Machine Learning (ML) techniques have become valuable tools in healthcare for tasks such as disease diagnosis, treatment planning, and clinical decision-making. However, their application often requires specialized expertise and considerable development time. Automated Machine Learning (AutoML) has emerged to address these challenges by automating key steps of the ML pipeline, thereby enhancing the efficiency and accessibility of ML integration into clinical workflows. This paper presents a systematic literature mapping following a structured protocol across multiple academic databases. A total of 244 studies published between 2016 and 2025 were analyzed from 171 distinct sources. Most studies employed AutoML for diagnosis prediction (52.8%) and prognosis prediction (31.9%), primarily using tabular (43.4%) and image (31.5%) data. Classification tasks dominated (81.1%), followed by regression tasks (9.4%). Model selection emerged as the primary challenge (25.3%) among the studies not utilizing managed AutoML tools, while data preprocessing remained critical (13.7%) even when using managed services. Although Explainable AI (XAI) was incorporated in only 30.7% of the studies, its adoption showed a notable increase in 2024. These findings indicate that while the application of AutoML in medicine is growing, its black-box nature continues to limit its adoption in interpretability-critical domains. The integration of XAI techniques with AutoML is an emerging trend that seeks to address this limitation. However, further research is necessary to evaluate the clinical impact of these combined approaches and to enhance trust in AutoML-driven decision support systems.

Four-dimensional (4D) flow magnetic resonance imaging (MRI) has evolved into an advanced non-invasive imaging technique that enables comprehensive assessment of blood flow in cardiovascular system. Hemodynamic data from 4D flow MRI provide key biomarkers such as wall shear stress (WSS), turbulent kinetic energy (TKE), and viscous energy loss, which aid in the analysis of cardiovascular diseases. However, its clinical application is limited by challenges such as prolonged scan times and limited spatiotemporal resolution. Conventional algorithms have been developed to automate the reconstruction of sparse data and perform image segmentation for hemodynamic quantification, but these methods are often time-consuming and require user expertise for accurate postprocessing. To address these limitations, artificial intelligence (AI), particularly deep learning (DL) techniques, has been introduced. DL models have shown promise in accelerating scan times by reconstructing sparsely sampled data into fully sampled datasets and enhancing image resolution by combining computational fluid dynamics (CFD) with 4D flow MRI data, as well as improving data quality through noise reduction. In addition, automated segmentation techniques have been developed to reduce user intervention, enabling more consistent and efficient analysis. Many researchers are working on DL-based approaches to 4D flow MRI using limited datasets. Recently, the lack of systematic methodologies has made it difficult to identify appropriate approaches. This paper aims to provide a comprehensive review of the latest AI applications using 4D flow MRI, enabling researchers to access and evaluate existing research more effectively.

To synthesize existing evidence and compare the predictive performance of conventional statistical (CS) models versus machine learning (ML) methods in pharmacoepidemiology.

Ultrasound imaging reveals eye morphology and aids in diagnosing and treating eye diseases. However, interpreting diagnostic reports requires specialized physicians. We present a labeled ophthalmic dataset for the precise analysis and the automated exploration of medical images along with their associated reports. It collects three modal data, including the ultrasound images, blood flow information and examination reports from 1,0361 patients at an ophthalmology hospital in Shenyang, China, during the year 2016 to 2020, in which the patient information is de-identified for privacy protection. To the best of our knowledge, it is the only ophthalmic dataset that contains the three modal information simultaneously. It incrementally consists of 2,2173 images with the corresponding free-text reports, which describe 10 typical imaging findings of intraocular diseases and the corresponding anatomical locations. Each image shows three kinds of blood flow indices at three specific arteries, i.e., nine parameter values to describe the spectral characteristics of blood flow distribution. The reports were written by ophthalmologists during the clinical care. In addition, the knowledge fusion cross modal network (KFCMN) is proposed to generate report according to the proposed dataset. The experimental results demonstrate that our dataset is suitable for training supervised models concerning cross-modal medical data.

Accurate estimation of polyp size during colonoscopy is critical for risk assessment and surveillance planning in colorectal cancer prevention. However, current methods often rely on subjective visual judgment, leading to inconsistencies and potential misclassification. This study proposes a deep learning framework that enables automated and objective polyp size classification by integrating RGB and depth information. The approach leverages a modified Af-SfM module to generate refined and rectified depth maps, which are combined with RGB inputs to support classification into clinically relevant size categories. The model was trained and validated on a dataset of over 10,000 annotated colonoscopic images curated by expert gastroenterologists. Experimental results demonstrate that incorporating rectified depth information significantly improves classification performance over RGB-only baselines. For polyps measuring 10 mm or larger, the system achieved a precision of 91.5% and a recall of 93.6%. These findings highlight the potential of depth-enhanced deep learning methods to support more consistent and accurate polyp size estimation in clinical endoscopy.

The increasing digitisation of healthcare data and the rapid development of Artificial Intelligence (AI) pave the way for innovative strategies for infectious disease management. This study aimed to systematically retrieve and summarize current evidence on the use and performance of AI-based models for healthcare-associated infection (HAI) detection (i.e., identifying infections already present in available data) and prediction (i.e., estimating future risk based on earlier patient information).

The rapid development of large language models (LLMs) has accelerated research into applying artificial intelligence (AI) to domains such as medical question answering and clinical decision support. However, LLMs face substantial limitations in medical contexts due to challenges in understanding specialized terminology, complex contextual information, hallucination issues (i.e., generating incorrect responses), and the black-box nature of their reasoning processes. To address these issues, methods like retrieval-augmented generation (RAG) and its graph-based variant, GraphRAG, have been proposed to incorporate external knowledge into LLMs. Nonetheless, these approaches often rely heavily on external resources and increase system complexity. In this study, we introduce MedSumGraph, a medical question-answering system that enhances GraphRAG by integrating structured medical knowledge summaries and optimized prompt designs. Our method enables LLMs to better interpret domain-specific knowledge without requiring additional training, and it enhances the reliability and interpretability of responses by directly embedding factual evidence and graph-based reasoning into the generation process. MedSumGraph achieves competitive performance on two out of eight multiple-choice medical QA benchmarks, including MedQA (USMLE), outperforming closed-source LLMs and domain-specific foundation models. Moreover, it generalizes effectively to open-domain QA tasks, yielding significant gains in reasoning over common knowledge and evaluating the truthfulness of answers. These findings demonstrate the potential of structured summarization and graph-based reasoning in enhancing the trustworthiness and versatility of LLM-driven medical AI systems.

Accurate pixel-level segmentation of ultrasound (US) images is vital for computer-aided disease screening, diagnosis, and treatment response evaluation. The weakly supervised methods have the potential to reduce the time-consuming and labor-intensive workload for radiologists, paving the way for further automation in the quantitative analysis of US images. Among these methods, the multiple instance learning (MIL) has proven effective and is often applied to prediction tasks with insufficiently labeled data. In US examinations, the elliptical region formed by intersecting lines used by radiologists for target annotation serves as a crucial prior information. Therefore, we propose a novel weakly supervised method called elliptical shape prior constraint MIL (ESPC-MIL) for pixel-level segmentation of US images. ESPC-MIL incorporates an elliptical shape prior constraint into the MIL framework, delivering more accurate foreground and background candidate regions for MIL, which enhances its predictive performance for tissues and organs with approximately elliptical shapes. Furthermore, the method utilizes elliptical shape prior information for global supervision, improving edge segmentation and localization accuracy. Compared to other weakly supervised methods, ESPC-MIL achieves state-of-the-art results on four US image datasets: Achilles tendon dataset, median nerve dataset, private breast tumor dataset, and public breast ultrasound image dataset, with Dice similarity coefficients of 0.855, 0.849, 0.876, and 0.748, respectively. It demonstrates performance comparable to fully supervised segmentation methods while significantly reducing annotation requirements. Notably, the method demonstrates a more significant performance improvement in segmenting objects with approximately elliptical shapes compared to those with complex shapes. Source codes and models are available at https://github.com/CYWang-kayla/ESPC-MIL-Model.

Autism spectrum disorder (ASD) is a complex neurodevelopmental condition whose rising prevalence places increasing demands on a lengthy diagnostic process. Machine learning (ML) has shown promise in automating ASD diagnosis, but most existing models operate as black boxes and are typically trained on a single dataset, limiting their generalizability. In this study, we introduce a transparent and interpretable ML approach that leverages BioBERT, a state-of-the-art language model, to analyze unstructured clinical text. The model is trained to label descriptions of behaviors and map them to diagnostic criteria, which are then used to assign a final label (ASD or not). We evaluate transfer learning, the ability to transfer knowledge to new data, using two distinct real-world datasets. We trained on datasets sequentially and mixed together and compared the performance of the best models and their ability to transfer to new data. We also created a black-box approach and repeated this transfer process for comparison. Our transparent model demonstrated robust performance, with the mixed-data training strategy yielding the best results (97 % sensitivity, 98 % specificity). Sequential training across datasets led to a slight drop in performance, highlighting the importance of training data order. The black-box model performed worse (90 % sensitivity, 96 % specificity) when trained sequentially or with mixed data. Overall, our transparent approach outperformed the black-box approach. Mixing datasets during training resulted in slightly better performance and should be the preferred approach when practically possible. This work paves the way for more trustworthy, generalizable, and clinically actionable AI tools in neurodevelopmental diagnostics.

Surgical workflow recognition has the potential to accelerate training initiatives through the analysis of surgical videos, improve intraoperative efficiency, and support preemptive postoperative care. Unlike well-explored minimally invasive surgeries, where surgical workflows are consistent across patients, automating endoscopic pituitary surgery workflow recognition is challenging. Pituitary surgery involves a large number of steps, diverse sequences, optional steps, and frequent transitions, making it challenging for current state-of-the-art (SOTA) methods, which struggle with transferability. Progress is largely limited by the lack of annotated data that captures the complexity of pituitary surgery, and obtaining such annotations is both time-consuming and resource-intensive. This paper presents SurgflowNet, a novel spatio-temporal model for consistent pituitary workflow recognition leveraging unannotated data. We utilise a limited yet fully annotated dataset to infer quasi-labels for unannotated videos and curate a balanced dataset to train a robust frame encoder using the student-teacher framework. A spatio-temporal network that combines the resulting frame encoder and an LSTM network is trained with a consistency loss to ensure stability in step predictions. With a 5% improvement in macro F-score and 13.4% in Edit Score over the SOTA, SurgflowNetdemonstrates a significant improvement in workflow recognition for endoscopic pituitary surgery.

Abdominal aortic aneurysms (AAAs) are progressive focal dilatations of the abdominal aorta. AAAs may rupture, with fatal consequences in >80% of cases. Current clinical guidelines recommend elective surgical repair when the maximum AAA diameter exceeds 55 mm in men or 50 mm in women. Patients that do not meet these criteria are periodically monitored, with surveillance intervals based on the maximum AAA diameter. However, this diameter does not take into account the complex relation between the 3D AAA shape and its growth, making standardized intervals potentially unfit. Personalized AAA growth predictions could improve monitoring strategies. We propose to use an SE(3)-symmetric transformer model to predict AAA growth directly on the vascular model surface enriched with local, multi-physical features. In contrast to other works which have parameterized the AAA shape, this representation preserves the vascular surface's anatomical structure and geometric fidelity. We train our model using a longitudinal dataset of 113 computed tomography angiography (CTA) scans of 24 AAA patients at irregularly sampled intervals. After training, our model predicts AAA growth to the next scan moment with a median diameter error of 1.18 mm. We further demonstrate our model's utility to identify whether a patient will become eligible for elective repair within two years (acc = 0.93). Finally, we evaluate our model's generalization on an external validation set consisting of 25 CTAs from 7 AAA patients from a different hospital. Our results show that local directional AAA growth prediction from the vascular surface is feasible and may contribute to personalized surveillance strategies.

Data Augmentation (DA) aims to create a new dataset to address the lack of data in various domains. Particularly in few-shot scenarios of the biomedical Named Entity Recognition (NER) domain, an effective DA method can enhance data diversity, reduce overfitting, and significantly improve the model's generalization ability. In this work, we propose a novel DA method for NER tasks, which uses ChatGPT and prompt learning to extract high-quality data from large language models. The entity recognition tasks are then performed via transfer learning and efficient decoding strategies. Moreover, this study conducted extensive experiments on four publicly available biomedical datasets (BC5CDR, NCBI, BioNLP11EPI, and BioNLP13GE), demonstrating that our methods exhibit strong stability and entity recognition capabilities even in extremely limited scenarios. In the 5-shot, 20-shot, and 50-shot scenarios, the average F1 scores of the four datasets reached 72.96%, 75.05%, and 77.42%, respectively.

The integration of artificial intelligence (AI) into the field of mental health diagnosis has garnered increasing scholarly and clinical attention, particularly in relation to the early detection and classification of depression. This study offers a comprehensive review of the current landscape of AI-driven approaches for depression diagnosis, examining the methodologies, data modalities, and performance metrics employed across recent empirical investigations. Emphasizing machine learning and deep learning techniques, the study critically evaluates the utility of linguistic, behavioral, and physiological data sourced from social media, clinical interviews, speech recordings, and wearable devices. The findings suggest that AI systems, particularly those incorporating multimodal data fusion and advanced neural network architectures, demonstrate promising diagnostic accuracy and the potential to augment traditional psychiatric assessments. However, the study also identifies significant methodological, ethical, and practical challenges, including issues of dataset bias, algorithmic transparency, and clinical applicability. In response, the paper outlines key future directions aimed at improving model generalizability, enhancing interpretability, and fostering ethically responsible deployment in real-world settings. This review not only elucidates the transformative capacity of AI in mental health diagnostics but also provides a roadmap for advancing the development of robust, transparent, and clinically integrated AI systems for the detection of depression.

Alzheimer's disease (AD) is a serious neurodegenerative brain disorder with complex pathophysiology. While currently available drugs can provide symptomatic benefits, they often fail to cure the disease. Thus, there is an urgent need to explore new therapeutic agents. In this study, we developed DTIP (Drug-Target Interaction Prediction), a machine learning-based approach to search novel drugs for AD by utilizing the information of drug-target interaction (DTI). By training a Skip-gram model on drug-target sequences derived from known DTI information, the algorithm learned the drug-target relationship embeddings and to predict potential drug candidates for diseases like AD. For AD, we compiled 917 risk genes and identified 292 potential drugs via the new algorithm. We further performed molecular docking by AutoDock Vina and conducted Inverted Gene Set Enrichment Analysis (IGSEA) on these drug candidates. Our results identified that several drugs could be promising for AD treatment, including human C1-esterase inhibitor, quetiapine, dasatinib, miconazole, aniracetam, chlorpromazine, hypericin, entrectinib, torcetrapib, bosutinib, sunitinib, aniracetam, rosiglitazone, tarenflurbil, milrinone, and MITO-4509. Results from this study also provided insights for understanding the molecular mechanisms underlying AD. As a systematic and versatile method, our approach can also be applied to identify efficacious therapies for other complex diseases.

The invasiveness prediction in renal cell carcinoma (RCC) is of significant importance for the decision of clinical surgical plans and the patients' prognosis. Currently, besides invasive pathological assessment, it mainly relies on observation through computed tomography (CT) imaging. However, limitations of human vision and qualitative descriptions restrict the accuracy of the diagnosis of renal sinus invasion (RSI). Recently, artificial intelligence approaches have shown promising prospects in cancer diagnosis. Due to the complex imaging characteristics of invasiveness, prediction models that only focus on tumor regions are inadequate, requiring comprehensive evaluation of intratumoral heterogeneity, peritumoral information, and the kidney in which the tumor resides. Therefore, in this study, we propose a context-aware heterogeneous graph neural network for multi-level description and invasiveness prediction in RCC. The superiority of the proposed model lies in its ability to integrate imaging features at multi-level, and to learn disturbance invariant features through a data-driven diffusion perturbation strategy. To evaluate the effectiveness and generalization of our model, we conduct extensive experiments on a multi-center dataset (including CT scan images of 437 patients) to compare our model with a series of state-of-the-art (SOTA) classification models. The experimental results show the superiority of our model for RSI classification (AUC=0.88). Additionally, we also perform a comparative study with clinical experts, and the proposed method is significantly better than existing assessment methods and clinical experts (p<0.05). In general, our work provides an effective assessment tool for automated diagnosis of RSI in RCC and also offers new insights for constructing more precise tumor prediction models.

Rheumatoid arthritis (RA) is a chronic systemic autoimmune disorder characterized by progressive destruction of synovial joints, for which precise early diagnosis is critical to effective clinical management. Although current computer-aided diagnosis (CAD) systems show promising potential, their practical deployment remains challenged by heterogeneity in multimodal data and the inherent complexity of pathophysiological manifestations. A key limitation of existing approaches is the absence of dedicated mechanisms to reconcile inter-modal conflicts and to adequately leverage the diagnostic information present in contradictory samples, such as discordances between laboratory findings and clinical descriptions, which often leads to suboptimal diagnostic performance. To overcome these challenges, this study introduces a novel auxiliary diagnostic framework for RA based on Contradiction-Aware Orthogonal Contrastive Learning (CA-OCL) and a Confidence-guided Hierarchical Attention Network (CHAN). The proposed architecture incorporates three major innovations. First, in the data processing stage, a lightweight feature-crossing network is employed to derive robust representations from structured data, while a multi-task adapted extension of the BERT model is utilized to extract rich semantic features from unstructured textual inputs. Second, the CA-OCL module is designed to explicitly identify and learn from contradictory negative samples-a capability largely absent in conventional contrastive learning frameworks. Additionally, orthogonal constraints are applied to minimize feature redundancy across modalities, thereby preserving discriminative modality-specific information that is often obscured by methods promoting excessive feature alignment. Finally, the CHAN module dynamically modulates inter-modal contributions using confidence estimates, mitigating the risk of unilateral dominance by any single modality, a common drawback in attention-based fusion mechanisms and facilitating refined integration of conflicting information. Comprehensive experimental evaluations demonstrate that the proposed framework achieves superior performance compared to state-of-the-art methods. These results not only validate the efficacy of our approach in handling multimodal conflicts and exploiting contradictory evidence, but also highlight its significant clinical utility through effective multimodal data integration. This work addresses critical limitations in conventional CAD systems and provides an advanced paradigm for intelligent diagnostic assessment of RA.

The purpose of this review was to comprehensively explore the landscape of recently published literature on the applications of artificial intelligence (AI) in predicting individualized patient waiting times in an emergency department (ED) and identify pertinent considerations for practitioners and hospital decision-makers.

Improving the AI-driven automated medical encoding of clinical text plays a vital role in gathering information on the occurrence of diseases to improve population-level health. This work presents a novel attention mechanism designed to enhance text classification models and ensure appropriate classification of medical concepts in unstructured electronic health records.

Large Language Model-Driven Generative Recommender Systems (LLM-GRSs) are playing a growing role in healthcare, particularly in clinical question-answering. This study reviews their corpus sources, customization techniques, and evaluation metrics.

Depression is a significant global health issue with increasing prevalence. Current diagnostic methods rely on subjective observations and questionnaires, often resulting in underestimation of the condition and insufficient treatment. This study investigates voice-based markers for detecting depressive symptoms through a novel system of virtual humans (VHs) capable of engaging in open-ended talks, unlike previous research which relied primarily on structured clinical interview formats. A total of 101 participants (42 with depressive symptoms) engaged in six casual social interactions with VHs simulating basic emotions, forming the DEPTALK dataset. Speech recordings and their automatic transcriptions were processed using state-of-the-art pre-trained transformer-based models to generate embeddings. We first employed a conversation-level aggregation strategy, combining embeddings across each dialogue and classifying them with Extreme Gradient Boosting. A single model trained on all six conversations per participant outperformed emotion-specific models, achieving F1 scores of 0.566 for speech, 0.329 for text, and 0.648 for the multimodal fusion, indicating that aggregating emotionally diverse interactions exposes stronger depression cues. To capture temporal dynamics, we further implemented a turn-level aggregation strategy using Gated Recurrent Units and training on all conversations. This approach improved performance for text (F1 = 0.505) and maintained competitive results for speech (F1 = 0.541), although the multimodal GRU model (F1 = 0.556) did not surpass the best conversation-level model. Overall, findings suggest that in casual conversations, depressive symptoms are primarily conveyed through prosody, with the addition of semantic context further enhancing detection. This study advances the understanding of speech-based depression patterns in simulated social interactions and highlights the potential of using VHs for more objective depressive symptoms detection.

Early detection of Alzheimer's disease (AD) is essential for effective clinical intervention and disease management. However, conventional Deep Learning (DL) methods face limitations in analyzing complex brain magnetic resonance imaging (MRI), especially when training data are scarce. In this study, we propose a Quantum-Enhanced Neural Network Architecture (QENNA) that integrates quantum convolutional layers with classical deep learning to improve diagnostic accuracy in early AD detection. The model also incorporates quantum data augmentation strategies, including Quantum Generative Adversarial Networks (QGANs) and quantum random walks, to generate high-fidelity synthetic MRI scans and address training data limitations. Experiments on two public MRI datasets demonstrate that QENNA achieves up to 93.0 % accuracy and 96.0 % Area Under the Curve (AUC), outperforming state-of-the-art classical models. Ablation studies confirm that the quantum components substantially enhance performance. These results suggest that quantum-enhanced learning frameworks can significantly advance Artificial Intelligence (AI)-driven diagnostic tools for neurodegenerative disorders and support scalable, early-stage AD screening in clinical practice.