Background Evidence on explainable machine learning (ML) for colorectal cancer (CRC) classification using multicenter routine hospital data remains limited, particularly in Southeast Asia. Objective To develop and evaluate explainable ML models for CRC classification using multicenter routine hospital data in Thailand. Methods A retrospective multicenter case–control study was conducted across 11 provincial hospitals between June 2016 and June 2017, including 504 newly diagnosed CRC cases and 997 controls. The dataset was randomly divided into training (70%) and testing (30%) sets. Model robustness was assessed using five-fold cross-validation and repeated analyses across five random seeds. Adaptive Synthetic Sampling (ADASYN) was applied to address class imbalance. Logistic Regression, K-Nearest Neighbor, Random Forest (RF), Extreme Gradient Boosting (XGBoost), and AdaBoost models were evaluated. Model interpretability was assessed using Shapley Additive Explanations (SHAP). Results Tree-based ensemble models demonstrated superior classification performance compared with traditional ML approaches. In the ADASYN with LASSO feature selection, XGBoost achieved the highest overall performance (accuracy 0.94 ± 0.006; AUC 0.98 ± 0.002; sensitivity 0.91 ± 0.005; specificity 0.95 ± 0.007), while RF showed comparable discrimination performance (accuracy 0.93 ± 0.003; AUC 0.98 ± 0.001). SHAP analysis identified hypertension, blood pressure measures, fasting plasma glucose and behavioral indicators as influential features associated with case–control classification Conclusions Tree-based ensemble models demonstrated strong and robust performance for CRC classification using routine hospital data. SHAP analysis identified key socioeconomic, metabolic, clinical, and behavioral features associated with classification. External and prospective validation studies are needed to assess generalizability and support future implementation.
Retrieval-augmented generation (RAG) is widely assumed to improve medical question answering, yet the conditions under which retrieval helps, leaves performance unchanged, or actively harms it remain poorly characterized. This paper is a diagnostic study of that question rather than the proposal of a new retrieval framework. We evaluate a deliberately text-only (image-free) pipeline in which the model receives the question and retrieved text but never the medical image, using SLAKE and VQA-RAD as sources of clinically diverse question types, and we validate externally on PubMedQA (n=1000). We compare six retrieval conditions across two language models. The conditions are parametric-only No-RAG, dense RAG, hybrid BM25+dense, two cross-encoder rerankers (general and biomedical), and CAF-RAG, a training-free clinical filter whose auditable keep/remove decisions make it a diagnostic instrument. The two models are GPT-4o-mini and the open medical model BioMistral-7B, which read byte-identical retrieved context. Our central finding is negative and consistent: text-based retrieval does not significantly outperform parametric knowledge on any group we tested, and on external data it is significantly worse (PubMedQA, −0.120, p<0.0001). No passage-selection method recovers it: neither reranker significantly beats plain RAG, and CAF-RAG is statistically indistinguishable from a cross-encoder reranker (＋0.002, p=0.91) while recovering part of the accuracy that retrieval costs (＋0.034 over RAG on SLAKE-CLOSED, p=0.027). The two models agree to within 0.002 on both effects, indicating these are properties of retrieval rather than of one model. We identify a behavioral mechanism: retrieved context displaces the model from its parametric answer prior. It moves the model toward affirmation when the context is topically related to the question (both models swing sharply toward “Yes” on SLAKE), and toward hedging when it is not (both swing toward “Maybe” on PubMedQA). We also report an audit of our own evaluation: a substring-matching artifact in lexical scoring inflated retrieval conditions more than the terse parametric baseline, and we replace it with a scorer validated against 197 blinded human gradings (91.9% agreement, κ=0.79). All per-question outputs, the rebuilt knowledge base, retrieval contexts, and the scorer with its documented failure modes are released.
Tuberculosis (TB) triage from chest radiographs is influenced by image representation and training-data variability in addition to model architecture. This study quantified the independent and interaction effects of these workflow factors using 3522 examinations from 3100 patients in a single-institution multi-vendor PACS/DICOM cohort. A representation-matched 2 × 2 factorial design compared a conventional DICOM-to-PNG representation (P0) with DICOM-aware HyFusion preprocessing (P1), and the original training set (A0) with a deterministic clinically constrained augmented pool (A1). Four DenseNet121 conditions used identical patient-level partitions, optimization settings, three random seeds, and a fixed raw-probability threshold of 0.5. Evaluation included discrimination, operational burden per 1000 examinations, validation-only temperature scaling, 5000 paired patient-level bootstrap resamples, and adjusted generalized estimating equations. C3 (P1A0) achieved the highest mean AUROC (0.8732 ± 0.0085) and AUPRC (0.8809 ± 0.0194). Relative to C1 (P0A0), the AUROC and AUPRC differences were ＋0.0185 (95% CI −0.0070 to ＋0.0451) and ＋0.0217 (95% CI −0.0023 to ＋0.0487), respectively. At the fixed threshold, C3 increased specificity and PPV and reduced referrals by 78.5 and false positives by 53.5 per 1000 examinations. A1 reduced referrals and false positives but also reduced sensitivity and increased false negatives. Temperature scaling improved NLL, Brier score, and ECE10 across all conditions. The study establishes a reproducible workflow-factor attribution framework and identifies P1A0 as the DenseNet121 configuration with the highest mean discrimination and a favorable calibration and referral-efficiency profile in this cohort.
Early detection of Parkinson’s disease remains difficult because symptoms overlap with related disorders, and there is no definitive diagnostic test. We present an interpretable end-to-end screening framework that unites a structured preprocessing protocol with a transparent stacking ensemble. Using a dataset of 2,105 individuals with diverse demographic, lifestyle, and clinical profiles, we performed outlier detection and removal, class rebalancing with ADASYN, and Chi-Square feature selection to ensure high-quality inputs. Two high-performing gradient boosting classifiers, XGBoost and CatBoost, are optimized separately and then combined in a stacking architecture to leverage complementary strengths while reducing variance. Using an 80:20 stratified train–test split and stratified cross-validation, the stacked model achieved 97.15% accuracy, a 0.9806 AUC score, and a 0.9411 Cohen’s Kappa Score and outperformed the individual baseline classifiers under the selected-feature evaluation setting. Importantly, the pipeline emphasizes interpretability without sacrificing performance. SHAP provides global attribution that highlights stable cohort-level prediction drivers, and LIME provides instance-specific explanations that clinicians can examine and interpret. Furthermore, the combined SHAP and LIME analyses highlighted UPDRS, MoCA, and Bradykinesia among the influential clinical features associated with diagnostic outcomes. The Mann–Whitney U test indicated statistically significant differences between the predicted score distributions of the two diagnostic classes (p < 0.05). The novelty of this work lies in three elements: a reproducible pipeline from preprocessing to prediction that couples rigorous outlier control, class rebalancing, and statistically grounded feature selection; a focused stacking design that uses two modern boosting learners to deliver dependable gains; and a clinician-oriented interpretability framework that provides both global and local explanations to support clinical decision-making. By bringing together data curation, stacked gradient boosting, and actionable explanations, this study offers a promising foundation for interpretable Parkinson’s disease screening at the initial clinical assessment. However, prospective clinical validation is required before real-world deployment.
Clinical notes in Electronic Health Records (EHRs) can provide enormous amounts of data for research purposes. Extraction of data from clinical notes can be automated by Pretrained Language Models (PLMs). However, erroneous PLMs can extract incorrect data potentially impacting study results. We evaluated the impact of PLM-induced errors in cohort selection on subsequent clinical research with a focus on prognostic prediction model development. We used an EHR database of over 40,000 patients and deliberately decreased the performance of an PLM such that the model selected increasingly inaccurate cohorts of patients. Eligibility was defined by the presence of a target disease/procedure in the clinical notes. We used these inaccurate cohorts to develop prognostic prediction models and evaluated their discrimination and calibration performance. We found that PLMs with decreasing cohort selection performance (expressed by the F1-score), selected increasingly inaccurate cohorts. This resulted in aberrant regression coefficients in the prediction models developed on those cohorts. However, it did not affect the performance of the prediction models in terms of discrimination or calibration. While inaccurate cohort selections by erroneous PLMs affected the prediction model's coefficients, discrimination and calibration remained largely unaffected by PLM error. This is possibly due to the insensitivity of these measures to slight changes in the study cohort. Our results suggest that prognostic models may be robust to minor PLM-induced cohort selection errors. For clinical research focused on general prediction model performance rather than interpretation of individual coefficients, PLMs with moderate errors may still be applicable for selection of cohorts.
Enhancing mammographic images while preserving subtle structures is challenging because mammograms often exhibit low contrast, noise, and small clinically relevant features. This study proposes the Fuzzy Genetic Enhancement Technique (FGET), which combines Wiener filtering, a parametric fuzzy membership function (FMF), and genetic-algorithm optimization for pixel-level mammogram enhancement. FGET was evaluated using 50 mammograms from the Curated Breast Imaging Subset of the Digital Database for Screening Mammography (CBIS-DDSM) and compared with histogram equalization (HE), fractional integral (FI), and fuzzy image enhancement (FIE). Performance was assessed using the average gradient (AG), Fog-Aware Density Evaluator (FADE), Blind Image Quality Model Evaluator (BIQME), and execution time. FGET achieved the highest mean AG value (2.21 ± 1.61), compared with 1.20 ± 0.92 for the original images, 1.64 ± 1.17 for HE, 1.74 ± 0.94 for FI, and 1.98 ± 1.41 for FIE. Welch's t-tests indicated significant AG differences between FGET and the original images and between FGET and HE; the differences between FGET and FI and between FGET and FIE were not statistically significant. Consistent with the direction of the FADE metric, FI produced the largest reduction (8%), whereas FGET produced a moderate reduction (5%). BIQME results were mixed and did not establish the general superiority of FGET. FGET required 1.50 min, compared with 0.78 min for HE, 1.34 min for FI, and 1.65 min for FIE. Overall, FGET improved AG-based edge and detail enhancement at a moderate computational cost. Its effect on microcalcification detection requires future validation by radiologists or task-specific detection experiments. The single-dataset design also precludes conclusions about cross-dataset generalizability.
Artificial Intelligence-enabled software systems are increasingly integrated across various sectors, driven by the promise of improved decision-making and operational efficiency. However, ensuring such system-driven decisions are fair and unbiased poses significant challenges in some high-stakes domains such as healthcare. The current work considers this issue through the specific example of the automated classification of retinal images acquired by Optical Coherence Tomography. A system development pipeline is used to evaluate fairness-focussed practice in that system. Using the OCT2017 dataset, images from four classes of recognised types were first downsampled to ensure the dataset contains equal numbers per class (a balanced dataset) and resized to 128x128 (with one cohort of images also resized to 224x224 to evaluate the effect of image size on model performance). Three data pipelines are defined incorporating different training/validation/testing splits (∼97:<1:<3, 80:10:10 and 70:10:20) along with the absence or presence of certain preprocessing steps (noise reduction and contrast enhancement). In all cases a customised DenseNet121 deep learning model is employed via transfer learning to classify retinal conditions. Other factors including alternative class balancing strategies and alternative deep learning architectures are not explored in this study. Results indicate that dataset balancing by downsampling when used with the configured DenseNet121 architecture improves fairness at the expense of performance metrics for their comparable imbalanced studies. Within the constraints of this work and subject to further validation an exploratory study suggests that preprocessing steps such as contrast enhancement and noise reduction may offer minor benefits, but the data splitting strategy for training and testing has a greater influence on model accuracy and reliability. This study underscores the importance of embedding fairness practices into the artificial intelligence development lifecycle to promote more equitable and trustworthy systems.
Accurate NSCLC subtype diagnosis by CT imaging is critical for effective personalized treatment; however, challenges exist related to class imbalance, model opacity, and computational overhead in clinical practice. We propose a two-step deep learning method with progressively updated versions of VGG16, VGG19, MobileNetV2, ResNet50, EfficientNetB5, and ensembles of deep networks to predict binary cancer and classify it into NSCLC subtypes (adenocarcinoma, squamous cell, and large cell carcinoma). The preprocessing of data incorporates CLAHE for contrast enhancement and a standardized 224 × 224 image size. Dynamic class weight adjustment, one-class classification, and various data augmentation procedures help alleviate the severe class imbalance problem in a dataset of 1000 CT images (33.8:18.7 distribution). VGG16 achieves an accuracy of 98.55% with a binary detection precision–recall of 0.99, while MobileNetV2 achieves 97.97% accuracy with perfect cancer recall (1.0) using limited computational resources. ResNet50 achieves an accuracy of 95% (F1- 0.96) for classifying the cancer classes. Crucially, the combined use of Grad-CAM and LIME explainability supports the reliability of the predictions. External validation using a separate dataset (IQ-OTH/NCCD) demonstrates 100% binary accuracy (VGG16, AUC 1.00) and 99.09% multiclass accuracy (ResNet50). This real-time prototype surpasses the current literature in terms of accuracy, interpretability, and practicality, improving the decision-making process for radiologists while reducing variability and enabling precise NSCLC subtyping to support molecular therapy.
Molecular subtyping of colorectal cancer (CRC) provides valuable insights into tumor heterogeneity and has the potential to guide treatment strategies. However, its clinical integration remains limited, largely due to the cost and complexity of RNA sequencing required for subtype assignment. In this study, we investigated the feasibility of using deep learning to classify CRC consensus molecular subtypes (CMSs) directly from histopathology images. We utilized diagnostic whole-slide images and corresponding RNA-seq data from the TCGA COAD and READ cohorts, focusing on tumor regions. These tumor regions were tessellated into tiles and labeled according to transcriptome-derived CMS classifications. We evaluated two architectures: a ResNet-34 convolutional neural network and a Vision Transformer (ViT-B-16), both trained on 128,181 tiles across both cohorts. Both models achieved comparable classification performance; however, ResNet-34 required significantly fewer computational resources. The ResNet-34 model was evaluated using 10-fold cross-validation and an independent test set, achieving an accuracy of ∼92% on the test set, with micro and macro average AUC values reaching ∼0.99. Despite the inherent class imbalance in CRC datasets, the model demonstrated strong generalization across all subtypes, highlighting the potential of image-based classification for CRC molecular subtyping. By enabling subtype identification from widely available histopathology slides, this method may support broader integration of molecular subtyping into precision oncology workflows.
Obstetric and gynaecological records are among the most privacy-sensitive categories of personal health data. Conditions such as cervical cancer, preeclampsia, gestational diabetes mellitus, and preterm birth contribute substantially to maternal and perinatal morbidity and mortality, particularly in low-resource settings. Federated learning (FL) enables collaborative model training while raw patient records remain within institutional boundaries, making it a promising approach for privacy-preserving artificial intelligence (AI) in women’s health. This systematic review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 statement and synthesises FL research in obstetrics and gynaecology (OB/GYN), characterising clinical applications, architectures, privacy-enhancing technologies (PETs), model performance, methodological quality, and clinical translation barriers. Eight electronic databases were searched from inception to March 2026. Eligible studies applied a genuine federated, split, swarm, or privacy-preserving decentralised learning architecture to an obstetric or gynaecological task and reported at least one quantitative performance metric. Risk of bias was assessed using the Prediction model Risk Of Bias ASsessment Tool (PROBAST) and the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2). Quantitative synthesis used descriptive subgroup area under the receiver operating characteristic curve (AUC–ROC) summaries, a visual forest plot, and exploratory AUC–ROC pattern analysis rather than formal meta-analysis. Forty-two records were retained after full-text assessment across five clinical subdomains, of which 31 contributed to the quantitative synthesis. After methodological reclassification, 28 studies satisfied the operational definition of true FL for architecture and PET coding. Horizontal FL with federated averaging (FedAvg) was a commonly identified design pattern. Differential privacy was the most frequently reported PET, while 10 of 28 architecture/PET-coded studies reported no formal privacy mechanism. The overall descriptive AUC–ROC summary was 0.878. PROBAST indicated high risk of bias in most synthesis studies, and evidence certainty was very low to low across all subdomains. The evidence suggests that FL is scientifically promising for privacy-preserving AI in women’s health, but remains largely at the proof-of-concept stage. Future work should prioritise real multi-institutional validation, stronger privacy engineering, calibration and uncertainty reporting, robustness to non-independently and non-identically distributed (non-IID) data, and clinically grounded translation pathways.
Background and objective: Accurate liver segmentation from volumetric computed tomography underpins surgical planning, volumetric assessment and patient-specific reconstruction, yet anatomical variability and low soft-tissue contrast hinder automation, and published comparisons rarely control training conditions. This work delivers a rigorous, reproducible and externally validated benchmark of nine 3D convolutional and transformer architectures, designed so that measured differences reflect architectural design rather than experimental confounds. Methods: Every architecture was re-implemented and trained from scratch under one identical protocol, with shared preprocessing, augmentation, optimisation, hardware and a single learning-rate configuration, on two co-primary public datasets evaluated independently: the Multi-Atlas Labelling Beyond the Cranial Vault (BTCV) and the larger, multi-institutional and pathology-bearing Medical Segmentation Decathlon Task03 (Liver). The trained weights were then validated, in a zero-shot regime, on an independent single-centre clinical cohort, the Álvaro Cunqueiro Hospital Liver Dataset (Vigo, Spain). Segmentation quality was quantified with five volumetric and boundary-aware measurements, and inter-architecture differences were tested using the Friedman test with paired Wilcoxon signed-rank tests. The complete benchmark is released as a configuration-driven PyTorch Lightning framework (LightningMedSeg3D). Results: On BTCV, boundary accuracy was shared between 3D SwinUNETR (best MASD ≈0.824) and Res-UNet (best HD95 ≈2.5 mm), with Attention-UNet best on overlap (DSC ≈0.959). On Task03 the ranking re-ordered (Friedman p<10−7): Res-UNet ranked first (DSC 0.919 ± 0.053), ahead of Attention-UNet (0.887), and 3D V-Net failed (DSC 0.331), so no architecture led on both datasets. Under external validation the training source proved decisive: BTCV-trained weights transferred significantly better than Task03-trained weights for eight of nine architectures (paired Wilcoxon p≤0.013), with BTCV-trained SwinUNETR, Attention-UNet and Res-UNet the most reliable (DSC 0.952, 0.949 and 0.947). Conclusions: No single architecture is universally superior; attention-gated convolutional models and 3D SwinUNETR generalise most reliably, and the training distribution should match the intended clinical population. The reproducible, externally validated benchmark provides a dependable and openly available basis for model selection in clinical liver segmentation.
Background ASA Physical Status (ASA-PS) classification is widely used in preoperative risk assessment but has substantial inter-rater variability. Large language models (LLMs) may improve classification consistency. Methods We assessed five early-generation and six advanced reasoning LLMs on published vignettes using an optimized prompt. Set A (Hurwitz 2017, n = 10) served for prompt development; Set B (Owens/Mak/Riley, n = 10) and Set C (De Cassai 2019, n = 8, ASA I–V) provided independent validation. DeepSeek R1 was also tested on 150 Cheng et al. (2024) vignettes for aggregate distributional concordance. Set A was evaluated in English and Italian. Results On Set A, advanced LLMs reached 97.5% accuracy (117/120; bootstrap CI 95.0%–100.0%) versus 77.0% for early-generation models (77/100; p = 0.0008) and the human baseline. On Set B, the four selected advanced LLMs converged for 9/10 cases. On held-out Set C, DeepSeek R1 achieved 100% accuracy (8/8) with a seven-call self-consistency ensemble. On Cheng et al., DeepSeek R1's aggregate distribution showed no significant difference from experts (p = 0.927; mean 2.30 vs. 2.32), but underrepresented ASA IV cases (2.7% vs. 8.0%). Conclusions Advanced reasoning LLMs performed strongly on standardized ASA-PS benchmarks, and DeepSeek R1 classified held-out ASA I–V cases accurately while showing aggregate-level concordance with experts. Self-consistency agreement scores flagged borderline cases. ASA IV underrepresentation supports pairing any deployed classifier with an independent high-risk escalation rule. Findings are limited to synthetic vignettes and one language pair; prospective validation on real-world electronic health records is required.
Parkinson's disease (PD) is a progressive neurological disorder that affects speech production, making speech-based analysis an effective tool for early diagnosis. Although machine learning (ML) techniques have shown promising performance in PD detection, their performance is often affected by high-dimensional feature spaces and the presence of redundant or irrelevant voice features. However, conventional feature selection methods often produce unstable feature subsets and may not consistently identify the most informative voice features. To address these challenges, this study proposes a Consensus-Based Multi-Metaheuristic Feature Selection framework integrated with lightweight supervised machine learning, termed CMFS-LSML. The proposed framework combines Particle Swarm Optimization (PSO), the Grey Wolf Optimization (GWO), and Whale Optimization Algorithm (WOA) within a consensus-driven voting mechanism, where features selected by at least two of the three optimization algorithms are retained and evaluated using the combined classification performance of Decision Tree (DT), Random Forest (RF), K-Nearest Neighbors (KNN), and Logistic Regression (LR) classifiers as the optimization objective. Experimental results demonstrate effective feature reduction while maintaining high classification performance. It achieves 98.31% accuracy using 9 features (60.87% reduction) on the Parkinson Disease Classification Dataset and 95.83% accuracy using 10 features (77.27% reduction) on the Parkinson Dataset with Replicated Acoustic Features. Compared with recent studies, the proposed method improves the reduction ratio by approximately 13.05% and 18.18%, respectively, while maintaining higher or comparable accuracy. Furthermore, compared with using the full feature set, the proposed approach reduces the average execution time by 77.25%, resulting in lightweight ML models suitable for real-time PD diagnosis and resource-constrained healthcare environments.
The widespread use of histopathology images, particularly those stained with hematoxylin and eosin (H&E), has proven pivotal for detecting tumors. However, advancing cancer cases to appropriate treatment pathways requires accurate quantification of human epidermal growth factor receptor 2 (HER2) protein expression. Predicting both low- and high-level expressions of HER2 remains challenging. Moreover, the joint analysis of H&E and immunohistochemistry (IHC) stained images for HER2 scoring is difficult. Although several deep learning methods have been investigated to address this clinical task, they fail to provide pixel-level localization of HER2 status. In this study, we propose a single end-to-end pipeline using a system of vision transformers with HER2 status scoring on whole slide images of WSIs. The method includes patch-wise processing of H&E WSIs for tumor localization. A novel mapping function is proposed to correspondingly identify correlated IHC WSI regions with malignant regions on H&E. A clinically inspired HER2 scoring mechanism is embedded in the pipeline and allows for automatic pixel-level annotation of 4-way HER2 scoring (0, 1+, 2+, and 3+). Also, the proposed method accurately returns HER2-negative and HER2-positive. Privately curated datasets were collaboratively extracted from 13 different cases of WSIs of H&E and IHC. A thorough experiment was conducted on the proposed method. Results obtained showed good classification accuracy during tumor localization. Also, a classification accuracy of 0.94 and a specificity of 0.933 were obtained for the prediction of HER2 status, scoring in the 4-way methods. The applicability of the proposed pipeline was investigated using WSIs patches as comparable to human pathologists. Findings from the study showed the usability of jointly evaluated H&E and IHC images on end-to-end ViT-based models for HER2 scoring.
Computational depression detection has reached a saturation point with standard end-to-end deep learning paradigms. To navigate the rapid proliferation of highly complex models, this systematic scoping review analyzes 87 primary post-LLM studies, spanning both “in-the-wild” social media and structured clinical interactions, published primarily between January 2024 and April 2026. Evaluated alongside 7 foundational 2023 baselines (totaling 94 core methodology papers out of 111 cited references), these works mark a distinct methodological transition from isolated feature engineering to multimodal and architectural synthesis. We introduce a novel taxonomy that categorizes these emerging approaches into four information fusion paradigms: (1) Architecture Fusion, which hybridizes heterogeneous neural components to capture complementary spatiotemporal dependencies; (2) Knowledge Fusion, which injects structured external clinical knowledge (e.g., DSM-5 ontologies) into neural pipelines to bridge the semantic gap; (3) Generative Fusion, which integrates in-context learning (ICL), retrieval-augmented generation (RAG), and parameter-efficient fine-tuning (PEFT); and (4) Data-Level Fusion, which addresses multimodal alignment, dynamic fusion timing, and temporal modeling. Our critical analysis reveals that while zero-shot large language models (LLMs) offer broad semantic reasoning capabilities, fine-tuned encoders and hybrid architectures integrated with domain-specific knowledge currently provide a more rigorously validated balance of reliability, explainability, and performance. We conclude by exposing the “illusion of accuracy” prevalent in proxy-labeled benchmarks and outlining strategic directions for developing clinically valid, privacy-preserving detection systems suitable for real-world clinical workflows.
Calcium salt deposits, commonly referred to as calcifications, can form in various body tissues. Among them, calcifications in the cervical carotid arteries are of particular clinical significance due to their strong association with cardiovascular disease. These calcifications are often detected as incidental findings, that is, unexpected observations that may carry greater clinical importance than the original purpose of the imaging. Given their diagnostic relevance, automated detection and segmentation of such calcifications in medical images are critical for enabling early diagnosis and improving clinical decision-making. Despite their importance, detecting carotid artery calcifications presents several challenges: their small size, sparse distribution, and visual similarity to surrounding anatomical structures make them difficult to identify accurately. To address these challenges, this study proposes a multi-stage workflow for segmenting carotid artery calcifications. First, we introduce a novel 3D dataset comprising Cone Beam Computed Tomography (CBCT) scans that contain incidental neck calcifications. For the segmentation task, we adapt 3D UNETR, a vision transformer-based model. We then employ a 3D Grad-CAM technique to generate class activation maps, which are subsequently used in a post-processing step to refine the segmentation results. The proposed workflow achieves a Dice coefficient of 64% for segmentation, despite calcifications occupying only 0.04% of the total image volume. Beyond its strong performance on such sparse targets, the approach remains computationally efficient, with modest memory requirements and fast inference times, making it practical for integration into clinical imaging workflows.
Artificial Intelligence (AI) has emerged as a transformational tool for breast cancer diagnosis, facilitating swift and precise interpretation of medical imaging while mitigating human error. This systematic review explores AI-driven approaches for breast cancer diagnosis using multiple imaging modalities, including mammography, ultrasound, magnetic resonance imaging, and thermography. In addition to summarizing imaging techniques, this study outlines breast cancer progression patterns and stage classification based on standardized tumor size criteria, providing foundational insights for early-stage researchers. Following PRISMA guidelines, 60 peer-reviewed studies published between 2015 and 2026 were analyzed to evaluate data modalities, algorithmic frameworks, and validation strategies used in AI-based detection systems. This review highlights advances in deep learning models, such as convolutional neural networks, transformer-based architectures, hybrid frameworks, and multimodal fusion techniques. Publicly available datasets and evaluation metrics were also examined. Furthermore, this study proposes an AI-driven multimodal healthcare analytics framework for accurate breast cancer diagnosis. By synthesizing recent advances and practical considerations, this study offers clinicians and researchers a valuable resource on current AI-driven methods for breast cancer diagnosis.
Background: Healthcare systems face challenges scaling predictive analytics across institutional boundaries while maintaining data privacy. Federated learning enables collaborative model training without sharing raw patient data, but prior studies have not isolated the effect of federation from confounding factors such as algorithm switching. Methods: We implemented federated XGBoost across 47 U.S. healthcare institutions (January 2023–June 2024). The final analytic cohort comprised 992,000 adult admissions after applying exclusion criteria and privacy-preserving deduplication. To isolate the federation effect from algorithm switching, we conducted a controlled comparison on the 7 sites that already used single-institution XGBoost as baseline, holding algorithm, features, and hyperparameters constant. Results: In the controlled comparison (same XGBoost algorithm, with vs. without federation), federation contributed ＋0.02 AUROC (95% CI: –0.01 to ＋0.05, p = 0.18), which was not statistically significant. The headline improvement of 0.08 AUROC over all baselines was primarily attributable to algorithm switching (logistic regression→ XGBoost, which contributed ＋0.11 for those 28 sites). Resource optimization achieved 70.9% (net after differential privacy overhead) network bandwidth reduction. Maximum pairwise subgroup AUROC disparity was 6.6% (uninsured vs. privately insured), which decreased to 3.6% by study end after three mitigation interventions. External validation on MIMIC-IV achieved AUROC 0.742 (95% CI: 0.730–0.754). The system maintained 99.1% operational uptime with central differential privacy guarantees (ϵ=1.0, δ=10−5 as primary conservative bound). Conclusion: Federated learning for readmission prediction demonstrated modest additional benefit beyond using XGBoost locally (＋0.02 AUROC, not statistically significant). The primary practical advantages were facilitating algorithm standardization across heterogeneous sites and enabling privacy-preserving multi-institutional collaboration. Researchers should carefully isolate federation effects from algorithm switching when evaluating federated learning systems.
Background and objective Our previous work demonstrated that a Random Forest model could successfully automate the 3-class classification of clinical research activities, significantly improving efficiency. Building on this success, this project aimed to enhance classification granularity and accuracy by expanding the dataset, refining the classification schema to four categories, and evaluating a new range of state-of-the-art Machine Learning (ML) models, including Large Language Models (LLMs) and Convolutional Neural Networks (CNNs). Methods The original dataset was expanded to approximately 1230 clinical studies. The classification schema was refined to distinguish between Interventional Studies (IS), Primary Data Collection Non-Interventional Studies (PDC NIS), Secondary Data Use Non-Interventional Studies (SDU NIS) and Real-World Evidence Scientific Projects (RWE SP). We conducted an extensive evaluation of 17 models, including ClinicalBioBERT, RoBERTa and ResNet, against our Random Forest baseline. Both a direct 4-class classification and a 2-step hierarchical approach were tested. Advanced techniques for overfitting mitigation, including early stopping, class-weighted losses, regularisation and Bayesian hyperparameter optimisation, were systematically implemented. Results While advanced LLMs demonstrated strong initial performance, their advantage diminished after rigorous regularisation, a 2-step hierarchical Random Forest model emerged as the superior solution after rigorous testing and optimisation. It achieved an overall accuracy of 95.12% and an F1-score of 95.03%. This approach outperformed both direct 4-class models and the more computationally intensive LLMs, offering an optimal balance of performance, stability and a 41% reduction in execution time. The final model was successfully integrated into a classification application. Conclusion This research confirms that a more granular and robust classification of clinical research can be achieved through ML classification techniques. The success of the hierarchical Random Forest model provides compelling evidence that for this specific, domain-constrained task, a tailored, simpler architecture offers a superior balance of accuracy, speed and interpretability over complex deep learning models accelerating regulatory compliance in the biopharmaceutical industry.
One of the common types of cancer is that of the bladder, and it is important to detect this condition early for effective treatment, as well as ensuring survival rates are higher. In this paper, we suggest a comprehensive deep learning approach for automatic detection and classification of bladder cancer using T2-weighted Magnetic Resonance Imaging (MRI). This was done by initially processing the first set of bladder MRI, with resizing augmentations applied to equalize the sizes of the images, as well as adding variance to the samples. A Lightweight Context-Aware Network (LCANet) was later used to precisely trace the bladder and the tumor within the neural network. The tumor layer was classified with a Multi-Stacked Layered Graph Attention Network (MSL-GAT) because it is proficient at obtaining features and able to distinguish between non-muscle-invasive (NMIBC) and muscle-invasive (MIBC) bladder cancer. The Leopard Seal Optimization (LSO) algorithm was also utilized for further optimization of major parameters within the network to ensure its convergence and accuracy. The developed model achieved 98.06% AUC and proved to be accurate in its classification of the data from two clinical centers, showing low error rates during classifying. These results suggest that the suggested model is highly generalizable and applicable. The study shows that it may be possible to use the proposed approach to enable radiologists to early diagnose and manage the condition.
