and F1-score of 95.5% on the Taiwan dataset. Additionally, it achieved an average mAP@50 of 87%, mAP@50-95 of 48%, precision of 83.9%, recall of 70.3%, and F1-score of 76.4% on the TLDDCV dataset. These results demonstrate the generalizability and effectiveness of the proposed system in real-world agricultural situations.
Chemical analog design in the hit-to-lead and lead optimization stages of drug discovery relies on systematic structural modifications, often guided by medicinal chemistry intuition. Although a matched molecular pair (MMP) provides an interpretable framework to capture intuition, models trained on individual MMP instances face significant limitations, such as bias toward frequent transformations in historical data. This work introduces a novel approach to transform the concept of "how to design chemical analogs like a medicinal chemist" into a primary training objective for generative models by focusing on matched molecular pair transformations (MMPTs) as the fundamental unit of chemical modifications. This allows for a more generalizable and context-independent representation of medicinal chemistry intuition, enabling the application of the same transformation priors across different projects, regardless of the target or indication. Using a consistently curated ChEMBL-derived data set, we compared a transformation-centric foundation model (MMPT-FM) with multiple MMP-based generative formulations trained on the same underlying data. Furthermore, performance is assessed through challenging within-patent and cross-patent real-world test cases derived from drug discovery patents. The MMPT-FM model achieves comparable or improved recall metrics across all test cases, demonstrating particularly strong performance for low-frequency and previously unseen transformations. This work not only shifts the paradigm of learning and utilizes medicinal chemistry intuition in an efficient and scalable manner in the AI for drug discovery era but also establishes a significant competitive advantage by enabling the drug discovery industry to encode decades of collective medicinal chemistry knowledge─both public and proprietary─into a scalable foundation generative model that could help researchers design chemical analogs.
Abdominal ultrasound is widely used for the routine screening of hepatobiliary and renal diseases because it is safe, inexpensive and broadly accessible. In daily clinical practice, however, lesion detection remains highly operator-dependent, particularly when sonographers must identify subtle abnormalities across different abdominal organs under low-contrast, noisy and continuously changing views. Missed or delayed recognition may affect subsequent diagnostic assessment and patient management. Although deep learning has shown promise in ultrasound analysis, most existing methods focus on single-organ or static-image settings, which limits their applicability to real-time abdominal screening. These challenges motivate the development of a unified real-time detection framework that can better support lesion identification in multi-organ ultrasound videos. We propose the Organ-Guided Lightweight Multi-frame Integration (OGLMFI) framework based on YOLOv11 for unified lesion detection across the liver, gallbladder and kidney in ultrasound videos. The framework incorporates an Organ-Guided Feature Filtering module that uses organ segmentation priors to suppress background interference and enhance lesion discrimination. It also includes a Lightweight Multi-frame Integration module, which adopts a dual-branch fusion strategy to efficiently integrate temporal information from consecutive frames using only historical context, thereby preserving causal real-time inference. On the test set of 205 clinical videos (12,964 annotated frames), OGLMFI achieved a Recall of 0.691, a mean average precision at an intersection-over-union threshold of 0.5 (mAP50) of 0.703, an mAP50-95 of 0.510 and an inference speed of 52.3 frames per second. Compared with the YOLOv11-L baseline, OGLMFI improved Recall by 10.7%, mAP50 by 3.2% and mAP50-95 by 8.1%, while maintaining real-time performance. Among the evaluated methods, OGLMFI achieved the highest Recall, mAP50 and mAP50-95. Category-wise analysis further showed the highest average precision at an intersection-over-union threshold of 0.5 across all seven lesion categories, including 66.8% for gallbladder stone and 53.7% for gallbladder polyp. By integrating organ-guided feature filtering and causal multi-frame feature fusion, OGLMFI improves real-time lesion detection in abdominal ultrasound videos. The proposed framework provides a practical unified solution for lesion detection across three abdominal organs and may serve as a useful computer-aided tool for routine abdominal ultrasound screening.
Improving athletic performance and minimizing the risk of injury due to overtraining and/or muscle damage in young tennis players is a key aspect of biomedical monitoring of their training. The purpose of the study was to evaluate biochemical markers of adaptation to physical activity and their correlations with body composition indicators in young tennis players, taking into account their dietary habits. The study involved 59 tennis players aged 7 to 17 years. Using the AU 680 biochemical analyzer (Beckman Coulter), the level of the following biochemical markers in blood serum was determined: total protein, urea, creatinine, uric acid, glucose, cholesterol and triglycerides, the activity of creatine phosphokinase (CPK), alanine (ALT) and aspartate aminotransferase (AST); the muscle damage index (CPK/AST) and the de Ritis coefficient (AST/ALT) were calculated. Somatometric parameters were measured by anthropometry, the body composition was determined by bioimpedancemetry using the ABC-01 Medass analyzer. The analysis of the diet structure was carried out using the 24-hour dietary recall for 3 days of the survey (2 training days and 1 day off). Creatinine levels exceeding the reference norms were revealed in 20 girls and 9 boys, in most cases observed at a younger age, which indicates a greater share of the anaerobic energy supply mechanism compared to the older age group. Serum creatine phosphokinase activity was elevated in 14 of 33 female athletes, while in boys it was within the normal range, which indicates the depletion of adaptation reserves, and therefore requires adjustment of the training schedule with a sufficient recovery period. The strongest relationship was observed for creatinine and uric acid serum level and all studied parameters of body composition (p<0.05), except for the proportion of body fat, which may be due to eating habits and the structure of the diet - optimal protein intake. The relationship was also found between blood level of triglycerides and fat mass (ρ=0.336; p<0.01) and the proportion of body fat (ρ=0.348; p<0.01), which is associated with an excess of fat in the diet. An assessment of biochemical markers of adaptation to physical activity allowed us to determine the current state of physical performance in young tennis players. For athletes with biochemical marker deviations from reference values, additional screening tests are recommended, as well as consultation with the coaching staff regarding adjustments to the volume and intensity of physical activity, as well as the recovery period between training sessions.