Haitao Song

Exercise is a foundational intervention for preventing and managing chronic diseases, yet substantial inter-individual variation in physiological and molecular responses limits translation into personalized clinical practice. Meanwhile, wearables, multi-omics profiling, imaging phenotypes, and large-scale cohorts are generating multi-scale data that characterize exercise responses from molecular programs to real-world behavior. Artificial intelligence (AI) offers computational tools to integrate these modalities, learn predictive representations, and derive clinically relevant phenotypes. In this review, we summarize recent advances in AI-enabled translational exercise biomedicine. We first describe the evolving data ecosystem, including wearable-derived digital phenotypes, imaging biomarkers, and exercise-responsive multi-omics signatures. We then discuss key modeling paradigms – time-series learning, multimodal fusion, causal inference, and reinforcement learning – with emphasis on their roles in response prediction and personalized intervention design. We highlight emerging applications across cardiometabolic disease, metabolic liver disease, neurodegeneration, cancer survivorship, and chronic kidney disease. We also delineate persistent barriers to translation, including data heterogeneity, reliability under distribution shift, interpretability and uncertainty communication, and regulatory and ethical constraints. We argue that progress will depend on frameworks that link mechanistic insight with scalable digital monitoring and safety-aware adaptive prescriptions, enabling precision exercise biomedicine to move beyond empirical guidelines toward evidence-based, individualized decision-making.

Introduction: Multi-agent AI systems are believed to bring significant improvements in digital health, but it also brings new and more serious ethical issues. Such systems distribute the decision-making process among multiple interacting agents, and this decentralized decision-making system has raised ethical concerns in the medical field. On the one hand, it continues the ethical issues of traditional AI tools; on the other hand, the interaction processes within complex systems have also brought about new dilemmas. This narrative review aims to synthesize the ethical issues related to multi-agent AI systems in healthcare presented and explore the corresponding mitigation strategies. Methods: The study outcomes were synthesized using a narrative approach. Relevant records were gathered through Boolean searches in databases such as PubMed, Scopus, and Web of Science. A total of 21 articles related to multi-agent AI, healthcare, and ethical issues are included in this review. Results: Seven key ethical challenges were identified: (1) compound opacity, where interacting AI agents create layers of inscrutable decision-making; (2) error propagation and attribution difficulties, complicating accountability for clinical harm; (3) increased clinician dependence and automation bias, leading to potential deskilling and overreliance; (4) erosion of human oversight, as multi-agent AI systems operate beyond effective human control; (5) privacy and data security risks, stemming from complex data flows among agents; (6) threats to patient autonomy and informed consent, due to opaque or paternalistic AI recommendations; and (7) contextual blindness, reflecting a loss of individualized patient understanding in modular AI workflows. Furthermore, this review also summarized solutions proposed in the existing literature for these ethical issues. Conclusions: Multi-agent AI systems intensify existing ethical concerns in healthcare by distributing decision-making and blurring responsibility. To mitigate these issues, recent research advocates for the development of adaptive governance models, clear accountability frameworks, human–AI collaboration structures that preserve clinician authority, enhanced systems for explainability, and privacy-centered designs. In order to successfully incorporate agentic AI into healthcare, it is essential to maintain transparency, protect patient rights, and ensure that human-centered values continue to guide clinical decision-making in an era dominated by autonomous, interacting AI systems.

Fetal growth restriction (FGR) contributes to over 30% of late-pregnancy stillbirth, yet its diagnosis is challenging because current methods rely on indirect surrogate markers (estimated fetal weight and umbilical artery) that often fail to detect fetal compromise, particularly in late-onset cases. We hypothesized that fetal cardiac remodeling could provide a more robust basis for prediction. This study aimed to develop and validate the cardiac remodeling for FGR prediction model (CR-FGR), a first-in-class machine learning approach designed to operationalize the concept of fetal cardiac remodeling as a direct marker for FGR prediction. This multicenter study of singleton pregnancies included retrospective development (n = 663) and prospective validation in two independent cohorts (internal, n = 224; external, n = 51). The primary outcome was FGR (birth weight < 10th percentile). From 938 echocardiography videos, 222 cardiac parameters were extracted. A machine learning process selected the five most predictive parameters for the final logistic regression model (CR-FGR): right ventricular stroke volume/kg (RVSV/kg), cardiac output/kg (RVCO/kg), cardiac output (RVCO), left ventricular cardiac output (LVCO), and end-systolic area (RVESA). The CR-FGR model showed robust performance, with an area under the curve (AUC) of 0.872 (95% confidence interval (CI), 0.780–0.935) in the prospective internal testing set and 0.831 (95% CI, 0.674–0.947) in the external testing set. Its performance was comparable to a conventional EFW and Doppler model. Critically, the CR-FGR excelled in identifying challenging subgroups: it was highly effective for late-onset FGR (AUC 0.876, 95% CI, 0.748–0.951) and successfully detected FGR in many cases with normal umbilical artery Doppler, demonstrating its ability to capture pathology missed by traditional assessment. We developed and validated the first machine learning model for FGR prediction based on fetal cardiac remodeling. This model establishes a new diagnostic strategy, offering a powerful, complementary tool that captures direct evidence of fetal compromise. It significantly enhances risk stratification, particularly for the clinically challenging late-onset and Doppler-normal phenotypes of FGR. The Chinese Clinical Trial Registry, TRN: ChiCTR2000034182, Registration date: 27 June 2020.

This article presents UGache , a unified multi-GPU cache system designed for embedding-based deep learning (EmbDL). UGache is primarily motivated by the unique characteristics of EmbDL applications, namely read-only and skewed embedding accesses with affinity and predictability. UGache introduces a novel factored extraction mechanism that avoids bandwidth congestion to fully exploit high-speed cross-GPU interconnects (e.g., NVLink and NVSwitch). Based on a hotness metric, UGache also provides a near-optimal cache policy that balances local and remote access to minimize the extraction time for diverse GPU interconnect topologies. We have implemented UGache and integrated it into two representative frameworks, TensorFlow and PyTorch. Evaluation using two typical types of EmbDL applications, namely graph neural network (GNN) training and deep learning recommendation (DLR) inference, shows that UGache outperforms state-of-the-art replication and partition designs by an average of 1.93× and 1.63× (up to 5.25× and 3.45×), respectively. Furthermore, we demonstrate the applicability of UGache ’s principle beyond embedding-based deep learning, with an example of text-to-image generation on an inference cluster.

In recent years, paper-based humidity sensors have emerged as a highly promising technology for humidity detection. In this work, a polymerizable deep eutectic solvent (PDES) was prepared via a one-step blending method, which was applied to modify filter paper. The modification process did not alter the overall structure of the paper cellulose but rather targeted only its internal cellulose channels, thereby minimizing any impact on the paper’s original moisture-independent properties. The filter paper functioned both as the substrate and the humidity-sensing material in the fabricated sensor. The finger-like electrodes were designed using AutoCAD 2018 software and then printed onto the modified paper using screen-printing technology to fabricate the humidity sensor. Different saturated salt solutions were used to simulate corresponding humidity environments and evaluate the humidity performance of sensors. Compared with that of the blank paper-based humidity sensor, the sensitivity of the sensor modified by the PDES was significantly greater, and the recovery time was greatly shorter. Specifically, the sensitivity increased from 1.34 to 10.36 at 54% RH and from 166.24 to 519.2 at 98% RH. Additionally, the sensor response time was reduced from 728 s to 137 s. PDES modification significantly improved the moisture-sensitive characteristics and detection performance of the sensor.

Surface-confined DNA computing has emerged as a powerful information processing paradigm, offering enhanced specificity and accelerated reaction kinetics. Artificially designed DNA origami serves as a key enabler for such systems by providing a highly programmable platform for positioning computational components with nanometer resolution. However, conventional monolayer DNA origami circuits often exhibit non-negligible signal leakage, attributed to structural fluctuation-induced crosstalk between surface-confined molecules. Here, we utilize a rigidified double-layered uniaxial DNA origami platform to suppress the intrinsic structural fluctuation, thereby achieving high-fidelity signal propagation via minimizing fluctuation-mediated leakage. The rigidified DNA origami serves as a reliable computing platform to provide site arrangements that are narrowly distributed and closer to theoretical design. Basic propagation modules of varying orientations and spacings are implemented with reduced leakage and increased on-off ratios. We further demonstrate high-performance parallel transmission lines and logic gates on rigidified DNA origami. This approach establishes a generalizable strategy for engineering reliable platforms for surface-confined DNA computing.

In this study, a terahertz microfluidic multi-band sensor was designed. Unlike previous microfluidic absorption sensors that rely on dipole resonance, the proposed sensor uses a physical mechanism for absorption by exciting higher-order lattice resonances in microfluidic structures. With a Fabry–Perot cavity, the sensor can form an absorption peak with a high quality factor (Q) and narrow full width at half maximum (FWHM). A high Q value and a narrow FWHM are valuable in the field of sensing and provide strong support for high-precision sensing. On this basis, the sensing performance of the device was investigated. The simulation results clearly show that the absorption sensor has ultra-high sensitivity, which reaches 400 GHz/Refractive Index Unit (RIU). In addition, the sensor generates three absorption peaks, overcoming the limitations of a single frequency band in a composite resonance mode and multidimensional frequency response, which has potential application value in the field of volatile organic compound (VOC) sensing.

Seafood, especially marine fish, is highly prone to spoilage during processing, transportation, and storage. It releases pungent trimethylamine (TMA) gas, which severely affects food quality and safety. Metal–oxide–semiconductor (MOS) gas sensors for TMA detection offer a rapid, convenient, and accurate method for assessing fish freshness. Indium oxide (In₂O₃) has shown potential as an effective sensing material for the detection of TMA. In this work, one-dimensional In₂O₃ nanowires with different grain sizes and levels of crystallinity were synthetized using the electrospinning technique and underwent different thermal calcination processes. Gas-sensing tests showed that the In₂O₃–3 °C/min–500 °C gas sensor exhibited an outstanding performance, including a high response (R_a/R_g = 47.0) to 100 ppm TMA, a short response time (6 s), a low limit of detection (LOD, 0.0392 ppm), and an excellent long-term stability. Furthermore, the sensor showed promising experimental results in monitoring the freshness of Larimichthys crocea (L. crocea). By analyzing the relationship between the grain size and crystallinity of the In₂O₃ samples, a mechanism for the enhanced gas-sensing performance was proposed. This work provides a novel strategy for designing and fabricating gas sensors for TMA detection and highlights their potential for broad applications in real-time fish freshness monitoring.

The recent success of large language models (LLMs) in performing natural language processing tasks has increased interest in applying generative artificial intelligence (AI) to scientific research. However, a common problem of LLMs is their tendency to produce inaccurate and sometimes "hallucinated" outputs. Here, we established a generative AI tool, NanoSafari, to automatically extract knowledge from the biomedical nanoscience literature and address scientific queries. We developed the Grouped Iterative Validation based Information Extraction (GIVE) method to extract contextual information on nanoparticle characteristics from >20,000 published articles and established a database that was incorporated into the generative LLM to provide accurate nanomaterial design parameters. Blinded evaluation by biomedical nanoscientists showed that NanoSafari outperformed the baseline model in providing more reliable parameters for nanomaterial design tasks, as further validated by bench experiments. Together, these findings demonstrate the utility of AI-based methods for automated learning from "real-world" published work to provide accurate and reliable scientific references for biomaterial and bioengineering applications.

Background: Medical abstract sentence classification is crucial for enhancing medical database searches, literature reviews, and generating new abstracts. However, Chinese medical abstract classification research is hindered by a lack of suitable datasets. Given the vastness of Chinese medical literature and the unique value of traditional Chinese medicine, precise classification of these abstracts is vital for advancing global medical research. Objective: This study aims to address the data scarcity issue by generating a large volume of labeled Chinese abstract sentences without manual annotation, thereby creating new training datasets. Additionally, we seek to develop more accurate text classification algorithms to improve the precision of Chinese medical abstract classification. Methods: We developed 3 training datasets (dataset #1, dataset #2, and dataset #3) and a test dataset to evaluate our model. Dataset #1 contains 15,000 abstract sentences translated from the PubMed dataset into Chinese. Datasets #2 and #3, each with 15,000 sentences, were generated using GPT-3.5 from 40,000 Chinese medical abstracts in the CSL database. Dataset #2 used titles and keywords for pseudolabeling, while dataset #3 aligned abstracts with category labels. The test dataset includes 87,000 sentences from 20,000 abstracts. We used SBERT embeddings for deeper semantic analysis and evaluated our model using clustering (SBERT-DocSCAN) and supervised methods (SBERT-MEC). Extensive ablation studies and feature analyses were conducted to validate the model’s effectiveness and robustness. Results: Our experiments involved training both clustering and supervised models on the 3 datasets, followed by comprehensive evaluation using the test dataset. The outcomes demonstrated that our models outperformed the baseline metrics. Specifically, when trained on dataset #1, the SBERT-DocSCAN model registered an impressive accuracy and F₁-score of 89.85% on the test dataset. Concurrently, the SBERT-MEC algorithm exhibited comparable performance with an accuracy of 89.38% and an identical F₁-score. Training on dataset #2 yielded similarly positive results for the SBERT-DocSCAN model, achieving an accuracy and F₁-score of 89.83%, while the SBERT-MEC algorithm recorded an accuracy of 86.73% and an F₁-score of 86.51%. Notably, training with dataset #3 allowed the SBERT-DocSCAN model to attain the best with an accuracy and F₁-score of 91.30%, whereas the SBERT-MEC algorithm also showed robust performance, obtaining an accuracy of 90.39% and an F₁-score of 90.35%. Ablation analysis highlighted the critical role of integrated features and methodologies in improving classification efficiency. Conclusions: Our approach addresses the challenge of limited datasets for Chinese medical abstract classification by generating novel datasets. The deployment of SBERT-DocSCAN and SBERT-MEC models significantly enhances the precision of classifying Chinese medical abstracts, even when using synthetic datasets with pseudolabels.