OBJECTIVE: This survey aims to systematically map the rapidly evolving landscape of AI agents in healthcare. It addresses the critical need to adapt general-purpose agentic frameworks characterized by autonomy, planning, and tool use to the high-stakes, safety-critical constraints of medical decision-making and patient care. METHODS: We conducted a comprehensive review of over 200 recent studies, synthesizing literature from major academic databases. We developed a holistic taxonomy that traces the full lifecycle of healthcare agents, analyzing perception modalities, core technical architectures, and evaluation protocols specific to autonomous systems. RESULTS: The review presents a quantitative landscape analysis showing exponential growth in the field. We structure the domain into three pillars: (1) Perception of multi-modal clinical data (e.g., EHR, imaging, genomics); (2) Agent Capabilities, including tool use, reasoning, memory, and multi-agent collaboration; and (3) an Application Ecosystem organized by stakeholder roles (clinicians, patients, researchers, and administrators). Additionally, we categorize evaluation frameworks, and discuss the deployment readiness of current systems across technical, evidentiary, and governance dimensions. Finally, we identify challenges for advancing healthcare agents from controlled evaluation toward real-world clinical integration. A continuously updated repository of related papers is available at https://github.com/AgenticHealthAI/Awesome-AI-Agents-for-Healthcare. CONCLUSION: AI agents offer significant potential to enhance healthcare through autonomous reasoning and workflow integration. However, current research remains largely concentrated in benchmark and controlled evaluation settings, and the translation into clinical practice will require advances in reliability, privacy protection, governance, and operational integration.

Dealing with noisy labels (LNL) emerges as a critical challenge when applying deep learning (DL) in practical settings. Previous methodologies primarily concentrated on harnessing model predictions to mitigate the impact of noisy labels. Nevertheless, their efficacy is strongly contingent on the accuracy of model predictions, a factor that cannot be assured in the context of LNL. Our empirical analysis shows that in noisy datasets, the spatial information of latent feature representation combined with original noisy labels is more robust than the methods using model predictions. To mitigate the unreliability introduced by model predictions, we propose a novel Feature Representation method, which utilizes noisy labels of nearest neighbors for label Correction and sample Selection (FRCS). Extensive experiments on various benchmark datasets demonstrate the superiority of FRCS compared with SOTA methods. Our codes are available at https://github.com/tianfangjh/FRCS-Noisy-Labels.

Modern programs, characterized by their complex functionalities, high integration, and rapid iteration cycles, are prone to errors. This complexity poses challenges in program analysis and software testing, making it difficult to achieve comprehensive bug coverage during the development phase. As a result, many bugs are only discovered during the software’s production phase. Tracking and understanding these field bugs is essential but challenging: the uploaded field error reports are extensive, and trivial yet high-frequency bugs can overshadow important low-frequency bugs. Additionally, application codebases evolve rapidly, causing a single bug to produce varied exceptions and stack traces across different code releases. In this paper, we introduce Foliage, a bug tracking and clustering toolchain designed to trace and characterize field bugs in JavaScript applications, aiding developers in locating and fixing these bugs. To address the challenges of efficiently tracking and analyzing the dynamic and complex nature of software bugs, Foliage proposes an error message enhancement technique. Foliage also introduces the verbal-characteristic-based clustering technique, along with three evaluation metrics for bug clustering: V-measure, cardinality bias, and hit rate. The results show that Foliage’s verbal-characteristic-based bug clustering outperforms previous bug clustering approaches by an average of 31.1% across these three metrics. We present an empirical study of Foliage applied to a complex real-world application over a two-year production period, capturing over 250,000 error reports and clustering them into 132 unique bugs. Finally, we open-source a bug dataset consisting of real and labeled error reports, which can be used to benchmark bug clustering techniques.

Text-to-image diffusion has attracted vast attention due to its impressive image-generation capabilities. However, when it comes to human-centric text-to-image generation, particularly in the context of faces and hands, the results often fall short of naturalness due to insufficient training priors. We alleviate the issue in this work from two perspectives. 1) From the data aspect, we carefully collect a human-centric dataset comprising over one million high-quality human-in-the-scene images and two specific sets of close-up images of faces and hands. These datasets collectively provide a rich prior knowledge base to enhance the human-centric image generation capabilities of the diffusion model. 2) On the methodological front, we propose a simple yet effective method called Mixture of Low-rank Experts (MoLE) by considering low-rank modules trained on close-up hand and face images respectively as experts. This concept draws inspiration from our observation of low-rank refinement, where a low-rank module trained by a customized close-up dataset has the potential to enhance the corresponding image part when applied at an appropriate scale. To validate the superiority of MoLE in the context of human-centric image generation compared to state-of-the-art, we construct two benchmarks and perform evaluations with diverse metrics and human studies. Datasets, model, and code are released at https://sites.google.com/view/mole4diffuser/.

We demonstrate that AI models can accurately segment liver tumors without the need for manual annotation by using synthetic tumors in CT scans. Our synthetic tumors have two intriguing advantages: (I) realistic in shape and texture, which even medical professionals can confuse with real tumors; (II) effective for training AI models, which can perform liver tumor segmentation similarly to the model trained on real tumors—this result is exciting because no existing work, using synthetic tumors only, has thus far reached a similar or even close performance to real tumors. This result also implies that manual efforts for annotating tumors voxel by voxel (which took years to create) can be significantly reduced in the future. Moreover, our synthetic tumors can automatically generate many examples of small (or even tiny) synthetic tumors and have the potential to im-prove the success rate of detecting small liver tumors, which is critical for detecting the early stages of cancer. In addition to enriching the training data, our synthesizing strategy also enables us to rigorously assess the AI robustness.

Well-annotated medical datasets enable deep neural networks (DNNs) to gain strong power in extracting lesion-related features. Building such large and well-designed medical datasets is costly due to the need for high-level expertise. Model pre-training based on ImageNet is a common practice to gain better generalization when the data amount is limited. However, it suffers from the domain gap between natural and medical images. In this work, we pre-train DNNs on ultrasound (US) domains instead of ImageNet to reduce the domain gap in medical US applications. To learn US image representations based on unlabeled US videos, we propose a novel meta-learning-based contrastive learning method, namely Meta Ultrasound Contrastive Learning (Meta-USCL). To tackle the key challenge of obtaining semantically consistent sample pairs for contrastive learning, we present a positive pair generation module along with an automatic sample weighting module based on meta-learning. Experimental results on multiple computer-aided diagnosis (CAD) problems, including pneumonia detection, breast cancer classification, and breast tumor segmentation, show that the proposed self-supervised method reaches state-of-the-art (SOTA). The codes are available at https://github.com/Schuture/Meta-USCL.

Modern RESTful services expose RESTful APIs to integrate with diversified applications. Most RESTful API parameters are weakly typed, which greatly increases the possible input value space. Weakly-typed parameters pose difficulties for automated testing tools to generate effective test cases to reveal web service defects related to parameter validation. We call this phenomenon the type collapse problem. To remedy this problem, we introduce FET (Format-encoded Type) techniques, including the FET, the FET lattice, and the FET inference to model fine-grained information for API parameters. Inferred FET can enhance parameter validation, such as generating a parameter validator for a certain RESTful server. Enhanced by FET techniques, automated testing tools can generate targeted test cases. We demonstrate Leif, a trace-driven fuzzing tool, as a proof-of-concept implementation of FET techniques. Experiment results on 27 commercial services show that FET inference precisely captures documented parameter definitions, which helps Leif discover 11 new bugs and reduce$72\% - 86\%$fuzzing time compared to state-of-the-art fuzzers. Leveraged by the inter-parameter dependency inference, Leif saves$15\%$fuzzing time.

Abstract Modern RESTful services expose RESTful APIs to integrate with diversified applications. Most RESTful API parameters are weakly typed, which greatly increases the possible input value space. This poses difficulties for automated testing tools to generate effective test cases to reveal web service defects related to parameter validation. We call this phenomenon the type collapse problem. To remedy this problem, we introduce FET (Format-encoded Type) techniques, including the FET, the FET lattice, and the FET inference to model fine-grained information for API parameters. Enhanced by FET techniques, automated testing tools can generate targeted test cases. We demonstrate Leif, a trace-driven fuzzing tool, as a proof-of-concept implementation of FET techniques. Experiment results on 27 commercial services show that FET inference precisely captures documented parameter definitions, which helps Leif to discover 11 new bugs and reduce $$72\% \sim 86\%$$ 72 % ∼ 86 % fuzzing time as compared to state-of-the-art fuzzers.

In this paper, a robot, namely iLeg, is designed for the purpose of rehabilitation of patients with hemiplegia or paraplegia. The iLeg is composed of one reclining seat and two leg orthoses, and each leg orthosis has three degrees of freedom, which correspond to the hip, knee, and ankle. Based on this robotic system, two controllers, i.e., passive training controller and active training controller, are proposed. The former takes advantage of the proportional-integral control method to solve the trajectory tracking problem, and the latter employs the surface electromyography signals to achieve active training. Two simplified impedance controllers, i.e., damping-type velocity controller and spring-type position controller, are designed for active training. A perceptron neural network detects movement intentions. The performance of the controllers was investigated with one able-bodied male. The results showed that the leg orthosis tracked the predefined trajectory based on the passive training controller, with the error rates of 0.45%, 0.44%, and 0.27%, respectively, for the hip, knee, and ankle. The active training controller whose loop rate is 6.67 Hz can move the leg orthosis smoothly, and the average recognition error of the perceptron neural network is less than 5%.

A lower limb rehabilitation robot, namely iLeg, has been developed recently. Since active exercises have been proven to be effective for neurorehabilitation and motor recovery, they are suggested to be implemented on iLeg. To this goal, patients' motion intention should be recognized. Therefore, a method based on the dynamic model of the human-robot interface (HRI) is designed to recognize the human motion intention. This paper is devoted to modeling and identifying the dynamics of the HRI. Firstly, the dynamic model of the HRI is designed by combining the dynamic models of the human leg and iLeg, where the human leg dynamic model (HLDM) is mainly concerned. By considering the motion trajectories during the rehabilitation exercises provided by iLeg, the human leg can be taken as a manipulator with two degrees of freedom; meanwhile, the joint angles and torques of the human leg can be measured indirectly by using the position and torque sensors mounted on the joints of iLeg. As a result, an 8-parameter HLDM can be designed by using the Lagrangian method. Then, the dynamic model of the HRI is identified by respectively and independently identifying the undetermined dynamic parameters of iLeg and the HLDM, where the dynamic parameters of the HLDM are mainly considered. Finally, the feasibility of the dynamic model of the HRI is validated by experiments.

This paper is a continuation of previous published work by the same authors on Personalized Modelling and Evolving Spiking Neural Network Reservoir architecture (PMeSNNr). The focus is on improvement of predictive modeling methods for the stroke occurrences case study utilizing an enhanced NeuCube architecture. The adaptability of the new architecture leads towards understanding feature correlations that affect the outcome of the study and extracts new knowledge from hidden patterns that reside within the associations. Through this new method, estimation of the earliest time point for stroke prediction is possible. This study also highlighted the improvement from designing a new experimental dataset compared to previous experiments. Comparative experiments were also carried out using conventional machine learning algorithms such as kNN, wkNN, SVM and MLP to prove that our approach can result in much better accuracy level.

The paper is a feasibility analysis of using the recently introduced by one of the authors spiking neural networks architecture NeuCube for modelling and recognition of complex EEG spatio-temporal data related to both physical and intentional (imagined) movements. The preliminary experiments reported in the paper suggest that NeuCube is much more efficient for the task than standard machine learning techniques, resulting in high recognition accuracy, a better adaptability to new data, a better interpretation of the models, leading to a better understanding of the brain data and the processes that generated it.

We present a new hybrid method for solving constrained numerical and engineering optimization problems in this paper. The proposed hybrid method takes advantage of the differential evolution (DE) ability to find global optimum in problems with complex design spaces while directly enforcing feasibility of constraints using a modified augmented Lagrangian multiplier method. The basic steps of the proposed method are comprised of an outer iteration, in which the Lagrangian multipliers and various penalty parameters are updated using a first-order update scheme, and an inner iteration, in which a nonlinear optimization of the modified augmented Lagrangian function with simple bound constraints is implemented by a modified differential evolution algorithm. Experimental results based on several well-known constrained numerical and engineering optimization problems demonstrate that the proposed method shows better performance in comparison to the state-of-the-art algorithms.