Amidst a rising global incidence of Alzheimer's Disease (AD) and an aging population, the search for predictive structural brain biomarkers and effective treatments is critical. Traditional multi-task learning (MTL) methods for predicting AD progression often focus on single cognitive scores across multiple time points. In contrast, our innovative approach, the Adaptive Multi-Cognitive Objective Temporal (AMCOT) mechanism, predicts AD progression by leveraging multiple cognitive assessments. This method not only integrates diverse cognitive data but also captures temporal correlations between different cognitive tasks, enhancing the model's ability to forecast disease progression. Unlike conventional MTL approaches that focus on isolated cognitive scores or single-time series predictions, AMCOT dynamically identifies temporal correlations across various cognitive-objective scores at different time points. We developed a novel MTL methodology using sparse group lasso techniques to pinpoint biomarkers linked to cognitive assessments accurately. A robust algorithm designed for large datasets significantly outperforms existing models in both overall and task-specific performance. Moreover, we applied stability selection to determine stable MRI biomarkers, analyzing their temporal patterns to deepen our understanding of AD progression. This comprehensive approach offers significant advancements in predictive accuracy and biomarker identification for AD. The implementation source can be found at https://github.com/XuanhanFan/AMCOT.

Parkinson's Disease (PD) is a prevalent and progressive neurodegenerative disorder. Wearable sensors enable automated analysis of human activities in daily life, providing a promising solution for continuous monitoring of motor symptoms. However, the scarcity of labeled data limits the effectiveness of supervised learning approaches in this context. To address this limitation, we propose a novel cycle-aware masked self-supervised learning framework that leverages unlabeled wearable sensor data for pre-training. Unlike conventional masking strategies that operate on individual time points or random continuous spans, our method introduces cycle-level masking that exploits the intrinsic periodicity of human activities, encouraging the model to learn structured temporal representations aligned with motor patterns. Extensive experiments on real-world wearable datasets demonstrate that our approach consistently outperforms state-of-the-art self-supervised learning baselines across multiple clinical and non-clinical tasks. In particular, our method achieves 3%-5% absolute accuracy improvement over the strongest competing masking-based methods across clinical diagnostic tasks, and delivers consistent gains across multiple daily activity recognition settings. These results validate the effectiveness of incorporating cycle-aware inductive bias into self-supervised learning for time-series clinical analysis, highlighting its potential for improving PD diagnosis in real-world scenarios.

Plant diseases and pests cause significant annual crop losses, severely threatening global food security. To mitigate crop losses, large language models (LLMs) demonstrate the ability to alleviate challenges in accessing precise farming support by acting as intelligent assistants that provide farmers with timely and precise decision support services. However, existing efforts still fall short in precision due to the neglect of time-sensitive nature in agricultural practices. They implicitly treat domain knowledge as time-irrelevant, ignoring the impact of crop phenology and pest life stages in their generated recommendations. To address this challenge, we propose a Time-Aware Retrieval-Augmented Generation framework (TARAG), comprising a time-aware knowledge base construction module, a hybrid retrieval module, and a time-based generation module to provide precise pest management suggestions. Firstly, the time-aware knowledge base construction module constructs a time-annotated knowledge base from unstructured documents to provide supplementary agricultural knowledge. Secondly, the hybrid retrieval module performs coarse-grained sparse retrieval to ensure relevance, with a time-sensitive re-ranking stage that refines the results to achieve both semantic relevance retrieval and time alignment with user queries. Finally, the time-aware generation module leverages the top-k retrieved documents and an instruction prompt to produce the final suggestion. To validate the effectiveness of the proposed framework, we contribute TAQA, the first bilingual, time-annotated agricultural question-answering dataset. Experiments demonstrate that TARAG significantly outperforms state-of-the-art RAG frameworks in retrieval precision and suggestion quality, with 99.14% retrieval recall and an F1 score of 66.85% for generated suggestions. The implementation and datasets for this work are available in https://github.com/lee-hash1/agri_rag.

Alzheimer’s disease (AD) is a neurodegenerative disease, well-known for its incurability, and is common among the elderly population worldwide. Previous studies have demonstrated that early intervention positively influences disease progression, leading to increased research into pathological analysis and disease trajectory prediction through machine learning (ML) methods. Given the similarities across different neurodegenerative disorders, a diagnosis relying solely upon a single modality of data is inadequate. Consequently, current research predominantly focuses on multimodal analysis, integrating medical imaging and clinical patient information, with continuous identification of new data types potentially aiding AD diagnosis. Multimodal approaches have been explored extensively over the past two decades, with significant advances observed following the introduction of Deep Learning (DL) techniques. Deep neural networks can adaptively extract and fuse features directly from input data, significantly broadening the scope of multimodal analysis. However, earlier classification studies have primarily concentrated on traditional ML, often neglecting the rapid advancements in DL networks. This article provides a comprehensive description of the acquisition pathways based on modalities, discusses the modalities currently used for research in neuroimaging, human body fluids, and other relevant sources. Additionally, it classifies fusion methodologies utilised in both DL and traditional ML contexts, highlights existing challenges, and outlines potential directions for future research.

An essential challenge of continual learning(CL) is that the knowledge gained from old tasks might be erased as new tasks are learned, which is called "catastrophic forgetting" (CF). Most CL methods assume tasks are drawn independently and identically distributed (i.i.d.); however, they ignore that as the number of tasks grows, some tasks inevitably share similarities, which could be leveraged for knowledge protection and transfer. This paper aims to enhance the adaptation of continual learning by leveraging task similarity. Specifically, we define task similarity as the change in empirical loss before training and show how it correlates with CF. As task similarity decreases, gradient norms increase, causing more severe forgetting. Building on this insight, we propose incor porating a Gradient Norm Regularisation(GNR) approach into the CL process. During training on a new task, we slow down the update magnitude for parameters deemed dissimilar. Meanwhile, gradient directions are adjusted to improve adaptability. Once training is complete, dissimilar parameters are consolidated to prevent them from being easily altered in future tasks. Experiments on four benchmark datasets show that our method outperforms other regularisation techniques and nearly matches the performance of multi-task learning. By plugging into existing regularisation and replay methods, GNR enhances the adaptation of continual learning. These results confirm that gradient norms can serve as an effective supervisory signal to balance knowledge protection and transfer. An extended experiment on a real-world pest classification dataset validates GNR's generalisability and the soundness of the task-similarity assumption in practical applications. We make the code of GNR publicly available at https://github.com/wang-xulong/GNR.git.

Large language models are increasingly deployed in agricultural decision-support settings, yet high-stakes crop protection in smallholder agriculture requires more than output-quality benchmarks. Over a two-year design and evaluation programme, we formalise policy-constrained large language model assessment as an adaptive compute allocation problem that jointly captures safety compliance, helpfulness, operational latency, and expert supervision workload. We introduce P-EVAL (Policy-grounded Expert-calibrated VALidation protocol), a unified evaluation framework for policy-grounded decision support, evaluating the architecture on a simulated field query database consisting of 1,240 cases. The protocol is instantiated on the Pezego advisory architecture (Pezego-HITL) and evaluated in Ghana. Following offline judge calibration against gold-standard human expert decisions (κ= 0.77), we evaluate the architectural performance under simulated query workloads. Under P-EVAL, our memory-routed architecture improves the Policy Alignment Rate (PAR) to 0.94 and the Agronomic Utility Rate (AUR) to 0.95, while reducing P95 latency by 55

Nitrogen rates for winter wheat are set before the season, under unknown prices and weather. The standard UK advice does not respond to prices, yet recent price swings moved the most profitable rate by tens of kilograms per hectare. Machine learning is often proposed as the fix. However, it is usually judged on prediction accuracy, and accurate prediction does not by itself make the recommended rate more profitable. Our insight is to score nitrogen advice directly by the profit it forgoes on measured yield response curves. We build a test bench on 892 such curves from two long running UK experiments, and sweep the nitrogen to grain price ratio to cover all price scenarios. On this bench, machine learning fails as a predictor. No model recovers the best rate within farm tolerance, and the benchmark noise shows none can. At normal prices, every model also loses to the standard advice on profit. The gain sits elsewhere. A simple correction step applied after the model cuts profit losses by a quarter, while better models and extra features give no gain. The same frozen correction cuts losses by 43

Detecting abnormal events in physiological signals such as EEG and ECG is critical for early diagnosis of neurological and cardiovascular disorders. However, existing unsupervised anomaly detection methods often suffer from limited representation capacity and weak generalization across diverse signal domains. To address these challenges, we propose MAGE, a novel unsupervised anomaly detection framework that integrates multi head memory gating, self-supervised learning, and adversarial training within a unified convolutional autoencoder architecture. The proposed memory-augmented gating mechanism selectively preserves and adaptively integrates salient features, improving discriminability and robustness over prior memory-based approaches. To further enhance representation learning, a self-supervised auxiliary task based on multiple signal transformations is introduced to encourage structure-aware feature extraction. In addition, transformation-aware adversarial perturbations are incorporated during training to enhance robustness against distribution shifts. Extensive experiments on multiple benchmark EEG and ECG datasets demonstrate that MAGE consistently outperforms state-of-the-art baselines in standard within-dataset evaluation settings, achieving over 98% detection accuracy and superior F1-scores. These results highlight the effectiveness and clinical potential of MAGE for early anomaly detection and continuous health monitoring under realistic deployment conditions. Code available at https://github.com/wzxmodel/MAGE

OBJECTIVE:Achieving effective and robust free-living PD severity assessment with wearable intelligence technologies requires a deep understanding of clinically relevant features, representative activities, and machine learning algorithms. METHODS:We designed a unified analytic framework (PDWearML) to optimise wearable ML approaches with simple daily activities for fast assessment of PD severity. It comprises annotation criteria, feature importance analysis, representative activity combination, and PD severity assessment. We conducted a 12-month study, developing a supervised PD wearable dataset containing 100 PD patients and 35 age-matched healthy controls using Huawei smartwatches and Shimmer. PD severity, assessed by trained physicians using the Hoehn and Yahr (H&Y) scale. RESULTS:The results reveal that through optimising multi-level feature extraction and combining three representative daily activities (WALK, ARISING-FROM-CHAIR, and DRINK), our smartwatch-based machine learning approach can assess PD severity in supervised settings within 2 minutes with an accuracy of up to 84.7%. SIGNIFICANCE:This work holds significant clinical value, offering a potential auxiliary tool for faster, more tailored interventions in PD healthcare.

The IEEE Journal of Translational Engineering in Health and Medicine (JTEHM) exists at the intersection of biomedical engineering and clinical practice. Published articles go beyond laboratory proof-of-concept to provide tangible, real-world evidence of translation into clinical settings. This editorial provides the rationale for manuscripts submitted to IEEE JTEHM to demonstrate evidence of clinical translation. It also provides examples of acceptable forms of evidence and offers guidance to authors on how to meet this expectation. Clinical and Impact—By requiring demonstrated clinical translational evidence IEEE JTEHM endeavours to publish high-quality research with scientific novelty and practical clinical impact. This expectation strengthens the journal’s aim to accelerate the adoption of innovative solutions into healthcare systems and ultimately deliver quantifiable benefits to patients.

Multi-predecessor following (MPF) platoons are a critical technology in autonomous transportation systems, enabling the alleviation of traffic congestion and enhancement of traffic efficiency. However, missing state information caused by communication delays, sensor faults, and packet losses in wireless networks poses a reliability challenge in MPF platoons for Connected and Automated Vehicles (CAVs), undermining future transportation safety. To solve this challenge, this study examines the effectiveness of data imputation strategies under controlled network conditions of CAVs and contributes to efficient solutions. Specifically, a configurable simulated network framework is established as an experimental test-bed for CAV platoons under unreliable wireless connections, where a CAV accesses incomplete velocity and spacing data from its predecessors under varying loss rates. In addition, a mask-aware CNN-LSTM model is proposed to impute missing values for guiding CAV platoons. The CNN-LSTM model is compared with statistical mean imputation and Transformer model in the simulation framework. Experimental results show that the CNN-LSTM consistently achieves the lowest imputation errors across missing rates from 10% to 90% in 10% increments, and significantly outperforms other methods in maintaining safe time-to-collision, velocity stability, and average speed. Remarkably, it even exceeds the fully observable baseline in certain scenarios. The proposed CNN-LSTM approach effectively mitigates information loss in MPF platoons, enhancing control resilience across diverse communication loss conditions and showing strong potential for real-world deployment. The source code and simulation data supporting this study are publicly available at: https: //github.com/LukeZhu0227/CNN-LSTM_for_vehicle_information_reconstruction.

Digital mobility outcomes (DMOs) derived from wearable sensors characterise mobility in daily life and offer a promising means of monitoring disease progression. Yet most DMO studies examine one disease at one visit; they do not model how multivariate DMO relationships with multiple clinical outcomes evolve jointly across diseases. Technically, existing temporal multi-task frameworks can model progression within an individual disease, but they do not jointly model multiple prediction outcomes across diseases, particularly when disease cohorts do not share participants. To address these gaps, we propose DeMMO, an interpretable framework for longitudinal, multi-disease, and multi-outcome learning. DeMMO represents each disease-outcome objective by a longitudinal DMO coefficient matrix and combines temporal regularisation with stable and visit-specific feature selection. Its central technical contribution is an automatic cross-disease and cross-outcome relation-learning mechanism that learns signed relations directly from these longitudinal mappings, enabling selective information sharing without paired participants. We evaluate DeMMO on the recently released, large-scale, multicentre Mobilise-D dataset, which provides a new opportunity to study 24 harmonised real-world DMOs over five visits across multiple mobility-limiting conditions. Against nine strong linear, longitudinal, and deep-regression baselines, DeMMO achieves the best overall and outcome-specific prediction performance, with significant improvements over the strongest baselines. Stability selection further identifies reliable longitudinal DMO patterns for subsequent clinical validation and disease monitoring. The implementation code and experimental results are available at https://github.com/menghui-zhou/DeMMO.

Pests significantly threaten global agricultural production, which causes severe yield losses through feeding and virus transmission. To mitigate yield losses caused by pests, timely and precise pest management practices are critical. Although previous efforts have advanced automated solutions for real-time environmental monitoring in agriculture, implementing precise pest management decision-making and suggestion generation remains a challenge due to complex reasoning processes in practice. In response, an enhanced pest management system, PEZEGO, is proposed to provide precise management suggestions through multimodal environmental data, a fine-tuned open vocabulary detector (OVD), and large language models (LLMs). Specifically, a mobile application and low-cost Internet of Things (IoT) devices are developed to capture images and environmental information. A hybrid convolutional low-rank adaptation method (HCLoRA) is proposed to fine-tune pretrained OVDs, enabling zero-shot pest detection for converting images to pest species and quantity information. In addition, a structured data-based retrieval augmented generation (SRAG) workflow for LLMs is proposed to provide precise pest management suggestions through automatically extracted agriculture management knowledge and Chain-of-Thought. The effectiveness of PEZEGO is validated in a case study of pest management in the U.K., including pest detection in field scenarios and management suggestion generation. Compared to advanced model fine-tuning methods, HCLoRA for YOLOWorld achieves the highest detection performance with 0.1759 AP(h) on pest detection. Additionally, the proposed SRAG workflow demonstrates the ability to support pest management with a 68.7% average F1 score for knowledge extraction and 77.33% accuracy for suggestion generation. Eventually, a mobile application demonstrates the practical effectiveness of the proposed system.

The effectiveness of complete multi-modal neuroimaging data in the diagnosis of Alzheimer’s disease has been extensively demonstrated and applied. Dealing with incomplete modalities poses a common challenge in multi-modal neuroimaging diagnosis. The mainstream approaches aim to synthesize missing neuroimaging data in order to make full use of all available samples. However, these methods treat image synthesis and disease diagnosis as two independent tasks, overlooking the potential feature of cross-modality image synthesis for downstream tasks. To this end, we propose the Joint Image Synthesis and Classification Learning method to jointly optimize image synthesis and disease diagnosis using incomplete neuroimaging modalities. Our approach comprises a submodule for synthesizing missing neuroimaging data and a decision fusion submodule that integrates features from different modalities and the high-level/converted features generated during synthesis. Experimental results demonstrate that our joint optimization approach outperforms conventional two-stage methods. Our method is capable of handling arbitrary neuroimaging modality missing scenarios and achieves state-of-the-art performance in both Alzheimer’s Disease identification and mild cognitive impairment conversion classification tasks. Finally, we further explored the importance of different converted features. This highlights the effectiveness of our approach in addressing the challenges of Alzheimer’s Disease diagnosis and provides insights for future research in multi-modal medical image analysis.

Alzheimer's disease (AD) is the most prevalent form of dementia, and its early diagnosis is essential for slowing disease progression. Recent studies on multimodal neuroimaging fusion using MRI and PET have achieved promising results by integrating multi-scale complementary features. However, most existing approaches primarily emphasize cross-modal complementarity while overlooking the diagnostic importance of modality-specific features. In addition, the inherent distributional differences between modalities often lead to biased and noisy representations, degrading classification performance. To address these challenges, we propose a Collaborative Attention and Consistent-Guided Fusion framework for MRI and PET based AD diagnosis. The proposed model introduces a learnable parameter representation (LPR) block to compensate for missing modality information, followed by a shared encoder and modality-independent encoders to preserve both shared and specific representations. Furthermore, a consistency-guided mechanism is employed to explicitly align the latent distributions across modalities. Experimental results on the ADNI dataset demonstrate that our method achieves superior diagnostic performance compared with existing fusion strategies.

Parkinson’s disease (PD) is a progressive neurodegenerative disorder that significantly diminishes patients’ quality of life. Early and accurate diagnosis is critical for reducing both individual and societal burdens. Although current diagnostic methods can effectively differentiate between PD patients and healthy individuals, they tend to ignore the diversity of PD symptoms and the differences with other similar diseases, such as essential tremor or multiple system atrophy, leading to a higher risk of misdiagnosis. Additionally, existing supervised learning methods rely on subjective labeling by physicians, which is both time-consuming and subjective. To overcome these limitations, we collect multi-sensor activity data from 102 participants in free-living environments, and propose a novel self-supervised learning framework that redefines PD diagnosis as an anomaly detection problem. Additionally, we utilize two large public PD datasets as external cohorts to verify their validity. Extensive experiments demonstrate that our framework not only learns more discriminative features but also significantly enhances the model’s generalization ability, providing a promising solution to reduce misdiagnosis in PD.

Objective: Recent quantification research on Parkinson's disease (PD) integrates wearable technology with machine learning methods, indicating a strong potential for practical applications. However, the effectiveness of these techniques is influenced by environmental settings and is hardly applied in real-world situations. This paper aims to propose an effective feature assessment framework to automatically rate the severity of PD motor symptoms from short-term motor tasks, and then classify different PD severity levels in the real world. Methods: This paper identified specific PD motor symptoms using a novel feature-assessment framework at both segment-level and sample-level. Features were selected after calculating SHapley Additive exPlanation(SHAP) value, and verified by different machine learning methods with appropriate parameters. This framework has been verified on real-world data from 100 PD patients performing Unified Parkinson's Disease Rating Scale(UPDRS)-recommended short motor tasks, each task lasting 20-50 seconds. Results: The sensitivity for recognizing motor fluctuations reached 88% in tremor recognition. Additionally, LightGBM achieved the highest accuracy for early detection(92.59%) and achieved 71.58% in fine-grained severity classification using 31 selected features. Conclusion: This paper reports the first effort to assess multi-level and multi-scale features for automatic quantification of motor symptoms and PD severity levels. The proposed framework has been proven effective in assessing key PD information for recognition during short-term tasks. Significance: The explanatory analysis of digital features in this study provides more prior knowledge for PD self-assessment in a free-living environment.

With the rapid proliferation of edge devices, such as those in the Internet of Things (IoT), which generate critical data for machine learning applications, it is essential to enable their participation in privacy-preserving Federated Learning (FL) systems. Given their limited computational resources, an effective approach is to adapt and reduce their training workload to align with their capabilities. Previous FL research has focused primarily on workload reduction through lightweight models at the edge, with limited attention given to optimizing on-device training efficiency by reducing the amount of data required for training. In this work, we propose FedFT-EDS, a novel approach that combines Fine-Tuning of partial client models with Entropy-based Data Selection to reduce training workloads on edge devices. By actively selecting the most informative local instances for learning, FedFT-EDS significantly reduces the training data in FL and demonstrates that not all user data are equally beneficial across training rounds. We show that FedFT-EDS uses only 50% of the available training data while improving the global model performance compared to the baseline methods, FedAvg and FedProx. Importantly, FedFT-EDS improves the learning efficiency of client models by up to 3x, to achieve a similar performance to the baselines in only one third of their training time. This work underscores the critical role of data selection in Federated Learning and offers a promising direction for achieving scalable and efficient FL systems.

Precision fertilisation is essential for fertiliser conservation, soil nutrient balancing, emission reduction and yield improvement. Precision fertilisation prediction refers to the simultaneous prediction of timing and rate for fertilisation. Current research focuses on the precision prediction for a single type of fertiliser, or the application rate prediction of multiple fertilisers. However, interactions of multiple fertilisers can affect plant growth, soil structure and environmental balance. The lack of corresponding agricultural data and the resulting barriers to designing temporal and spatial fertilisation prediction algorithms are significant obstacles to utilising machine learning approaches for precision application prediction of multiple fertilisers. In this research, we investigated and collected real-world agricultural dataset from ten genuine winter wheat farms in England, which encompassed a variety of agricultural variables including climate, soil, cropping data and multiple fertilisations. Based on this dataset, we proposed a novel spatio-temporal tensor multi-task learning approach to deal with the spatio-temporal characteristics and scarcity of agricultural data. The approach utilises the multidimensional tensor created from the raw data to efficiently predict the timing and rate of base and topdressing fertilisation of multiple fertilisers (nitrogen, phosphorus, potassium and sulphur). As there are complex interactions among different fertilisers, we constructed the continuous incremental learning framework based on this algorithm to enhance the prediction performance of precision application for multiple fertilisers. Extensive experiments were conducted to evaluate the performance of our proposed approach compared to standard regression models with real-world agricultural dataset. The results demonstrated that our proposed approach provides superior accuracy and stability in the precision application prediction of multiple fertilisers.

The increase in privacy concerns and the introduction of privacy and data protection legislation compel organisations to reevaluate their practices regarding traditional machine learning. The aggregation and management of users’ private data on the central server may contravene regulations if not properly administered. Federated learning provides a technique that eliminates the necessity of uploading users’ data to the server. It facilitates substantial learning by collaboratively training on each client’s devices and pooling the model gradient changes. Federated learning, augmented with a proxy as an intermediary and encrypted model parameters, will enhance anonymity, privacy, and data protection against malicious threats, including membership inference adversaries. Nonetheless, encrypted data incurs costs for customers’ communication and data size that exceed twice the original size. Our paper seeks to resolve these issues. We present two secure approaches for effective communication in an anonymous encrypted federated learning framework as our contribution. Additionally, our experiments demonstrated that it is feasible to attain equivalent communication costs as in non-encrypted scenarios. We provide recommendations in the conclusion for the effective implementation of privacy-preserving federated learning in the area of personal devices.