Zhizhong Fu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Zhizhong Fu has a total of 24 co-authors. Here are the top co-authors by number of shared publications:
- Xiaorong Ding
University of Electronic Science and Technology of China
5 shared publications - Yifan Chen
University of Electronic Science and Technology of China
3 shared publications - Xinyue Song
School of Life Science and Technology, University of Electronic Science and Technology of China, Sichuan, China
2 shared publications Yifan Chen
2 shared publications- Andrew Lowe
Auckland University of Technology
1 shared publication Deqi Chen
School of Life Science and Technology, University of Electronic Science and Technology of China, Sichuan, China
1 shared publication
Latest Publications
- 20 August 2026
- Published by IOP Publishing in Physiological Measurement
- Vol. 47 (8) , 085025
Slide-DML: sliding-window-based estimation of heterogeneous treatment effects
Objective. Heterogeneous treatment effect (HTE) estimation is essential for understanding individual differences in physiological responses and supporting personalized healthcare. However, existing non-parametric HTE estimation methods often rely on complex partition strategies and may have limited interpretability when applied to physiological measurements with continuous variations and non-uniform data distributions. This study aims to develop an adaptive and interpretable framework for HTE estimation in physiological measurement systems. Approach. We propose a sliding-window-based double machine learning framework (Slide-DML) for non-parametric HTE estimation. Slide-DML adaptively constructs quasi-homogeneous local windows based on treatment effect variation and estimates local linear HTE within each selected window. The local estimates are subsequently aggregated using adaptive weighting to obtain a smooth global HTE function while preserving interpretability. Main results. The performance of Slide-DML was evaluated using synthetic data, semi-synthetic clinical data, and real physiological measurements. In synthetic experiments, Slide-DML achieved a mean squared error (MSE) of 0.006, outperforming existing machine learning-based HTE estimation methods. In semi-synthetic experiments, Slide-DML achieved a root MSE of 3.526, demonstrating superior performance compared with both machine learning-based and deep learning-based approaches. Experiments on real photoplethysmogram and electrocardiogram data further showed that the estimated treatment effect curves were consistent with established cardiovascular knowledge and provided improved interpretability. Significance. Slide-DML provides an effective and interpretable approach for estimating HTE in physiological measurement systems. By capturing continuous variations in physiological states, the proposed framework may facilitate individualized analysis of cardiovascular responses and support personalized healthcare applications, such as cuffless blood pressure monitoring using wearable physiological devices.
- 20 August 2026
- Published by IOP Publishing in Physiological Measurement
- Vol. 47 (8) , 085025
Slide-DML: sliding-window-based estimation of heterogeneous treatment effects
Objective. Heterogeneous treatment effect (HTE) estimation is essential for understanding individual differences in physiological responses and supporting personalized healthcare. However, existing non-parametric HTE estimation methods often rely on complex partition strategies and may have limited interpretability when applied to physiological measurements with continuous variations and non-uniform data distributions. This study aims to develop an adaptive and interpretable framework for HTE estimation in physiological measurement systems. Approach. We propose a sliding-window-based double machine learning framework (Slide-DML) for non-parametric HTE estimation. Slide-DML adaptively constructs quasi-homogeneous local windows based on treatment effect variation and estimates local linear HTE within each selected window. The local estimates are subsequently aggregated using adaptive weighting to obtain a smooth global HTE function while preserving interpretability. Main results. The performance of Slide-DML was evaluated using synthetic data, semi-synthetic clinical data, and real physiological measurements. In synthetic experiments, Slide-DML achieved a mean squared error (MSE) of 0.006, outperforming existing machine learning-based HTE estimation methods. In semi-synthetic experiments, Slide-DML achieved a root MSE of 3.526, demonstrating superior performance compared with both machine learning-based and deep learning-based approaches. Experiments on real photoplethysmogram and electrocardiogram data further showed that the estimated treatment effect curves were consistent with established cardiovascular knowledge and provided improved interpretability. Significance. Slide-DML provides an effective and interpretable approach for estimating HTE in physiological measurement systems. By capturing continuous variations in physiological states, the proposed framework may facilitate individualized analysis of cardiovascular responses and support personalized healthcare applications, such as cuffless blood pressure monitoring using wearable physiological devices.
- Open Access
- 22 July 2026
- Published by Elsevier in SSRN Electronic Journal
Additive Causal Construction for Transferable and Reconfigurable Cross-system Learning in Multi-Source Image Fusion
In multi-source image fusion scenarios, heterogeneous inputs are typically driven by distinct generative mechanisms and can be viewed as a composition of multiple causal systems. However, cross-system discrepancy (CSD) and crosssystem entanglement (CSE) commonly arise during the fusion process, often leading to significant performance degradation under out-of-distribution (OOD) predictions. To address the CSD and CSE issues, we propose the additive causal construction (ACC) framework, which characterizes information fusion at two levels: firstly, it establishes causal "anchors" shared among multiple systems through intervention consistency to enable causal graph transferability (CGT); and secondly, it formalizes the fusion process as causal construction and models the reliability of constructed paths through uncertainty quantification to ensure causal graph reconfigurability (CGR). Building upon this, we revisit the traditional causal representation learning (CRL) with ACC and propose ACC-CRL as a learnable instantiation of the framework. The method explores joint causal content representations across systems via content-mechanism decoupling, and performs response alignment under shared anchors to mitigate CSD. Furthermore, it incorporates structural uncertainty to adaptively regulate the fusion process, thereby suppressing unstable CSE. We conduct systematic experiments on synthetic data (ColorMNIST) and real-world multi-center medical imaging tasks (microvascular invasion (MVI) prediction). The results demonstrate that the proposed method significantly improves OOD generalization while maintaining in-distribution (ID) performance, validating the effectiveness and robustness of the ACC-CRL strategy based on mechanism alignment and uncertainty modeling in open environments.
- 22 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessAdditive Causal Construction for Transferable and Reconfigurable Cross-system Learning in Multi-Source Image Fusion
In multi-source image fusion scenarios, heterogeneous inputs are typically driven by distinct generative mechanisms and can be viewed as a composition of multiple causal systems. However, cross-system discrepancy (CSD) and crosssystem entanglement (CSE) commonly arise during the fusion process, often leading to significant performance degradation under out-of-distribution (OOD) predictions. To address the CSD and CSE issues, we propose the additive causal construction (ACC) framework, which characterizes information fusion at two levels: firstly, it establishes causal "anchors" shared among multiple systems through intervention consistency to enable causal graph transferability (CGT); and secondly, it formalizes the fusion process as causal construction and models the reliability of constructed paths through uncertainty quantification to ensure causal graph reconfigurability (CGR). Building upon this, we revisit the traditional causal representation learning (CRL) with ACC and propose ACC-CRL as a learnable instantiation of the framework. The method explores joint causal content representations across systems via content-mechanism decoupling, and performs response alignment under shared anchors to mitigate CSD. Furthermore, it incorporates structural uncertainty to adaptively regulate the fusion process, thereby suppressing unstable CSE. We conduct systematic experiments on synthetic data (ColorMNIST) and real-world multi-center medical imaging tasks (microvascular invasion (MVI) prediction). The results demonstrate that the proposed method significantly improves OOD generalization while maintaining in-distribution (ID) performance, validating the effectiveness and robustness of the ACC-CRL strategy based on mechanism alignment and uncertainty modeling in open environments.
- 10 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Biomedical Engineering
- Vol. PP (00189294) , 1-11
Causal Decomposition of PPG Signals for Cuffless Blood Pressure Estimation
Objective: While photoplethysmogram (PPG) signals are physiologically linked to cardiac activity and widely used for cuffless blood pressure (BP) estimation, their correlation with BP remains limited in reliability due to inter-individual heterogeneity. This study aims to isolate the PPG components that have a strong causal relationship with BP, and leverage them to construct reliable cuffless BP estimation models. Methods: We propose a causal decomposition framework combining ensemble empirical mode decomposition (EEMD) with counterfactual inference to isolate physiologically causal components in PPG signals. First, PPG signals are adaptively decomposed into multi-scale intrinsic mode functions (IMFs) via EEMD. Then counterfactual PPG sequences are generated through sequentially excluding each IMF, and their causal relationships with BP are quantified via structure causal modeling to identify hemodynamically significant components. Subsequently, robust cuffless BP estimation models are constructed by selectively incorporating components demonstrating strong causal effects. To validate the framework, we benchmark our causality-based models against conventional approaches: pulse arrival time (PAT)-based physiological model, gradient-boosted regression tree (GBRT)-based feature model, and convolutional neural network (CNN)-based time-series model. Result: mid-frequency PPG components (IMF4-6) showed strongest BP causality, with CNN model achieving superior estimation performance over PAT and GBRT when utilizing these components. Using IMF45 components, the CNN model achieved BP estimation mean absolute errors of 5.55/3.45 mmHg (systolic/diastolic BP), improving accuracy by 26.39%/17.86% over original PPG. Conclusion: Specific PPG frequency bands exhibit physiologically meaningful causal BP relationships. Our causality-driven approach enhances both accuracy and interpretability in cuffless BP estimation. Significance:This work establishes a theoretical framework for causal feature selection in BP estimation, and a novel paradigm for physiological signal analysis through causal decomposition.
- 10 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Biomedical Engineering
- Vol. PP (00189294) , 1-11
Causal Decomposition of PPG Signals for Cuffless Blood Pressure Estimation
Objective: While photoplethysmogram (PPG) signals are physiologically linked to cardiac activity and widely used for cuffless blood pressure (BP) estimation, their correlation with BP remains limited in reliability due to inter-individual heterogeneity. This study aims to isolate the PPG components that have a strong causal relationship with BP, and leverage them to construct reliable cuffless BP estimation models. Methods: We propose a causal decomposition framework combining ensemble empirical mode decomposition (EEMD) with counterfactual inference to isolate physiologically causal components in PPG signals. First, PPG signals are adaptively decomposed into multi-scale intrinsic mode functions (IMFs) via EEMD. Then counterfactual PPG sequences are generated through sequentially excluding each IMF, and their causal relationships with BP are quantified via structure causal modeling to identify hemodynamically significant components. Subsequently, robust cuffless BP estimation models are constructed by selectively incorporating components demonstrating strong causal effects. To validate the framework, we benchmark our causality-based models against conventional approaches: pulse arrival time (PAT)-based physiological model, gradient-boosted regression tree (GBRT)-based feature model, and convolutional neural network (CNN)-based time-series model. Result: mid-frequency PPG components (IMF4-6) showed strongest BP causality, with CNN model achieving superior estimation performance over PAT and GBRT when utilizing these components. Using IMF45 components, the CNN model achieved BP estimation mean absolute errors of 5.55/3.45 mmHg (systolic/diastolic BP), improving accuracy by 26.39%/17.86% over original PPG. Conclusion: Specific PPG frequency bands exhibit physiologically meaningful causal BP relationships. Our causality-driven approach enhances both accuracy and interpretability in cuffless BP estimation. Significance:This work establishes a theoretical framework for causal feature selection in BP estimation, and a novel paradigm for physiological signal analysis through causal decomposition.
- Open Access
- 23 January 2025
- Published by JMIR Publications Inc. in JMIR Cardio
- Vol. 9 (1) , e60238
Causal Inference for Hypertension Prediction With Wearable E lectrocardiogram and P hotoplethysmogram Signals: Feasibility Study
Background: Hypertension is a leading cause of cardiovascular disease and premature death worldwide, and it puts a heavy burden on the healthcare system. Therefore, it is very important to detect and evaluate hypertension and related cardiovascular events to enable early prevention, detection, and management. Hypertension can be detected in a timely manner with cardiac signals, such as through an electrocardiogram (ECG) and photoplethysmogram (PPG) , which can be observed via wearable sensors. Most previous studies predicted hypertension from ECG and PPG signals with extracted features that are correlated with hypertension. However, correlation is sometimes unreliable and may be affected by confounding factors . Objective: The aim of this study was to investigate the feasibility of predicting the risk of hypertension by exploring features that are causally related to hypertension via causal inference methods. Additionally, we paid special attention to and verified the reliability and effectiveness of causality compared to correlation. Methods: We used a large public dataset from the Aurora Project , which was conducted by Microsoft Research. The dataset included diverse individuals who were balanced in terms of gender, age, and the condition of hypertension, with their ECG and PPG signals simultaneously acquired with wrist -worn wearable devices. We first extracted 205 features from the ECG and PPG signals, calculated 6 statistical metrics for these 205 features, and selected some valuable features out of the 205 features under each statistical metric. Then, 6 causal graphs of the selected features for each kind of statistical metric and hypertension were constructed with the equivalent greedy search algorithm. We further fused the 6 causal graphs into 1 causal graph and identified features that were causally related to hypertension from the causal graph . Finally, we used these features to detect hypertension via machine learning algorithms. Results: We validated the proposed method on 405 subjects. We identified 24 causal features that were associated with hypertension. The causal features could detect hypertension with an accuracy of 89%, precision of 92 % , and recall of 82%, which outperformed detection with correlation features (accuracy of 85%, precision of 88 % , and recall of 77%). Conclusions: The results indicated that the causal inference –based approach can potentially clarify the mechanism of hypertension detection with noninvasive signals and effectively detect hypertension. It also reveal ed that causality can be more reliable and effective than correlation for hypertension detection and other application scenarios.
- 23 January 2025
- Published by JMIR Publications Inc. in JMIR Cardio
- Vol. 9 (1) , e60238
Open AccessCausal Inference for Hypertension Prediction With Wearable E lectrocardiogram and P hotoplethysmogram Signals: Feasibility Study
Background: Hypertension is a leading cause of cardiovascular disease and premature death worldwide, and it puts a heavy burden on the healthcare system. Therefore, it is very important to detect and evaluate hypertension and related cardiovascular events to enable early prevention, detection, and management. Hypertension can be detected in a timely manner with cardiac signals, such as through an electrocardiogram (ECG) and photoplethysmogram (PPG) , which can be observed via wearable sensors. Most previous studies predicted hypertension from ECG and PPG signals with extracted features that are correlated with hypertension. However, correlation is sometimes unreliable and may be affected by confounding factors . Objective: The aim of this study was to investigate the feasibility of predicting the risk of hypertension by exploring features that are causally related to hypertension via causal inference methods. Additionally, we paid special attention to and verified the reliability and effectiveness of causality compared to correlation. Methods: We used a large public dataset from the Aurora Project , which was conducted by Microsoft Research. The dataset included diverse individuals who were balanced in terms of gender, age, and the condition of hypertension, with their ECG and PPG signals simultaneously acquired with wrist -worn wearable devices. We first extracted 205 features from the ECG and PPG signals, calculated 6 statistical metrics for these 205 features, and selected some valuable features out of the 205 features under each statistical metric. Then, 6 causal graphs of the selected features for each kind of statistical metric and hypertension were constructed with the equivalent greedy search algorithm. We further fused the 6 causal graphs into 1 causal graph and identified features that were causally related to hypertension from the causal graph . Finally, we used these features to detect hypertension via machine learning algorithms. Results: We validated the proposed method on 405 subjects. We identified 24 causal features that were associated with hypertension. The causal features could detect hypertension with an accuracy of 89%, precision of 92 % , and recall of 82%, which outperformed detection with correlation features (accuracy of 85%, precision of 88 % , and recall of 77%). Conclusions: The results indicated that the causal inference –based approach can potentially clarify the mechanism of hypertension detection with noninvasive signals and effectively detect hypertension. It also reveal ed that causality can be more reliable and effective than correlation for hypertension detection and other application scenarios.
- 1 October 2024
- Published by IOP Publishing in Physiological Measurement
- Vol. 45 (10) , 105007
The influence of heart rate on the relationship between pulse transit time and systolic blood pressure
Objective. Pulse transit time (PTT) is a popular indicator of blood pressure (BP) changes. However, the relationship between PTT and BP is somehow individual dependent, resulting in the inaccuracy of PTT-based BP estimation. Confounding factors, e.g. heart rate (HR), of PTT and BP could be the primary cause. In this study we attempt to explore the impact of HR as a window to look at the influence of confounding factors on the relationship between PTT and BP. Approach. We investigated the relationship between PTT and systolic BP (SBP) at different HR levels by introducing the heterogeneous treatment effects (HTE) as a quantitative indicator. Compared to the average HR calculated using traditional indicators (e.g. regression coefficient, correlation coefficient), the HTE calculation method can compute the relationship between PTT and SBP at different HR levels, and reduce the influence of confounding factors. Main results. We analyzed the HTE of PTT and SBP of 47 subjects who are resting healthy young people with varying levels of HR. The results showed that the strength of the HTE of PTT and SBP varied with HR, indicating that the strength of the causal relationship between PTT and SBP is subject to HR levels. Whereas the correlation between SBP and PTT was individual dependent; either the strength or the direction of the correlation can vary with HR. We further investigated the group in which PTT and SBP exhibited a negative correlation, and found that about 50% of the subjects showed enhanced strength of HTE with an increase in HR and the remaining showed the opposite. Significance. HR has an impact on the relationship between PTT and SBP. This means that HR needs to be considered when PTT is used as an indicator of SBP.
- 1 October 2024
- Published by IOP Publishing in Physiological Measurement
- Vol. 45 (10) , 105007
The influence of heart rate on the relationship between pulse transit time and systolic blood pressure
Objective. Pulse transit time (PTT) is a popular indicator of blood pressure (BP) changes. However, the relationship between PTT and BP is somehow individual dependent, resulting in the inaccuracy of PTT-based BP estimation. Confounding factors, e.g. heart rate (HR), of PTT and BP could be the primary cause. In this study we attempt to explore the impact of HR as a window to look at the influence of confounding factors on the relationship between PTT and BP. Approach. We investigated the relationship between PTT and systolic BP (SBP) at different HR levels by introducing the heterogeneous treatment effects (HTE) as a quantitative indicator. Compared to the average HR calculated using traditional indicators (e.g. regression coefficient, correlation coefficient), the HTE calculation method can compute the relationship between PTT and SBP at different HR levels, and reduce the influence of confounding factors. Main results. We analyzed the HTE of PTT and SBP of 47 subjects who are resting healthy young people with varying levels of HR. The results showed that the strength of the HTE of PTT and SBP varied with HR, indicating that the strength of the causal relationship between PTT and SBP is subject to HR levels. Whereas the correlation between SBP and PTT was individual dependent; either the strength or the direction of the correlation can vary with HR. We further investigated the group in which PTT and SBP exhibited a negative correlation, and found that about 50% of the subjects showed enhanced strength of HTE with an increase in HR and the remaining showed the opposite. Significance. HR has an impact on the relationship between PTT and SBP. This means that HR needs to be considered when PTT is used as an indicator of SBP.
- Open Access
- 21 February 2024
- Published by Elsevier in Biomedical Signal Processing and Control
- Vol. 92
Uncertainty quantification of cuffless blood pressure estimation based on parameterized model evidential ensemble learning
- 21 February 2024
- Published by Elsevier in Biomedical Signal Processing and Control
- Vol. 92
Open AccessUncertainty quantification of cuffless blood pressure estimation based on parameterized model evidential ensemble learning