Traditional cultural tourism interest point recommendation systems rely solely on single-modal data modeling, they fail to accurately uncover tourists' deeper preferences and dynamic travel needs, making it difficult to meet the evolving demands of contemporary cultural tourism consumption. This paper reviews the multimodal data architecture in the cultural tourism domain, categorizes cultural tourism data into five major types, and clarifies the applicable scenarios and application advantages of various domestic and international datasets. Furthermore, leveraging cutting-edge technologies, this paper constructs a comprehensive multimodal recommendation framework for cultural tourism. Subsequently, from the two core dimensions of data and algorithms, the paper conducts an in-depth analysis of the key bottlenecks and practical implementation challenges in current research and, in alignment with industry development trends, proposes several forward-looking research directions. This study enhances the theoretical and technical framework for multimodal recommendation in cultural tourism, providing a solid theoretical basis and engineering reference for numerous smart cultural tourism scenarios.

Accurate detection of intracranial aneurysms is essential for further imaging assessment, specialist referral, and clinical management. Computed tomography angiography and time-of-flight magnetic resonance angiography are commonly used non-invasive methods for cerebrovascular examination. However, small aneurysms may go unnoticed due to their limited size, complex vascular anatomy, and imaging artifacts. Artificial intelligence (AI) can automatically identify suspicious lesions and may improve reader performance. This review summarizes studies published between 2022 and 2026 on AI-assisted detection of intracranial aneurysms using computed tomography angiography and time-of-flight magnetic resonance angiography. It focuses on the detection performance for small lesions, common false-positive and false-negative patterns, and the clinical value of human–AI collaboration. Current evidence suggests that artificial intelligence provides satisfactory overall detection performance. In some studies, it also improved clinicians' detection sensitivity and reading efficiency. However, aneurysms measuring ≤3 mm and some atypical lesions remain difficult to detect. Increasing model sensitivity may also result in more false-positive markings. Incorrect artificial intelligence suggestions may further lead to automation bias. At present, AI is more suitable as a lesion-alert tool and a second reader than as a replacement for clinicians in final diagnosis. Future research should investigate the causes of detection errors in relation to dataset composition, model design, external validation, and human–AI interaction. Diagnostic performance, false-positive burden, and effects on patient management should also be evaluated in real clinical settings.

Precision oncology requires robust drug-response prediction because single biomarkers cannot fully explain therapeutic variation across tumors, cell lines, and drug structures. This study develops an AI-driven multi-omics model using GDSC, CCLE, DepMap, CTRP, and TCGA data, integrating mutation, copy-number variation, RNA expression, DNA methylation, drug SMILES, and IC50/AUC responses. Omics features are encoded separately, drug structures are modeled with a graph attention network, and cross-attention learns tumor-drug interactions. The model outperforms Elastic Net, Random Forest, XGBoost, and DeepCDR-style baselines, achieving RMSE of 0.684±0.014 and AUC of 0.872±0.011. Interpretability analysis highlights EGFR/ERBB, DNA repair, cell cycle, and PI3K-AKT pathways, demonstrating that multi-omics fusion and drug-structure modeling improve prediction stability and support interpretable precision-oncology drug screening.

Rolling bearings are key components in rotating machinery. Their wear and failure can cause the entire equipment to stop operating. Therefore, it is extremely necessary to predict their remaining useful life (RUL). However, in reality, the wear of bearings does not change linearly and the uncertainty over time makes predicting RUL challenging. To address this issue, this paper proposes an end-to-end RUL prediction framework that combines multi-dimensional time-domain feature extraction with stacked long short-term memory (LSTM) networks. By extracting eight time-domain statistical indicators (such as RMS and Kurtosis), and using the sliding window mechanism to construct the feature sequence, it is input into the stacked LSTM model to capture the continuous degradation trajectory. The 15 bearings in the XJTU-SY benchmark dataset were evaluated against the ground-truth degradation targets. Experimental results show that the proposed framework can capture the degradation trends under different working conditions. The predictions are less accurate in the early stage, mainly because the time-domain features remain relatively stable during this period. However, the model provides reasonable overall accuracy and follows the degradation trend well in the later stage.

Vision-based gesture pipelines run sub-30 ms per frame, yet closed-loop interaction often exceeds ~200 ms. This study treats latency as an end-to-end property spanning perception, inference, decision and feedback. This critical survey synthesizes a purposively screened corpus of gesture and adjacent interactive systems (2020–2026), plus foundational anchors. This study uses a three-layer taxonomy: system compensation, semantic adaptive inference, and hybrid semantic–system coordination. This study finds that system-level remedies dominate MediaPipe-era pipelines but remain blind to recognition uncertainty. Semantic adaptivity optimizes compute under uncertainty but is rarely bound to interactive latency service level objectives (SLOs). The missing link is the runtime interface between confidence and load. This paper proposes a three-layer taxonomy that defines hybrid coordination as a missing layer, identifies the perception–E2E latency gap as the central empirical blind spot, and formalizes Load–Confidence Coupling as a testable proposition with falsification criteria.

Probing tasks in complex and unknown working environments are placing increasingly demanding requirements on robots in terms of terrain adaptability and locomotion agility. Against this background, conventional rigid robots have shown obvious shortcomings, while bio-inspired robots—developed on the basis of biological body shapes and motion mechanisms—have gradually become one of the most compelling research directions in robotics. Taking operating scenarios as the organizing thread, this paper sorts out morphological bionic design methods and summarizes how representative bio-inspired robot configurations perform in terms of technical characteristics, locomotion capability, and suitable working conditions. The analysis shows that most existing bio-inspired robots are geared toward a single scenario, and their motion ability still falls noticeably short of the biological organisms they imitate. In response to these problems, follow-up research may draw on multimodal reconfigurable morphologies and deep-learning-assisted design to improve the overall performance of bio-inspired robots. The paper reviews the morphological design logic and functional integration schemes of bio-inspired robots under multi-scenario conditions, with the aim of offering a useful theoretical reference and technical route for the development of high-performance bio-inspired robots for specialized missions.

Individuals with severe limb impairments encounter substantial barriers during human‑computer interaction, while commercially available assistive input devices are often prohibitively expensive. This paper presents a low‑cost wearable assistive input device (bill‑of‑materials cost < ¥130) built around the nRF52832 Bluetooth‑Low‑Energy (BLE) system‑on‑chip and an MPU‑6050 inertial measurement unit. A real‑time‑clock‑based asynchronous event‑driven architecture reduces power consumption, and the sensor's on‑chip 5 Hz digital low‑pass filter suppresses involuntary muscle tremor at the signal source. A multi‑threaded PC‑side client maps real‑time posture data to discrete keyboard commands. A user study with three participants exhibiting distinct physiological profiles was carried out under two interaction modes. In discrete click‑trigger mode, the system achieved a peak directional‑selection throughput of 12 WPM, while continuous dwell‑hold control yielded an average throughput of 4.2 WPM. The participant with a full, unrestricted range‑of‑motion (ROM) reached a trigger accuracy of 98.2 %, whereas the participant with severely‑restricted head movement obtained only 64.0 % accuracy. Experimental results demonstrate that fixed‑cutoff on‑chip filtering carries inherent performance limitations for users with extremely limited motion range. Directions for future work include adaptive filtering algorithms and multi‑modal interaction schemes.

A substantial body of academic research has examined the determinants of undergraduate attrition and premature program departure. This study investigates whether logistic regression could identify students who discontinued their programs compared to those who graduated. An 80/20 stratified train/test split was used to evaluate three logistic regression models based upon 3,630 resolved outcomes, where 1,421 were designated as Dropout and 2,209 as Graduate. Model E included all enrolment variables. First-semester academic records were added to Model S1. Second-semester academic records were added to Model S2. The receiver operating characteristic curve (ROC-AUC) values were 0.836, 0.917 and 0.935, respectively. The Brier scores were 0.155, 0.099 and 0.080, respectively. Older age, male gender, and debtor status were positively related to dropout odds. As exact dropout timestamps are unavailable, semester-based models are treated as post facto. These results support targeted assistance to students experiencing difficulty rather than inferring causality.

Chest imaging report generation aims to automatically generate radiology reports from images such as chest X-rays and is an important task at the intersection of medical image analysis and natural language generation. In recent years, multimodal large language models have improved the fluency and organization of generated reports, but they still face limitations related to insufficient medical knowledge, inadequate modeling of anatomical relationships, and hallucination. This article reviews studies on chest imaging report generation, the integration of knowledge graphs and large language models, applications of graph neural networks in medical imaging, and the evaluation of multimodal foundation models for radiology. Existing studies indicate that knowledge graphs can provide structured medical concepts, relationships between diseases and anatomical sites, and traceable external knowledge, whereas graph neural networks are suited to relational propagation, selection, and compression in graph space.

The growth of digital financial media has led to the faster and larger dissemination of market information while varying levels of emotionality in the news reporting from different sources can create conflicting public signals and add to information asymmetry. While there is some existing literature on the connection between aggregate news sentiment and stock returns, volatility, and trading volume, the phenomenon of sentiment bias in cross-media news reporting has not been adequately studied yet. In order to fill the gap, financial news about S&P 100 firms published during 2021–2025 is gathered from GDELT and matched against the daily market data from Yahoo Finance. Headlines are analyzed using FinBERT to find the level of their sentiment while Sentence-BERT and HDBSCAN determine whether the news item refers to the same event. Event-level median sentiment becomes a reference point while deviations from the consensus sentiment measure the cross-media bias in public sentiment. The impact of sentiment biases on the market is assessed using the Corwin-Schultz bid-ask spread, Amihud illiquidity, and two-way fixed effects regressions. It was found that greater cross-media divergence in news sentiments causes the widening of next day bid-ask spread and increased illiquidity; moreover, negative deviations have a stronger impact than positive. These correlations were consistent after controlling for return, volatility, trading volume, and news attention as well as exclusion of bias outliers.