Zhiwen ZHANG

Zhiwen ZHANG


Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
5 March 2026 in Frontiers in Public Health
Verified
Journal Article
Open access
PDF file
ZWZhiqiang Wang
Zhiqiang Wang
XLXiaoya Li
Xiaoya Li
CMChao Ma
Chao Ma
ZZZhiwen ZHANG
Zhiwen ZHANG
Objective: To compare the differences in content quality, readability, and actionability of patient education texts for self-management of chronic heart failure (CHF) generated by five mainstream large language models (LLMs) in China, and to provide a basis for platform selection and assessment framework construction for clinical use. Methods: A standardized set of 20 questions was developed based on literature review, guidelines, and consensus from cardiovascular experts, covering disease awareness, diagnosis and classification, treatment and rehabilitation, daily management and prevention, and psychosocial dimensions. Using a uniform prompt, responses were generated by DeepSeek-R1, Doubao, ERNIEBot 4.5 Turbo, Qwen3-Max-Thinking-Preview, and Kimi K2. The PEMAT-P scale was used to assess understandability and actionability, 36-item expanded EQIP (EQIP-36 score) scale was used to evaluate information completeness and standardization, and Global Quality Score (GQS) was used to assess overall quality. Additionally, seven readability formulas, including Flesch Reading Ease Score (FRES) and Flesch–Kincaid Grade Level (FKGL), were computed for comparison. Results: Overall quality was high [GQS median 5.00 (4.00–5.00)] with significant between-platform differences (χ 2 = 14.47, P = 0.006). Doubao and Kimi K2 achieved the highest GQS [both 5.00 (5.00–5.00)]. DeepSeek-R1 showed the greatest information completeness [EQIP-36 39.20 (36.17–44.23); χ 2 = 25.07, P < 0.001] but the lowest readability [FRES 19.32 (17.94–36.89) and FKGL 14.28 (13.02–15.85); both P < 0.001]. ERNIEBot 4.5 Turbo and Qwen3-Max-Thinking-Preview were most readable (FRES ≈ 59; FKGL ≈ 8; both P < 0.001) but had lower EQIP-36 scores. Actionability was limited overall [PEMAT-P actionability 20.00% (0.00–40.00); χ 2 = 26.40, P < 0.001] and varied by topic, with daily management and prevention outperforming disease knowledge and diagnosis/classification (χ 2 = 20.86, P < 0.001). Conclusion: LLMs show potential for use in patient education for CHF, but there is a structural trade-off between information detail and readability, as well as gaps in actionability and verifiability. It is recommended to combine enhanced search and structured template generation strategies, and establish a governance feedback loop involving prompt engineering, clinical expert review, and continuous monitoring to improve readability alignment, completeness of action instructions, and patient safety.

1 February 2026 in Digital Health
Verified
Journal Article
Open access
MHMengqiu Hu
Mengqiu Hu
ZWZhiqiang Wang
Zhiqiang Wang
ZZZhiwen ZHANG
Zhiwen ZHANG
MLMuwei Li
Muwei Li
Objective: To assess the quality, readability, and actionability of hypertension patient education materials generated by six selected patient-facing large language model (LLM) platforms, and to characterize cross-platform heterogeneity to guide the optimization of AI-generated patient education materials. Methods: Six publicly accessible or commonly accessible platforms were evaluated using standardized prompts to generate patient education materials. Understandability and actionability were assessed using the Patient Education Materials Assessment Tool for Printable Materials (PEMAT-P). Information quality was evaluated using the expanded Ensuring Quality Information for Patients scale (EQIP-36), and overall quality was rated using the Global Quality Score (GQS). Readability was compared using seven metrics, including the Flesch Reading Ease Score (FRES) and the Flesch–Kincaid Grade Level (FKGL). Results: Overall educational quality and understandability were generally favorable, but substantial cross-platform heterogeneity was observed. Readability remained challenging, with low FRES values and grade-level indices generally exceeding commonly recommended thresholds for patient education materials. Qwen3-Max-Thinking-Preview achieved the highest PEMAT-P total score (77.00), followed by ChatGPT 5.2-Thinking (72.22). For EQIP-36, Qwen3-Max-Thinking-Preview scored highest (49.28), followed by DeepSeek-R1 (45.48). DeepSeek-R1 generated the most readable materials among the evaluated platforms, with a median FRES of approximately 42.14 and FKGL of approximately 10.10, whereas Qwen3-Max-Thinking-Preview showed lower readability, with a median FRES of approximately 16.82 and FKGL of approximately 14.38. Kimi K2 showed high PEMAT-P understandability (76.90) but low actionability (10.00). Post hoc analyses showed that Qwen3-Max-Thinking-Preview significantly outperformed ERNIE Bot 4.5 Turbo, Doubao, and Kimi K2 on PEMAT-P total score, and outperformed Kimi K2 and ERNIE Bot 4.5 Turbo on EQIP-36. DeepSeek-R1 also outperformed Kimi K2 and ERNIE Bot 4.5 Turbo on EQIP-36. Across content domains, actionability was significantly higher for Daily Care and Prevention than for Basic Understanding of the Disease and Complications, Psychological and Social Aspects. Conclusions: Generative AI shows promise for hypertension patient education, particularly in improving understandability. However, actionability remains a major limitation of current outputs, highlighting the need for platform-aware optimization strategies that explicitly strengthen step-by-step and action-oriented guidance.

10 June 2026 in Frontiers in Cardiovascular Medicine
Verified
Journal Article
Open access
PDF file
BDBo Deng
Bo Deng
ZWZhiqiang Wang
Zhiqiang Wang
TCTong Cheng
Tong Cheng
ZZZhiwen ZHANG
Zhiwen ZHANG
MLMuwei Li
Muwei Li
Background: Patients with cardiac myxoma require long-term follow-up, and the quality of patient education as well as the ability to recognize early symptoms may influence integrated postoperative cardiovascular and tumor-related management. Although large language models (LLMs) have been increasingly applied in medical education, cross-platform empirical evidence in this intersecting context remains limited. This study evaluated the performance and limitations of widely used LLMs in generating patient education texts for cardiac myxoma using a standardized, expert-curated educational question set. Methods: We constructed a standardized dataset of 60 expert-curated educational questions spanning the disease course of cardiac myxoma and used nine widely used LLMs to generate 540 patient education texts. Text quality, readability, and actionability were assessed using the Patient Education Materials Assessment Tool for Printable Materials (PEMAT-P), Ensuring Quality Information for Patients (EQIP-36), the Global Quality Score (GQS), and seven readability formulas. In addition, a regularized Gaussian graphical model-based partial correlation network analysis and latent profile analysis were performed to identify relationships among evaluation metrics and cross-platform text phenotypes. Results: Texts generated across platforms showed significant heterogeneity in information quality, understandability, and objective readability, whereas clinical actionability was generally low. Word count showed the strongest positive correlation with the total EQIP-36 score and occupied a central position in the network. Reading-difficulty indices were consistently negatively correlated with PEMAT-P actionability. Latent profile analysis identified three text phenotypes: moderate-quality/low-readability, high-quality/high-actionability, and low-quality/easy-to-read. Ideally suited patient education texts accounted for only a very small proportion of all outputs. Conclusions: Within this expert-curated educational question set, the current application of LLMs in patient education for cardiac myxoma is primarily limited by reading burden and insufficient behavioral guidance. Although longer outputs were generally associated with higher informational quality scores, greater syntactic complexity was associated with lower actionability of the materials. In addition, latent profile analysis suggested that only a small subset of outputs approached a more favorable quality–actionability profile.

3 February 2026 in IEEE Transactions on Consumer Electronics
Verified
Journal Article
Closed access
JYJing Yang
Jing Yang
VGVijay Govindarajan
Vijay Govindarajan
MKMuhammad Attique Khan
Muhammad Attique Khan
Zaffar Ahmed ShaikhZaffar Ahmed Shaikh
Zaffar Ahmed Shaikh
SAShrooq Alsenan
Shrooq Alsenan
QGQuan Guo
Quan Guo
Consumer-Centric Digital Twins (CCDT) are emerging as transformative enablers in Healthcare 5.0, offering real-time monitoring, predictive analytics, and patient-specific treatment optimization. This paper introduces CardioTwin-XAI, a novel AI-driven CCDT framework designed for early detection, risk prediction, and personalized management of Coronary Artery Disease (CAD) and related conditions such as atherosclerosis, myocardial infarction (MI), and heart failure (HF). The proposed system integrates IoT-enabled wearable sensors, continuous hemodynamic profiling, and federated multimodal learning to build a dynamic, patient-specific digital replica. The key innovation of CardioTwin-XAI is its unified integration of explainable AI (XAI) with privacy-preserving federated learning, wherein interpretable attribution mechanisms are embedded within distributed model updates, enabling transparent predictions without compromising data confidentiality across healthcare institutions. A hybrid transformer-GNN architecture captures complex spatiotemporal cardiovascular dynamics, while differential privacy and blockchain-based data sharing ensure regulatory compliance. Evaluation on multi-institutional datasets demonstrates 92.6% accuracy in CAD prediction, 31% improvement in treatment planning efficiency, and 40% reduction in unnecessary interventions. The proposed framework paves the way for scalable, secure, and explainable cardiovascular care in next-generation healthcare ecosystems.

24 February 2026 in IEEE Transactions on Consumer Electronics
Verified
Journal Article
Closed access
ZZZhiwen ZHANG
Zhiwen ZHANG
VGVijay Govindarajan
Vijay Govindarajan
XOXin Ouyang
Xin Ouyang
JYJing Yang
Jing Yang
BABayan Alabdullah
Bayan Alabdullah
MLMuwei Li
Muwei Li
Cardiovascular diseases (CVDs) are the most common causes of death in the world, especially coronary artery disease (CAD) and heart failure (HF). Prevention and sustained monitoring in the early stages are essential, yet conventional diagnostic methods are expensive, invasive, and limited to the clinical environment. Recent developments in consumer electronics (CE), including smartwatches and smartphones, are providing new opportunities for real-time, non-invasive cardiovascular monitoring. Nevertheless, the implementation of deep learning (DL) on such devices poses challenges for explainability, computational efficiency, and privacy. This paper introduces CardioGPT-ViT, a novel multimodal, explainable DL model for detecting CAD and HF on edge devices. The model combines Vision Transformers (ViT), GPT-style attention, Mamba state-space modeling, and a small You Only Look Once (YOLO) detection head and incorporates the explainability functionality of clinical trust. CardioGPT-ViT was trained and tested on three benchmark datasets: PulseDB, UCI Heart Disease, and an external electrocardiography (ECG)/photoplethysmography (PPG) cohort. The framework includes SHapley Additive exPlanations (SHAP), Local explainable Model-Agnostic Explanation (LIME), attention maps, and Gradient-weighted Class Activation Mapping (Grad-CAM) to explain results. Quantization, pruning, and latency/energy profiling were used to verify the feasibility of edge deployment on consumer devices. It was compared to performance with recent baseline models across convolutional neural networks (CNNs), ViTs, and state-space architectures. CardioGPT-ViT also scored over 0.90 across all datasets, reaching a maximum of 0.962 on PulseDB. The model was found to be more accurate (92.8%) and robust and had real-time feasibility (latency of less than 50 ms) and low energy consumption (less than 80 mJ per inference). Explainability results correlated with clinical biomarkers, increasing trust. CardioGPT-ViT provides a clinically explainable, resource-efficient, and edge-deployable solution to next-generation cardiovascular care, connecting the gap between consumer electronics and long-term preventive healthcare.

12 November 2025 in Reviews in Cardiovascular Medicine
Verified
Journal Article
Open access
QSQingbo Shi
Qingbo Shi
YGYang Gao
Yang Gao
ZZZhiwen ZHANG
Zhiwen ZHANG
ZSZhuocheng Shi
Zhuocheng Shi
HYHaosen Yu
Haosen Yu
MLMuwei Li
Muwei Li
Background: Previous studies have shown a strong link between sleep and cardiovascular disease. However, the association of sleep duration and quality with coronary atherosclerotic plaque vulnerability remains unclear. This study aimed to investigate the correlation between sleep duration, sleep quality, and coronary plaque vulnerability using optical coherence tomography (OCT). Methods: A total of 260 patients with stable angina who completed an OCT examination were included. Patients were divided into a thin-cap fibroatheroma (TCFA) group and a non-TCFA group according to the presence of TCFA on OCT. The sleep duration of the patients was recorded by questionnaire, and the sleep quality was evaluated using the Pittsburgh Sleep Quality Index (PSQI). Results: The TCFA group had significantly shorter sleep duration and higher PSQI values (p < 0.05). A multivariable logistic regression analysis revealed that sleep duration and PSQI were independent predictors of TCFA (p < 0.05). A receiver operating characteristic (ROC) study demonstrated that the area under the curve values for sleep duration and PSQI were 0.698 and 0.721, respectively, in predicting the presence of TCFA. Patients with a sleep duration ≤5.5 hours or a PSQI value >9 had a thinner fibrous cap thickness, a larger maximal lipid pool arc, and a higher incidence of TCFA and macrophage deposition (p < 0.05). Sleep duration was positively correlated with the thinnest fibrous cap thickness (r = 0.451; p < 0.001), and negatively correlated with the radian of the maximum lipid pool (r = –0.470; p < 0.001). The PSQI was negatively correlated with the thinnest fibrous cap thickness (r = –0.477; p < 0.001), and positively correlated with the radian of maximum lipid pool (r = 0.340; p < 0.001). Conclusion: Both sleep duration and sleep quality were significantly associated with coronary plaque vulnerability. Patients with either insufficient sleep duration or poor sleep quality exhibited significantly greater plaque vulnerability.