Progress in artificial intelligence (AI)-based analysis of surgical videos has been constrained by reliance on manual frame-level annotations rather than patient-level outcomes. In addition, concerns about data privacy restrict the exchange of laparoscopic video data and, thereby, multicenter collaboration. To address these limitations, we developed a pipeline that integrates weakly supervised deep learning with Swarm Learning, a decentralized machine learning approach that enables collaborative model training without data centralization. We evaluated our pipeline using a dataset of 397 laparoscopic appendectomy recordings from six international centers for two patient-level staging tasks: (1) laparoscopic grading of appendicitis and appendiceal perforation detection; and (2) histopathologic inflammation grading. We identified optimal modeling configurations (frame sampling rates and model architectures) using the binary perforation detection task, then compared Swarm Learning with single-center and centralized learning across the laparoscopic and histopathologic disease staging tasks. In addition, we surveyed participating centers to identify barriers to clinical implementation of our learning pipeline for surgical video analysis. For binary perforation detection, frame sampling at one frame per second and use of the SurgTempoNet architecture resulted in reliable classification performance, outperforming SurgFrameNet and Multiple Instance Learning. For both laparoscopic (area under the receiver operating curve [AUROC]: 0.818±0.092) and histopathologic disease staging (AUROC: 0.626±0.029), Swarm Learning consistently outperformed single-center training and achieved performance comparable to centralized learning on external validation (AUROC: 0.795±0.092 for laparoscopic grading; AUROC: 0.610±0.018 for histopathologic grading). The user survey identified hardware failure and limited integration of the decentralized learning pipeline with electronic patient records as key barriers to clinical implementation. Weakly supervised deep learning enables the prediction of patient-level labels directly from surgical video data. Swarm Learning facilitates privacy-preserving multicenter collaboration and achieves performance on par with centralized learning, highlighting its potential for advancing clinically relevant, collaborative AI development in surgical video analysis. (Funded by the European Union and others.).
Accurate and early detection of breast cancer in screening mammography remains a critical challenge in medical imaging, particularly due to variations in image quality, tissue density, and annotation standards across datasets. This study aims to design and validate a Hierarchical Transfer Learning (HTL) framework that leverages heterogeneous mammography datasets and multi‑stage pre-processing to achieve more accurate and generalizable deep learning-based detection of breast lesions in screening mammography. Our approach integrates five image enhancement techniques, CLAHE, Gamma Correction, Adaptive Histogram Equalization, Global Histogram Equalization, and Unsharp Masking into the learning process, allowing the model to adapt progressively to enhancement specific features. We evaluate the framework using two state-of-the-art object detection architectures, YOLOv8 and Faster R-CNN, across three publicly available datasets (INbreast, CBIS-DDSM, and MIAS) and a locally collected dataset. Our HTL model achieves a peak mean Average Precision (mAP) of 0.981 and accuracy of 0.998, outperforming previous methods in both generalizability and precision. Moreover, the model maintains robust performance on real-world local data, demonstrating its practical utility for deployment in diverse clinical environments. These results confirm that hierarchical learning combined with strategic image enhancement and pre-processing significantly improves detection performance and supports large scale breast cancer screening applications with robust cross-dataset generalization, enabling consistent performance across heterogeneous imaging domains with varying acquisition protocols, image quality, and annotation standards.
Medical imaging interpretation is central to clinical decision-making. However, a significant global radiologist shortage necessitates an expanded role for radiological technologists in providing image interpretation support. Preparing students for this role is challenging because undergraduate radiological technology curricula are notoriously crowded, restricting time for practical chest X-ray (CXR) interpretation training. To address this educational gap, we quantitatively evaluated the impact of a highly feasible, 180-minute e-learning intervention on students' CXR evaluation skills. Sixty-nine third-year radiological technology students completed a self-directed learning session using the diagnostic simulator simu.Doc. Accuracy rates and response times for a 50-case CXR test were measured pre- and post-intervention. To assess the method's versatility, participants were stratified into three proficiency levels based on pre-test scores. The overall mean accuracy rate significantly improved from 40.7 to 60.0% (p < 0.001), while the mean response time significantly decreased from 36.2 to 23.2 min (p < 0.001). Subgroup analysis demonstrated significant improvements across all proficiency levels. Furthermore, ANCOVA revealed no significant differences among the groups in post-test accuracy (p = 0.788) or response time (p = 0.653) when adjusted for pre-test scores. These results indicated that performance converged to a statistically uniform level regardless of baseline knowledge. This study provides quantitative evidence that a short, intensive e-learning program utilizing digital transformation can effectively and sustainably enhance CXR interpretation skills, facilitating the integration of image evaluation training into time-constrained undergraduate curricula.
Low birth weight (LBW) is a major public health concern associated with increased neonatal morbidity and mortality. This study was aimed to develop a predictive learning model using machine learning techniques to identify LBW from live birth data. A retrospective population-based study was conducted using data from the Brazilian Live Birth Information System (SINASC), which includes all live births registered in the state of São Paulo between 2019 and 2023. After data curation and exclusion of records with missing or unknown information, a total of 2,548,570 live births were analyzed. Predictors included maternal sociodemographic characteristics, obstetric history, prenatal care indicators, and newborn characteristics. Logistic regression and random forest models were trained using an 80/20 train-test split, with class imbalance addressed through class weighting. Model performance was evaluated using accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC-ROC). The analysis showed that the balanced logistic regression model performed best in terms of sensitivity (0.688), F1-score (0.567), and AUC-ROC (0.850). A proportional analysis of the variables between the LBW and normal-weight groups revealed a higher prevalence of LBW among mothers with low education levels, without a partner, of Black or Brown ethnicity, aged ≤ 19 years or ≥ 35 years, with multiple pregnancies, with a reduced number of prenatal consultations, and with late initiation of prenatal care. The findings highlight the importance of large-scale data analysis supported by machine learning techniques to inform public policies aimed at preventing LBW and promoting maternal and child health.
Functional materials-based chemical sensors play a crucial role in industrial process control, environmental monitoring, medical diagnostics, and safety assurance. Nearly all conventional chemical sensors rely on static material properties and specific operating parameters despite substantial advances in sensing materials and device fabrication, prompting minimal adaptability under inconsistent and complex environments. On the grounds of these constraints, there has been growing interest in developing intelligent chemical sensing, where adaptive behaviour is employed to improve selectivity, robustness, and prolonged stability. This review envisages intelligent chemical sensors as learning-enabled platforms for adaptive chemical detection while propounding a materials-centric yet system-aware vantage on intelligent chemical sensors. The integration of intelligence through the sensing pipeline, adaptive transduction strategies, encompassing hybrid materials and responsive ceramics, learning paradigms, and incorporated sensing architectures is discussed here. Functional materials are intended to enable selectivity, plasticity, drift mitigation, and dynamic sensitivity, while endorsed by system-level incorporation and learning-assisted signal interpretation. Reliable chemical detection in elaborate environments executed by intelligent material device system coupling is highlighted in representative examples. Key challenges associated with material stability, interpretability, data scarcity, and energy efficiency are critically examined besides emerging research directions such as memory-enabled sensing interfaces, chemically adaptive materials, and autonomous sensing ecosystems. This review intends to bridge materials innovation and intelligent system design, proposing perceptiveness for the evolution of next-generation adaptive chemical sensors.
The accelerating effects of climate change, particularly rising sea levels resulting from melting land ice, require accurate predictions of ice sheet dynamics. Traditional numerical models, such as the Parallel Ice Sheet Model (PISM), are widely used; however, they rely on simplifying assumptions and are computationally intensive, limiting their ability to make long-term predictions. This study addresses these limitations by employing machine learning techniques to emulate ice sheet dynamics, offering a faster, more efficient alternative to traditional numerical models. Using data generated by PISM for the Antarctic ice sheet spanning 2015 to 2100, we develop proof-of-concept machine learning models that predict key ice properties-ice mask, ice thickness, and ice velocity-based on climate inputs. Our models tackle both spatial and temporal challenges by incorporating the ice properties of neighbouring cells and past temporal values, thereby improving predictive accuracy. Additionally, explainable AI (XAI) techniques are applied to enhance trust and explainability. Our best-performing model, a Random Forest with an expanding window, achieves an average RMSE of 3.36 for thickness and 65.6 for velocity and misclassifies on average only 3 grid points per year for the ice mask out of 2257, demonstrating good accuracy comparable to that of the original simulator at a fraction of the computational cost. This approach not only improves the computational efficiency of scenario modelling but also contributes to the broader field of climate science by enabling more frequent and extensive analyses of climate change impacts on Antarctic ice sheets.
Neonatal sepsis (NS) is one of the leading causes of neonatal mortality. The nonspecific clinical manifestations and the limited timeliness of existing biomarkers (such as C-reactive protein) highlight the urgent need for highly accurate diagnostic tools. Neutrophils, as key effector cells of innate immunity, are closely involved in the progression of NS. This study integrated training (GSE69686) and validation (GSE25504) datasets from the GEO database. Neutrophil infiltration characteristics were analyzed utilizing CIBERSORT, and weighted gene co-expression network analysis (WGCNA) was introduced to determine neutrophil-related co-expression modules. Three machine learning algorithms-LASSO, SVM-RFE, and RF-were implemented to cross-screen core diagnostic genes. A combined diagnostic model was distributed based on these genes. NetworkAnalyst was utilized to predict miRNA-TF regulatory networks, and GSVA was conducted to interpret biological functions. Three algorithms identified IL1R2 and METTL7B as core diagnostic genes; the model showed strong reliability. IL1R2 high expression correlated with reduced CD8+ T cells, regulatory T cells, and neutrophils (p<0.05). METTL7B high expression linked positively to B cells and negatively to NK cells/neutrophils. The two genes synergistically cause immune cell dysfunction. Six miRNAs and 15 transcription factors (e.g., NFKB1/RELA, STAT3) regulating these genes were found, involved in inflammation and metabolic reprogramming. Integrating neutrophil infiltration and triple-machine-learning, this study first proposed an IL1R2/METTL7B two-gene panel. The model had high accuracy and generalizability, potentially contributing to NS pathogenesis via immune dysfunction and metabolic reprogramming, supporting rapid diagnostics and targeted interventions.
Particularly in highly tourist-active coastal locations, drowning is still a serious international public health concern. This study investigates the predictive value of machine learning approaches in estimating drowning-related mortality risk. This retrospective cohort study analyzed drowning incident data from the Emergency Management and Medical Urgency Center of Guilan Province, covering the period from 2018 to 2023. The data were preprocessed, missing values imputed using the K-Nearest Neighbors (KNN) algorithm, and balanced using the Synthetic Minority Over-sampling Technique (SMOTE). Three models including logistic regression, decision tree, and naïve Bayes were evaluated in predicting the risk of mortality following drowning and sensitivity, specificity, and accuracy of each model was calculated and compared. A total of 600 consecutive cases meeting the eligibility criteria were extracted for analysis, forming the final dataset. Logistic regression exhibited the highest predictive power, with an accuracy of 51.67% and an area under the curve (AUC) of 60.02%. The most influential variables in drowning-related mortality prediction were drowning location, drowning year, gender, and age. High-risk areas posed a 33-fold higher mortality risk than safe locations (p < 0.001). Age and gender were not statistically significant predictors of fatal drowning. Given its superior interpretability and predictive capability, logistic regression was identified as the most effective model for assessing drowning mortality risk. Preventative measures should focus on identifying high-risk areas, installing warning signs, implementing lifeguard teams, educating tourists, and enforcing strict coastal safety regulations to mitigate drowning fatalities.
To compare the performance of different Survival Machine Learning (SML) algorithms in predicting the survival of cancer patients. Data from the Registro Hospitalar de Câncer do Estado de São Paulo (São Paulo State Cancer Registry Hospital) were used, covering the five most incident types of cancer (breast, prostate, lung, colorectal and cervix). Six algorithms were evaluated: Gradient Boosting Survival (GBS), Random Survival Forest (RSF), Support Vector Machine Survival (SVM-Survival), XGBoost Cox, XGBoost Accelerated Failure Time (AFT), and LightGBM. Performance was measured by the Concordance Index (C-Index), C-Index IPCW and Integrated Brier Score (IBS) metrics. The XGBoost AFT model showed the best C-Index results for breast (0.7845), lung (0.7368), colorectal (0.7618), and cervix (0.7726), while GBS was superior for prostate (0.7574). Clinical staging was consistently the most important variable, according to the explainability analysis. The SML algorithms showed good predictive performance, regardless of cancer type, sample size and censoring proportion. These models show potential for subsidizing cancer planning and supporting strategic decisions in the organization of cancer care networks.
Introduction. Bloodstream infections (BSIs) are particularly problematic in the emergency department (ED) of hospitals, where patients often present with undiagnosed illness, one cause of which may be undetected BSI. Identifying whether a patient needs a blood culture (BC) performed is one component of this challenge with implications for diagnostic efficiency and avoidance of unnecessary resource expenditure.Gap statement. In Western Australia, there has been no previous study investigating methods to predict BC outcome in an ED patient cohort.Aim. This article assesses the feasibility of previously developed machine learning (ML) models for BC outcome prediction using a prospectively collected ED patient dataset.Methodology. An ML pipeline containing models previously trained using complete blood count (CBC), white blood cell differential (DIFF) and cell population data (CPD) generated by Sysmex XN-2000 haematology analysers was further evaluated using prospectively collected data containing patient sample results from the ED at Sir Charles Gairdner Hospital (SCGH), Perth, Western Australia. Blood samples used to produce CBC, DIFF and CPD were obtained at the same time as BC samples.Results. There were 59 samples from 58 unique patients. Forty-nine of those samples were associated with negative BC results and 10 with positive BC results. We evaluated previously developed XGBoost (XG) and random forest (RF) ML models for positive BC outcome prediction. The RF and XG models obtained mean area under the receiver operating characteristic curve scores of 0.865 (95% CI, 0.763-0.947) and 0.833 (95% CI, 0.683-0.953) with the ED dataset.Conclusion. The results presented in this study provide a foundation for further validation and shadow deployment of BC outcome prediction models in clinical settings and support future planning of clinical trials.
As the numbers of older people (65 +) rise globally, the pressure on acute hospitals to provide efficient and effective care while addressing resource inequities increases. In this study we introduce Recognising Episodes of Acute Complexity in Health (REACH), a novel Automatic Machine Learning (AutoML)-based predictive model that prioritises older patients and assigns them to complex or non-complex care pathways using routinely collected electronic health record (EHR) and interRAI data. Existing triage and patient-flow approaches are rarely designed for equity-sensitive deployment in settings where underserved populations experience systematically different access and outcomes; in our context, this gap motivated the development of REACH. Designed to be flexible and adaptable, REACH was developed and stress-tested in New Zealand, where Māori, Pasifika and other underserved older populations face inequitable access and outcomes, and is intended as a transferable approach for other health systems with similar equity and capacity challenges. The model estimates individual complexity probabilities and incorporates resource and capacity constraints to support decisions about patient prioritisation, bed allocation and placement into older-person-specific pathways. Applied to seven years of hospital data, REACH demonstrates improved identification of complex geriatric cases and more efficient use of specialised ward capacity compared with traditional FIFO and triage-based assignment. Using New Zealand data as a real-world case study, we demonstrate that equity-aware, capacity-constrained pathway assignment can be operationalised with routine EHR/interRAI inputs, with a design that can be retrained and recalibrated for other jurisdictions that use comparable assessment standards. The application of the model demonstrates its potential to improve patient prioritisation and placement processes while addressing challenges in healthcare access for underserved populations, highlighting the importance of data-driven decision support systems in older persons' care. The findings hold significant implications for healthcare practitioners, managers, policymakers and researchers seeking to design equitable and efficient models of acute care for older people.
Purpose To characterize Predetermined Change Control Plan (PCCP) adoption and documentation transparency among U.S. Food and Drug Administration (FDA)-cleared radiology artificial intelligence/machine learning (AI/ML)-devices (2015-2025). Materials and Methods A cross-sectional systematic scoping review was conducted with linked data from FDA AI/ML-enabled device databases through April 2026. PCCP documentation completeness was scored by two independent observers (intraclass correlation coefficient: 0.93) using an 8-point rubric (possible scores, 0-8) derived from FDA's final PCCP guidance (December 2024). Identified PCCP devices underwent manual verification against FDA regulatory summaries. Results Among FDA-listed AI/ML-device submissions, 1080/1394 (77.5%) were radiology submissions, and nearly all cleared via 510(k) review. Across all FDA panels, 170 devices were cleared with a PCCP. Radiology led AI/ML-specific PCCP adoption (34/37 radiology PCCP devices, 91.9%). Among the PCCP-cleared radiology AI devices, 22/34 (65%) were cleared in 2025 alone following the final FDA guidance. Discrepancies between FDA's public database and individual summaries required manual adjudication for 9/34 (27%) of devices. PCCP documentation scores ranged from 0 to 8 (mean, 5), with most modifications focused on data retraining, compatibility expansion, and algorithm optimization. Continuous monitoring of device performance and predefined drift triggers for retraining were absent from public summaries. Conclusion PCCP adoption increased in radiology after issuance of the final FDA guidance, yet public lifecycle controls, particularly monitoring performance metrics and trigger thresholds, were limited. Standardized PCCP reporting of lifecycle controls are suggested as a condition of PCCP authorization to enable systematic postmarket monitoring as this pathway scales. ©RSNA, 2026.
CRISPR-Cas9 has become a widely used tool for genome editing. However, its off-target cleavage caused by partial sequence matches with guide RNAs (gRNAs) remains a critical limitation. Recently, abasic gRNAs (ØXØ) have been developed to enhance target specificity, but their effects vary depending on the positional sequence context. Here, we present abCRISPR, a deep neural network (DNN) framework for the rational design of ØXØ sequences with minimized off-target activity. abCRISPR leverages informative few-shot training with paired datasets of abasic and unmodified gRNAs, using high-quality random mismatch target libraries, exhaustively sequenced for mismatched off-target substrates (n = 97,583) in in vitro CRISPR-Cas9 cleavage experiments. Predicted off-target activities for both abasic and unmodified gRNAs showed strong correlation with experimental data (r ≥ 0.95, 10-fold cross-validation). Notably, these comprehensive training sets provide robust ground-truth negatives, enabling accurate and sensitive prediction of off-targets. For unmodified gRNAs, abCRISPR (AUC = 0.98) was validated to outperform existing deep learning-based methods (AUC = 0.45-0.68). When applied to the human genome, abCRISPR generated ØXØ sequences, covering 58,875,004 potent CRISPR-targetable sites with improved target specificity. Together, this work provides a comprehensive bioinformatics resource for safe and precise CRISPR-Cas9 genome editing. The source code for abCRISPR and training data are available at https://doi.org/10.5281/zenodo.20398246. abCRISPR results for the human genome are available at http://clip.korea.ac.kr/abCRISPR/. Supplementary data are available at Bioinformatics online.
Systematic literature reviews offer high potential for efficiency gains from artificial intelligence (AI), now integrated into several systematic literature review software platforms. Validation studies show acceptable sensitivity, specificity, and accuracy for AI-assisted systematic literature reviews of clinical trial publications. Unlike trials, economic model publications lack consistency in content, terminology, and structure. We aimed to test the efficiency and accuracy of AI-assisted search, screening, and data extraction when applied to a systematic literature review of economic evaluations. A previously conducted manual systematic literature review of economic evaluations for chronic rhinosinusitis with nasal polyps was replicated using a machine learning-based inclusion prediction model (Robot Screener) and a large language model-based criteria screener (Smart Screener) within Nested Knowledge software, with performance benchmarked against the original human-conducted systematic literature review. The AI-generated search retrieved 22/43 (51%) PubMed articles from the original systematic literature review. Accuracy exceeded 95% for title/abstract screening but fell below 80% for full-text screening. Extraction was reliable for high-level model descriptors and general study characteristics, but less so for model structures, health states, outcomes, and distinguishing sensitivity from scenario analyses and complex modeling assumptions for duration of response, discontinuation, surgery, and mortality. Estimated time savings ranged from ~20% (data extraction) to 60% (title/abstract screening and searches), varying by task human validation requirement. Artificial intelligence-driven tools performed well for title/abstract screening and general data extraction but were less accurate for full-text screening and interpretation of modeling choices. They can increase systematic literature review efficiency for economic evaluations but fall below the reliability seen for systematic literature reviews of clinical trials. Artificial intelligence has the potential to speed up systematic literature reviews by assisting researchers with identifying articles and organizing information from the articles. In this study, artificial intelligence-assisted tools embedded within a commercial systematic literature review platform were applied to replicate a systematic review of economic evaluations, providing practical insights into their readiness for use in health economics and outcomes research and health technology assessment workflows. Artificial intelligence performed well in early-stage screening, achieving more than 95% accuracy when reviewing titles and abstracts but its accuracy dropped below 80% when screening full texts. Artificial intelligence also reliably extracted general data such as study perspective, comparators, geography, and time horizon, but it struggled with more complex elements and those that required some interpretation and experience with these types of studies, including model structure, health states, sensitivity analyses, and key modeling assumptions. Estimated time savings ranged from approximately 20% for data extraction to 60% for title and abstract screening and database searches, though savings were highly dependent on the approach taken and the extent of human validation required. Overall, the study shows that artificial intelligence can improve efficiency in aspects of the process of conducting a systematic literature review of economic evaluations by streamlining general tasks but still requires substantial human oversight.
Retrieving an implanted leadless pacemaker is a challenging component of its management. However, there are no dedicated tools specifically designed for retrieving the Micra leadless pacemaker. Given the lack of commercial retrieval systems, there is an urgent need to document and validate alternative interventional strategies to manage complications associated with device malfunction or migration. In this series, we present cases involving increased capture threshold, device embolization into the right pulmonary artery, and broken tines following Micra implantation. We successfully completed four retrieval procedures using a coordinated approach with existing devices, including the Micra transcatheter pacing system, coronary angiography catheter, and endovascular snare system. Notably, these cases exhibited different characteristics during the retrieval of the Micra leadless pacemaker, which may assist colleagues in dealing with such scenarios. Our experience demonstrates that a strategy combining the Micra delivery catheter with an endovascular snare system, without the need for a steerable sheath, was sufficient and effective for successful Micra retrieval.
暂无摘要（点击查看详情）
Mutation-induced drug resistance challenges both pandemic surveillance and drug discovery. While experimental assays are resource-intensive, current computational predictions remain limited by the scarcity of 3D mutant protein structures. We present DeepMutDTA, a structure-independent model pre-trained on 1.5 million data points to predict drug-target affinity and uncover underlying interaction mechanisms. However, like other sequence-based approaches, it often falls short in predicting mutant affinities due to the overwhelming sequence similarity between wild-type (WT) and mutant (MT) targets. To bridge this gap, we introduce SimSiam-MuTF, a novel fine-tuning framework to enhance the detection of resistance variants by explicitly aligning latent embedding distances with the corresponding shifts in binding affinity between WT and MT targets. Compared to representative baselines, our model exhibits remarkable robustness across varied sequence identities and unseen data splits, yielding average performance gains of 2.47% (PCC) and 5.10% (SCC) in regression tasks, alongside 4.00% (AUC) and 4.17% (AUPR) in classification tasks. Applications to SARS-CoV-2, HIV-1, and cancer-related targets highlight its generalization potential and utility in informing therapeutic strategies against drug resistance. Collectively, this robust computational pipeline and fine-tuning framework deepen our understanding of mutation-induced resistance and may serve as a powerful platform to accelerate drug discovery against mutant targets.
Language is a cognitive domain essential for communication; however, its responsiveness to acute physical exercise or activity is unexplored. This quasi-randomized trial examined the effects of different intervention modalities on verbal functions and attention. Ninety-eight participants were assigned to high-intensity interval training (HIIT), coordinatively demanding motor activity (COOR), or a sedentary control group (CG). A pre-post design with a 24-hour follow-up assessed semantic (VFsem) and category-switching verbal fluency (VFsw), verbal learning, and attention. No group × time effects were observed for verbal learning encoding (p = .258); however, HIIT showed a steeper learning slope compared to CG (p = .044). Following an interference list, proactive interference was unaffected (p = .428), whereas retroactive interference differed significantly between groups (p = .019), with CG showing lower susceptibility than both intervention groups. No group differences were found for consolidation after 20 min or 24 h (p's ≥ .271). Semantic verbal fluency improved significantly following HIIT relative to CG (p = .029), with no significant effects in VFsw (p = .221) or attention (p = .187). We conclude an intensity-dependent pattern following the interventions, manifesting as enhanced semantic lexical retrieval and an exploratory finding of a steeper encoding following HIIT, compared to CG, with COOR in between, accompanied by an increased susceptibility to retroactive interference.
Content-based retinal image analysis plays a crucial role in the early diagnosis of ocular diseases. In this study, we proposed a novel approach for efficient content-based retinal image retrieval and Diabetic Retinopathy (DR) detection using variants of local texture features derived from Local Binary Pattern (LBP), Local Ternary Pattern (LTP), and Gray Level Co-occurrence Matrix (GLCM). The methodology begins with meticulous image preprocessing to enhance feature extraction, followed by the extraction of LBP, LTP, and GLCM features, which capture intricate texture patterns and enrich the feature space for robust analysis. Subsequently, we trained machine learning models, including Support Vector Machine (SVM), Decision Tree, and Random Forest, on the extracted features to effectively retrieve retinal images and detect DR. A comparative analysis between preprocessed and raw images highlights the impact of preprocessing techniques on performance. A key innovation of this study lies in the fusion of multiple texture-based features, creating a comprehensive representation that integrates high-level semantic information with fine-grained local patterns. This hybrid approach enhances the system's capability to handle diverse retinal image variations, leading to improved retrieval accuracy and robustness. Further, a metaheuristic approach for feature selection and optimization is employed, comparing Differential Evolution, Genetic Algorithm, and Particle Swarm Optimization to identify the most effective features for retrieval. Differential Evolution achieved the highest precision of 90.67 % for retrieving the top 10 relevant images. The proposed hybrid approach demonstrates the effectiveness of integrating classical image analysis methods with machine learning for DR detection and content-based image retrieval. This research contributes to precision medicine and healthcare innovation by advancing ML-driven retinal image analysis.