电子科技大学（深圳）高等研究院

Time-series based deep learning methods have significantly improved performance of predictive healthcare tasks on electronic health records (EHR) data. However, mortality prediction for first admissions is a huge challenge due to the absence of historical visit sequences. Analysis of major healthcare databases reveals that 55.17% in MIMIC-IV and 83.80% in MIMIC-III present with single visit records only. Many of these single-visit patients are admitted directly to ICU, highlighting the critical need for early prediction models that can inform timely interventions and substantially impact patient outcomes, even without longitudinal history. On first-admission mortality prediction, Dual Ontology-enhanced Clinical Decision learning (DOCD) has been proposed in this paper where we effectively leverage clinical knowledge through two key components. Dual ontology-enhanced learning extracts hierarchical representations from both diagnosis and procedure taxonomies. Furthermore, a priori-guided attention mechanism integrates clinical knowledge through probability based regularization to model the decision-making process. An information fusion is then employed to seamlessly integrate the demographic data, vital signs, and knowledge enhanced medical codes. Extensive experiments on MIMIC III (AUROC: 0.9528, AUPRC: 0.8971) and MIMIC-IV (AUROC: 0.9817, AUPRC: 0.8857) datasets demonstrate DOCD's superior performance over all baselines, offering interpretable visualizations aligned with established clinical knowledge.

This article presents a 220GHz, scalable planar phased array with precise lambda /2 antenna spacing based on 0.13- mu m SiGe BiCMOS technology. A novel joint design incorporating an image rejection subharmonic mixer (IRSHM) and a vector modulator (VM) is proposed based on the current reuse technique to effectively minimize power consumption and occupied area. And this architecture simultaneously enables image rejection, subharmonic up-conversion, and intermediate frequency (IF) phase and amplitude control. The proposed phased-array chip employs a splicing architecture with a 2x2 array and on-chip antenna, where a ninth harmonic generator, a local oscillator (LO) signal distribution network, and four highly integrated transmitters (TXs) are integrated. Through the chip-edge wire bonding, a 4x8 planar phased array is realized with an RF 3-dB bandwidth of 212-224GHz. In the measurement, the peak effective isotropic radiated power (EIRP) is 36dBm of a 32-element phased array at 221GHz, with about 2dBm output power per channel. A 3-dB beam widths of 10 degrees (horizontal) and 17 degrees (vertical), with beam steering capabilities of +/- 60 degrees and +/- 40 degrees in the respective planes are achieved. Furthermore, a wireless communication link using a 32-element phase array and a 40-dBi dish antenna is demonstrated, achieving data rates of 4 Gbps (QPSK) and 6 Gbps (16-QAM) with corresponding error vector magnitude (EVM) values of 9.5% and 9.2%, respectively. An inverse synthetic aperture radar (ISAR) imaging experiment involving two corner reflectors is also demonstrated. The measured azimuth resolution is 8mm, and two corner reflectors can be clearly distinguished.

High-valence single-atom catalysts (HVSACs) generally exhibit superior catalytic activity owing to the fast electron transfer efficiency between HVSACs and their supportive substrate. Nevertheless, the thermodynamically unstable HVSACs are prone to spontaneous aggregation during their conventional synthesis processes, largely limiting their practical application demanding long-term stability. Herein, we demonstrate a universal approach to synthesize stable HVSACs by incorporating high-electronegativity fluorine (F) atoms to directly coordinate with the targeted metal atoms (M) of HVSACs on nitrogen-carbon supports. As a result, an asymmetric planar M-N2F2 structure is introduced, which boosts asymmetric polarized spin between M and F atoms with M stabilized at their high-spin and high-valence state. Leveraging this motif, we further demonstrate its high universality as a promising synthetic approach by obtaining stable HVSACs for 22 individual metal elements. Using high-valence Mn single-atom as a model system, the resulting MnN2F2 catalyst exhibits not only high activity but also superior durability in both the oxygen reduction reaction (ORR) and practical Zn-air batteries, outperforming conventional MnN4 single atoms.

Autonomous Underwater Vehicles (AUVs) provide a new prospect for data collection in underwater wireless sensor networks (UWSNs). For dynamic underwater environments, researchers typically apply deep reinforcement learning (DRL) to design multi-AUV collection schemes for UWSNs. However, these methods suffer from the following issues. 1) Overloaded observations. The importance of various observations for an AUV varies over time. Considering all observations equally complicates decision-making for subsequent actions. 2) Dynamic scale of AUVs and sensors. Once the number of AUVs or sensors is changed, traditional static neural networks require retraining, lacking scalability across diverse scenarios. To solve the above issues, we propose a Generalizable Attention-based Data collection scheme (GAMD) for Multi-AUV UWSNs, while enhancing AUVs’ collection efficiency. GAMD incorporates the attention mechanism with multi-agent DRL framework, which enables AUVs to prioritize observations more critical for action decisions. Moreover, we propose an adaptive information processing approach, enabling the AUV policy model to seamlessly adapt to various scenarios without retraining. Additionally, we develop a training paradigm with incremental complexity across different scale scenarios to simplify training process and accelerate convergence. Simulation results demonstrate that GAMD alleviates the training cost compared to the state-of-the-art methods, and simultaneously optimizes collection energy efficiency, collection time, and trajectory distance.

Current automated analog sizing relies heavily on simulation-based “opaque-box” optimization in the size domain. This forces the optimizer to directly navigate a highly non-linear mapping, failing to bridge the semantic gap between physical device sizes and high-level circuit specifications. To mitigate this non-linearity, we propose a behavior-centric sizing framework that decouples the optimization process into an efficient search over behavioral parameters (BPs) and a subsequent BP-to-size direct mapping for a given process. Specifically for amplifiers, the BP set (gm, fT ,AV ) encapsulates drive strength, high-frequency speed, and intrinsic gain, utilizing parameters fundamental to small-signal analysis and effectively transforming the optimization into a more linear and well-behaved search space. Exploiting this well-behaved space, we leverage Kolmogorov-Arnold Networks (KANs) to build data-efficient, symbolic surrogate models that map BPs to circuit performance, and employ a lookup-table (LUT)-based back-mapping to cleanly instantiate the optimized BPs into device sizes. Implemented in the open-source tool ‘Amplifier Copilot’, the framework generated a 7,200-point design database across 22 topologies, four process nodes, and both low-voltage and high-voltage device options. Silicon validation in a 22 nm CMOS process confirms that the measured silicon performance aligns with simulation predictions, validating the practical feasibility of the proposed framework. The Amplifier Copilot is available at https://github.com/AmpCopilot/Amplifier-Copilot.