BACKGROUND:Low back pain (LBP) is a leading cause of global disability. Most cases are non-specific (NSLBP) and lack identifiable causes. Early active management is endorsed by clinical guidelines; however, exercises are rarely customised despite substantial variability in impairments. Existing classification systems can support targeted rehabilitation but require extensive clinical training and lengthy assessment procedures, limiting timely personalised care. OBJECTIVE:This study used AI methods to identify the two most common motor control impairments (MCIs)-flexion and extension patterns (FP/EP)-in NSLBP. The approach used spinal silhouettes extracted from movement videos to enable self-phenotyping and guide personalised exercise selection. METHODS:Data were collected from a research fellowship involving ninety NSLBP participants classified by an expert physiotherapist (LS) into FP or EP MCIs. Participants performed standard forward- and backward-bend tasks recorded in the sagittal plane. Pose estimation and instance segmentation techniques were applied to extract motion features and spine silhouettes. From each participant, a curated set of 80 black-and-white back images captured at specific bending angles was produced. These features were used to train a feedforward neural network. Model performance was assessed using five-fold cross-validation with accuracy, sensitivity, specificity, F1 score and AUC. RESULTS:The model achieved a diagnostic accuracy of 91.91% (95% CI 84.8-99.1) for backward-bend videos, exceeding reported inter-examiner agreement rates for trained physiotherapists. Robustness was supported by a mean AUC of 0.9422. Accuracy was lower for forward-bend images (86.69%), combined tasks (86.29%), or PROMs alone (63.82%). Adding PROMs to forward- or backward-bend tasks provided only modest improvements (66.32% and 71.62%, respectively). CONCLUSION:The model reliably distinguished between FP and EP NSLBP subgroups, demonstrating the potential of AI to support timely personalised rehabilitation. The integration of PROMs with motion features reduced classification accuracy, suggesting that self-reported outcomes may provide limited benefit when tailoring exercises to specific physical impairments.

Accurate classification of Movement Impairment (MI) and Motor Control Impairment (MCI) in non-specific low back pain (NSLBP) is essential for targeted rehabilitation but remains challenging due to subjective assessments and subtle movement differences. We present MOVEXor, a lightweight and explainable multi-modal framework that integrates spinal curvature images and motion-derived features through a modality-aware attention gating mechanism. MOVEXor achieves high classification performance (up to 97.5% accuracy) while offering transparent decision-making via Grad-CAM and Integrated Gradients (IG). Our analysis shows that the model focuses on physiologically meaningful movement phases, particularly minimal flexion angle, and relies heavily on motion stability for classification. The fused attention-based design outperforms static fusion methods, especially when handling noisy inputs. With minimal hardware requirements and real-time explainability, MOVEXor holds strong potential as a clinical decision-support tool for both in-clinic and remote settings, enabling objective, interpretable, and personalised rehabilitation exercise of LBP subgroups.

Accurate fruit instance segmentation at different ripeness stages is critical for developing autonomous harvesting robots, particularly given the unstructured in-field conditions. In this paper, we combine two in-field fruit datasets of peaches and strawberries for multiple ripeness stages determination, and propose a lightweight query-based instance segmentation model named FruitQuery.The combined dataset contains 3 peach ripeness stages and 4 strawberry ripeness stages, covering various unstructured conditions of two popular fruits. The model FruitQuery consists of three parts: a backbone, a pixel decoder and Transformer decoders. Efficient multi-head self-attention modules are introduced to the backbone to reduce computational overhead, and a pyramid pooling module is added to the pixel decoder to enhance multi-scale feature fusion. Transformer decoders are then applied to learn a fixed number of queries from features and generate instance masks, avoiding postprocessing like non-maximum suppression. FruitQuery runs in an end-to-end way and incorporates the convolution and Transformer to capture fine-grained features related to different fruits at different ripeness stages.Extensive experiments on the combined fruit dataset demonstrate that our FruitQuery achieves the highest average precision of 67.02 with only 14.08M parameters, outperforming 13 state-of-the-art models with 33 variants. It is noted that FruitQuery surpasses three series of YOLO (v8, v9 and v10) by a large margin. Ablation studies and visualizations also show its robust feature extraction with fewer parameter usage, indicating that the query-based design is effective in localizing fruit. These results highlight FruitQuery's compelling balance between segmentation performance and model size, offering the potential for in-field application.

The growing need for standardised and automated cardiac ultrasound (US) acquisition has driven the integration of deep learning into echocardiographic workflows. While existing deep learning (DL) models have shown promising results in tasks such as view classification and image quality assessment, most of these approaches focus either on differentiating among standard views or grading image quality within a standard view. However, these methods lack the capacity to model the sequential spatial transitions that occur during the acquisition process, limiting their applicability to real-time probe guidance and robotic control. To address this gap, we propose a classification framework designed for the process of acquiring the parasternal long-axis (PLAX) view under a fixed scanning protocol. Based on extensive probe movement experiments across multiple patients, we identified four representative echocardiographic views that appear during the search for the optimal PLAX position. These views correspond to distinct probe positions and orientations and reflect varying levels of image completeness. A dataset of 7,200 annotated images was used to train a ResNet50-based deep network for multi-class classification. The model achieved robust performance with accuracy, sensitivity, specificity, and F1 scores above 89%, and AUC exceeding 97% in patient-level cross-validation. It effectively captures spatially relevant features, distinguishes subtle view differences, and generalizes well to unseen data. The outputs provide interpretable feedback correlating image quality with probe position, enabling real-time scanning assessment. Furthermore, this work introduces a novel problem formulation and multi-class view classification under a fixed acquisition protocol. It provides a foundation for developing the next generation of intelligent US systems. By linking image classification to probe position and orientation, the proposed framework enables real-time feedback that can ultimately support autonomous scanning agents in locating diagnostically optimal cardiac views.

The full view of the apples in the orchard is often obscured by leaves and trunks, making it challenging accurately determine their ripeness, whilst it is an essential yet difficult task for apple-harvesting robots. Within this context, we propose a novel method to address two critical challenges: ripeness determination and in-field occlusion. The proposed method is trained in a self-supervised manner on a dataset consisting of less than 1% labelled images and the rest of unlabelled images. It is made up of three key parts: a reconstructor, a feature extractor, and a predictor. The reconstructor is designed to reconstruct the missing parts of occluded apples. The feature extractor is introduced to learn ripeness-related features from the vast number of unlabelled images. Unlike the previous approaches classifying the fruit ripeness into several discrete categories, the predictor uses the learned features to generate a continuous ripeness score in the range between 0.0 and 1.0, thus eliminating the need to subjectively pre-define ripeness stages and offering end-users the flexibility to make their own decisions. Experimental results comparing our method to another method with different settings show that our method achieves the best Structural Similarity Index Measure (SSIM) of 0.75 and the second-best Peak-Signal-to-Noise Ratio (PSNR) of 25.36 for reconstructing missing apple parts, whilst using the fewest 86.3M parameters. Besides, our method outperforms 15 other self-supervised methods and even a supervised method in the ripeness score prediction, with the smallest score 0.0127 for fully unripe and the highest score 0.8933 for fully ripe apples. The results demonstrate the potential of our method to be incorporated with in-field robotic systems, enabling them to assess ripeness for selective harvesting effectively. It is helpful to monitor the overall ripeness of large orchards digitally, aid the decision-making processes and advance the goals of smart and precision agriculture.

The integration of artificial intelligence (AI) into healthcare holds significant promise for enhancing clinical decision-making and improving patient outcomes. However, general-purpose large language models (LLMs) frequently exhibit limitations such as hallucinations, lack of domain-specific accuracy, and opaque reasoning processes, posing risks in clinical applications. This study addressed these challenges by proposing and exploring an innovative integration of Chain-of-Thought (CoT) prompting, Retrieval-Augmented Generation (RAG), and parameter-efficient fine-tuning using Quantized Low-Rank Adaptation (QLoRA) specifically tailored for medical use.A distilled 14-billion-parameter variant of the DeepSeek R1 model was fine-tuned using a structured clinical dataset that emphasizes step-by-step reasoning. Additionally, external medical references were incorporated through RAG, employing embedding models for precise context retrieval. The combination of these techniques was systematically evaluated on a challenging set of open-ended medical questions, resulting in accuracy improvements—from a baseline accuracy of 55% to a final performance of 81%.Further qualitative evaluation involving three practising General Practitioners (GPs) and three fourth-year medical students from Cardiff University underscored the proposed system’s clinical utility and transparent reasoning capabilities while also identifying areas for improvement, such as conciseness and explicit adherence to national-specific clinical guidelines.This research demonstrated that integrating CoT, RAG, and QLoRA provides a practical pathway toward reliable, transparent, and clinically relevant AI support for healthcare professionals. Recommendations for future work include scaling models, incorporating comprehensive patient data, and enhancing customization for clinical application contexts.

Recognizing human intentions is a key challenge in human-robot interaction research. Much of the current work in this area centers on identifying human intentions within specific activities, often relying on a limited set of features. In contrast, this paper introduces a more versatile framework for intention recognition and introduces a novel model: the Spatial-Temporal Graph Attention Informer Neural Network (STGAIN). To recognize intentions, this model leverages spatial relationships between humans and objects in different scenes, along with their temporal evolution. In addition, to address an existing research gap this research developed a new dataset called Dynamic Scene Graph (DSG) with representative dynamic relationships, derived from 471 videos covering 20 categories of human intentions. This dataset represents people and objects in different scenes, and the relationships between them. The model was tested rigorously at different points in the videos to track how the scenes evolved and to assess prediction accuracy, comparing the results to a range of advanced algorithms. Our findings clearly demonstrate that STGAIN outperforms these models, showcasing its potential for advanced human intention recognition applications. This model represents a significant advance toward creating more human-centered robots, capable of understanding and adapting to human intentions in real-world situations.

Background Low back pain (LBP) is a major global disability contributor with profound health and socio-economic implications. The predominant form is non-specific LBP (NSLBP), lacking treatable pathology. Active physical interventions tailored to individual needs and capabilities are crucial for its management. However, the intricate nature of NSLBP and complexity of clinical classification systems necessitating extensive clinical training, hinder customised treatment access. Recent advancements in machine learning and computer vision demonstrate promise in characterising NSLBP altered movement patters through wearable sensors and optical motion capture. This study aimed to develop and evaluate a machine learning model (i.e., ’BACK-to-MOVE’) for NSLBP classification trained with expert clinical classification, spinal motion data from a standard video alongside patient-reported outcome measures (PROMs). Methods Synchronised video and three-dimensional (3D) motion data was collected during forward spinal flexion from 83 NSLBP patients. Two physiotherapists independently classified them as motor control impairment (MCI) or movement impairment (MI), with conflicts resolved by a third expert. The Convolutional Neural Networks (CNNs) architecture, HigherHRNet, was chosen for effective pose estimation from video data. The model was validated against 3D motion data (subset of 62) and trained on the freely available MS-COCO dataset for feature extraction. The Back-to-Move classifier underwent fine-tuning through feed-forward neural networks using labelled examples from the training dataset. Evaluation utilised 5-fold cross-validation to assess accuracy, specificity, sensitivity, and F1 measure. Results Pose estimation’s Mean Square Error of 0.35 degrees against 3D motion data demonstrated strong criterion validity. Back-to-Move proficiently differentiated MI and MCI classes, yielding 93.98% accuracy, 96.49% sensitivity (MI detection), 88.46% specificity (MCI detection), and an F1 measure of .957. Incorporating PROMs curtailed classifier performance (accuracy: 68.67%, sensitivity: 91.23%, specificity: 18.52%, F1: .800). Conclusion This study is the first to demonstrate automated clinical classification of NSLBP using computer vision and machine learning with standard video data, achieving accuracy comparable to expert consensus. Automated classification of NSLBP based on altered movement patters video-recorded during routine clinical examination could expedite personalised NSLBP rehabilitation management, circumventing existing healthcare constraints. This advancement holds significant promise for patients and healthcare services alike.

Over the past decade, automating ultrasound scanning has been the subject of intense research. However, training a robot to perform automated ultrasound examinations requires a substantial corpus of training data. Traditionally, researchers have sought to obtain such data either through publicly available datasets or by engaging professional sonographers in experiments aimed at dataset generation. The former approach often yields incomplete datasets insufficient for specialized research objectives, while the latter entails logistical challenges, including the necessity for frequent manual experimentation and ready access to medical professionals. Therefore, the acquisition of a comprehensive and suitable dataset remains an essential yet formidable challenge. Here, we propose a novel framework for achieving the automated acquisition of cardiac ultrasound datasets by controlling robot behaviour using a digital twin within a simulated environment. This framework consists of two modules: physical and virtual. Within the virtual simulation module, diverse body models of varying dimensions can be inputted, enabling the planning of robot arm path and ultrasound scanning manoeuvrers. Then the physical robot arm clones the actions of the robot in the simulation environment and updates its current state in the virtual module. The proposed framework was used to collect 43,000 cardiac ultrasound images from 8 patients with different pathologies and 1 healthy individual using a KUKA LBR Med robot and Intelligent Ultrasound Simulator. It is also expected to be feasible for a real-person dataset collection.

Human activity recognition, as a significant branch of artificial intelligence, requires increasingly generalized and precise methodologies due to growing demands. Therefore, this paper proposes a context-based method for recognising indoor human activities, which interlinks indoor human activities with interactions with objects, making the contextual relationship between humans and objects particularly crucial. In addition, this research has developed a new dynamic graph dataset based on publicly available video datasets and their associated descriptive scripts, instantiating the relationships between humans and objects. A novel architecture for human activity recognition is developed in this research. This architecture utilizes graph neural networks and self-attention mechanisms to learn the significance of the interactions between humans and objects and capture the relationships between video frames on a temporal level. The results demonstrate that the classification accuracy reaches 0.86 and it also performs better than other current advanced algorithms STGAT and STGCN. It is noteworthy that the approach also effectively reduces ambiguity in activity recognition.

Accurately segmenting strawberries within real-world production settings not only helps the development of automated harvesting robots but also allows for precise calculation of the number and size of strawberries, providing accurate yield information for agricultural planning and resource optimization. This paper proposes lightweight and efficient CNN models specifically designed for strawberry instance segmentation, consisting of an efficient self-attention-based backbone, a feature pyramid network (FPN), and a decoder with an instance branch and a mask branch. The proposed models surpass the original and simplified Mask R-CNN with significant 21.46 and 22.97 AP gains respectively, with the Base backbone achieving the highest AP of 70.22. Additionally, our models demonstrate efficiency by requiring much fewer parameters (17.42M) and floating-point operations (78.3G) compared to Mask R-CNN (35.08M / 877.4G), making them suitable for deployment on devices with limited computational resources.

Low Back Pain (LBP) is a prevalent musculoskeletal disorder affecting over 80% of the population over their lifetime and is a leading cause of disability globally. The most frequent type, non-specific LBP (NSLBP) does not have a clearly identifiable pathology cause. Current clinical guidelines advocate for tailored management and self-care approaches for NSLBP. The effectiveness of these personalised management plans significantly depends on accurate and on-going assessment of the patient’s spinal function. This presents considerable challenges for both clinicians and patients.This study introduces “SpineSighter”, an artificial intelligence (AI) model developed to tailor management of NSLBP by categorising patients based on their spinal function either into High Function (HF) and Low Function (LF) subsets. Utilising standard video recordings and computer vision technology, SpineSighter analyses motion features such as angular displacement, velocity, and acceleration during repeated forward flexion tests. The model showed high accuracy in classifying spinal function, achieving an accuracy of 95.13%, sensitivity of 93.81%, specificity of 96.00%, and an F1 score of 0.9442. This innovative use of AI highlights the importance of velocity as a critical indicator of spinal functional differences, opening new avenues for personalised clinical management, self-care and recovery strategies of NSLBP.

Peach instance segmentation is a crucial part to locate peaches and classify their ripeness stages to build an automatic peach harvesting or monitoring machine. This paper proposes a large and high-quality peach dataset called NinePeach, and a new one-stage instance segmentation model. The NinePeach dataset aims to reproduce real-world field conditions, encompassing various factors that can significantly influence the accuracy of peach detection, such as varying natural light intensity, instances of multiple fruit adhesion, and occlusion caused by stems and leaves. This is the largest and the most varied peach dataset among publicly available peach datasets to our best knowledge. Our proposed one-stage segmentation model does not require Region Proposal Network (RPN) to generate bounding box proposals, it directly identifies object instances by their centre locations and sizes and predict their category at the same time. The proposed model incorporates channel attention and spatial attention mechanisms to enhance object detection capabilities in crucial channels and spatial locations. Experimental results show that the state-of-the-art Mask RCNN performs 69.91% average precision (AP) with Swin-T backbone, our model surpasses it with the same backbone, achieving the highest 72.12% AP, and delivering more precise mask and boundary predictions. Specifically, our model is capable of accurately detect peaches under various conditions, such as peaches partially obscured by leaves, peaches partially exposed or overlapped. These advancements present promising prospects for the application of this technology to other fruits or crops.

In human-robot interaction, understanding human intention is important to smooth interaction between humans and robots. Proactive human-robot interactions are the trend. They rely on recognising human intentions to complete tasks. The reasoning is accomplished based on the current human state, environment and context, and human intention recognition and prediction. Many factors may affect human intention, including clues which are difficult to recognise directly from the action but may be perceived from the change in the environment or context. The changes that affect human intention are the triggers and serve as strong evidence for identifying human intention. Therefore, detecting such changes and identifying such triggers are the promising approach to assist in human intention recognition. This paper discusses the current state of art in human intention recognition in human-computer interaction and illustrates the importance of context change and triggers for human intention recognition in a variety of examples.

Traditionally, explainable artificial intelligence seeks to provide explanation and interpretability of high-performing black-box models such as deep neural networks. Interpretation of such models remains difficult, because of their high complexity. An alternative method is to instead force a deep-neural network to use human-intelligible features as the basis for its decisions. We tested this approach using the natural category domain of rock types. We compared the performance of a black-box implementation of transfer-learning using Resnet50 to that of a network first trained to predict expert-identified features and then forced to use these features to categorise rock images. The performance of this feature-constrained network was virtually identical to that of the unconstrained network. Further, a partially constrained network forced to condense down to a small number of features that was not trained with expert features did not result in these abstracted features being intelligible; nevertheless, an affine transformation of these features could be found that aligned well with expert-intelligible features. These findings show that making an AI intrinsically intelligible need not be at the cost of performance.

User experience (UX) analysis is essential for designers and companies when optimizing products or services as it can help designers to uncover valuable information, such as the hedonic and pragmatic qualities of a UX. While previous research has described the conventional methods of UX analysis, such as surveys or subjective determination, this paper proposes a data-driven methodology to automatically integrate hedonic and pragmatic qualities for UX from online customer reviews. The proposed methodology comprises the following steps. First, we combined a corpus-based approach, a dictionary-based approach and word embedding to generate a lexicon of hedonic and pragmatic qualities. Second, we filtered out the sentences that contained no hedonic or pragmatic information and classified the remaining review sentences. Third, we extracted and clustered the UX elements (such as product feature, context information and context clustering). Finally, we scored each UX element based on hedonic or pragmatic qualities and compared it against previous UX modeling. This study integrates hedonic and pragmatic qualities to enrich UX modeling in the field of UX. For a product designer, the UX analysis results may highlight a requirement to optimize product design. It may also represent a potential market opportunity in a UX state where most of the current products are perceived UX results by customers. This research also examines the invaluable relationship between UX and online customer reviews to support the prospective planning of customer strategy and design activities.

For user experience (UX) analysis in product design, the context of use (such as task, activities and environment) is a valuable element that enables context-awareness in accordance with the available context of use elements of users. Typically, conventional methods such as interviews or questionnaires are used to extract the context of use elements, but they are labour-intensive and time-consuming and thus do not scale well. On the other hand, the automatic extraction approaches from existing studies are not as effective as the extracted context of use phrases are not as informative. In this study, we present an automatic approach to exploit and understand the context of use elements from online customer reviews using BERT. Firstly, the context of use elements from online customer reviews is labelled using a BERT-based approach. Secondly, a syntactic-based post-processing is designed to check the labelled results and form the phrases related to the context of use. Finally, the customers’ preferences related to the contexts of use is analysed by studying and aggregating it with its relevant hedonic quality (such as positive or negative) which can then be used to enrich the UX modelling. For product designers, the modelling results can facilitate the optimisation of product design. A case study was conducted to understand and leverage the context of use elements in UX from online customer reviews to support customer strategy creation and design activities.

Research on user experience (UX) has attracted much attention from designers. Additionally, hedonic quality can help designers understand user interaction (such as attractive, original and innovative) when they experience a product. Realising the user’s interaction state is a significant step for designers to optimise product design and service. Previous UX modelling lacks exploration in user interaction state. Also, the lack of user interaction state factor will reduce the accuracy of the UX modelling. In this paper, we explore the interaction value of online customer review and introduce a new approach to integrating hedonic quality for UX modelling. Firstly, extracting word list from online customer review; Secondly, hedonic quality words are extracted from the word list and added as a hedonic quality part to UX modelling; Thirdly, we compared the analysis result with our previous study for the conclusion. This research combines hedonic quality with UX modelling to enrich modelling in the field of UX for the first time. The proposed data collection method is superior to the traditional collection methods in hedonic quality studies. Extracting hedonic quality factors from online customer reviews can in-depth provide reflections for designers to improve their product design. Furthermore, it also explored the valuable relationship between UX and online customer reviews to provide proactive thinking in user strategy and design activities.

Doppler ultrasound technology is widespread in clinical applications and is principally used for blood flow measurements in the heart, arteries, and veins. A commonly extracted parameter is the maximum velocity envelope. However, current methods of extracting it cannot produce stable envelopes in high noise conditions. This can limit clinical and research applications using the technology. In this article, a new method of automatic envelope estimation is presented. The method can handle challenging signals with high levels of noise and variable envelope shapes. Envelopes are extracted from a Doppler spectrogram image generated directly from the Doppler audio signal, making it less device-dependent than existing image-processing methods. The method's performance is assessed using simulated pulsatile flow, a flow phantom, and in vivo ascending aortic flow measurements and is compared with three state-of-the-art methods. The proposed method is the most accurate in noisy conditions, achieving, on average, for phantom data with signal-to-noise ratios (SNRs) below 10 dB, bias and standard deviation of 0.7% and 3.3% lower than the next-best performing method. In addition, a new method for beat segmentation is proposed. When combined, the two proposed methods exhibited the best performance using in vivo data, producing the least number of incorrectly segmented beats and 8.2% more correctly segmented beats than the next best performing method. The ability of the proposed methods to reliably extract timing indices for cardiac cycles across a range of signal quality is of particular significance for research and monitoring applications.

The Clinch Token Transfer Test (C3t) is a bi-manual coin transfer task that incorporates cognitive tasks to add complexity. This study explored the concurrent and convergent validity of the C3t as a simple, objective assessment of impairment that is reflective of disease severity in Huntington’s, that is not reliant on clinical expertise for administration. One-hundred-and-five participants presenting with pre-manifest (n = 16) or manifest (TFC-Stage-1 n = 39; TFC-Stage-2 n = 43; TFC-Stage-3 n = 7) Huntington’s disease completed the Unified Huntington’s Disease Rating Scale and the C3t at baseline. Of these, thirty-three were followed up after 12 months. Regression was used to estimate baseline individual and composite clinical scores (including cognitive, motor, and functional ability) using baseline C3t scores. Correlations between C3t and clinical scores were assessed using Spearman’s R and visually inspected in relation to disease severity using scatterplots. Effect size over 12 months provided an indication of longitudinal behaviour of the C3t in relation to clinical measures. Baseline C3t scores predicted baseline clinical scores to within 9–13% accuracy, being associated with individual and composite clinical scores. Changes in C3t scores over 12 months were small ( $$\Omega$$ ≤ 0.15) and mirrored the change in clinical scores. The C3t demonstrates promise as a simple, easy to administer, objective outcome measure capable of predicting impairment that is reflective of Huntington’s disease severity and offers a viable solution to support remote clinical monitoring. It may also offer utility as a screening tool for recruitment to clinical trials given preliminary indications of association with the prognostic index normed for Huntington’s disease.