Meina Li
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Meina Li has a total of 13 co-authors. Here are the top co-authors by number of shared publications:
Gengjia Zhang
2 shared publications- Jaehyo Jung
Department of Bigdata and Artificial Intelligence, Eulji University, Seongnam-si, South Korea
2 shared publications Zhiqiang Liu
2 shared publications- Gengjia Zhang
AI Healthcare Research Center, Department of IT Fusion Technology, Chosun University, Gwangju 61452, Korea
1 shared publication - Hongyan Liu
College of Instrumentation and Electrical Engineering, Jilin University, Changchun, China
1 shared publication Jaehyo Jung
1 shared publication
Latest Publications
- 24 August 2026
- Published by Elsevier in Knowledge-Based Systems
- Vol. 352
Personalized calibration for depression detection using knowledge-distilled facial video models
- 24 August 2026
- Published by Elsevier in Knowledge-Based Systems
- Vol. 352
Personalized calibration for depression detection using knowledge-distilled facial video models
- Open Access
- 18 February 2026
- Published by Elsevier in SSRN Electronic Journal
Continual Personalized Calibration for Depression Detection Using Knowledge-Distilled Facial Video Models
Personalized depression assessment requires models that can adapt to individual behavioral patterns while preserving data privacy. However, population-trained deep-learning models often fail to generalize to subject-specific facial dynamics. This study introduces a knowledge-based continual-learning framework that addresses this limitation by formulating personalization as a structured knowledge transfer and adaptation process. A teacher model, trained on large-scale population data, provides global behavioral knowledge, whereas a lightweight student model incrementally acquires user-specific cues from private calibration data. To strengthen the representational capacity of the teacher, we develop a multi-level fusion dual-stream network (MF-DSN) that integrates complementary spatiotemporal features through multi-scale fusion and deep supervision for robust depression-severity regression. Knowledge distillation enables the student model to retain global knowledge while specializing toward individual facial patterns under strict computational and privacy constraints. Experiments conducted on the AVEC2014, the DAIC-WOZ, and a real-world clinical datasets demonstrate the effectiveness of the proposed approach. The MF-DSN achieves competitive baseline performance (e.g., a mean absolute error of 2.95 for PHQ-9 prediction in the clinical data), and the personalized student model further reduces individual-level prediction error by up to 30% after calibration; this highlights its ability to correct subject-specific prediction bias. These results confirm that knowledge-based adaptive learning substantially enhances the accuracy, stability, and practical effectiveness of video-based depression-monitoring systems. The proposed framework offers a scalable and privacy-preserving solution for long-term mental-health assessment and represents a promising direction for knowledge-based personalized affective computing.
- 18 February 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessContinual Personalized Calibration for Depression Detection Using Knowledge-Distilled Facial Video Models
Personalized depression assessment requires models that can adapt to individual behavioral patterns while preserving data privacy. However, population-trained deep-learning models often fail to generalize to subject-specific facial dynamics. This study introduces a knowledge-based continual-learning framework that addresses this limitation by formulating personalization as a structured knowledge transfer and adaptation process. A teacher model, trained on large-scale population data, provides global behavioral knowledge, whereas a lightweight student model incrementally acquires user-specific cues from private calibration data. To strengthen the representational capacity of the teacher, we develop a multi-level fusion dual-stream network (MF-DSN) that integrates complementary spatiotemporal features through multi-scale fusion and deep supervision for robust depression-severity regression. Knowledge distillation enables the student model to retain global knowledge while specializing toward individual facial patterns under strict computational and privacy constraints. Experiments conducted on the AVEC2014, the DAIC-WOZ, and a real-world clinical datasets demonstrate the effectiveness of the proposed approach. The MF-DSN achieves competitive baseline performance (e.g., a mean absolute error of 2.95 for PHQ-9 prediction in the clinical data), and the personalized student model further reduces individual-level prediction error by up to 30% after calibration; this highlights its ability to correct subject-specific prediction bias. These results confirm that knowledge-based adaptive learning substantially enhances the accuracy, stability, and practical effectiveness of video-based depression-monitoring systems. The proposed framework offers a scalable and privacy-preserving solution for long-term mental-health assessment and represents a promising direction for knowledge-based personalized affective computing.
- Open Access
- 23 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 167315-167323
Evaluating Time-Series Deep Learning Models for Accurate and Efficient Reconstruction of Clinical 12-Lead ECG Signals
With the transformation of healthcare systems, the aging population, and the impact of the COVID-19 pandemic, interest in the healthcare industry has significantly increased. In particular, demand has grown for personal electrocardiogram (ECG) monitoring using wearable devices, such as smartwatches. However, these devices usually only record single-lead ECG signals, which are not as accurate for diagnosing cardiac conditions as the standard 12-lead ECGs used in clinical settings. Deep learning models have been actively explored to reconstruct standard 12-lead ECG signals from several leads to overcome this limitation. Nevertheless, comprehensive comparative studies on time-series-based deep learning models capable of generating clinically equivalent ECG signals are lacking. In this study, we designed and evaluated three recurrent neural network (RNN) architectures LSTM, Bi-LSTM, and GRU to reconstruct 12-lead ECG signals from a limited subset of leads and compared their reconstruction performance. Each model consists of three RNN layers, batch normalization, and dropout layers. Experimental results on the PTB-XL dataset show that the GRU model achieves the best reconstruction performance, with a PCC of 0.911, MAE of 0.033 and RMSE of 0.065, outperforming Bi-LSTM (PCC: 0.893, MAE: 0.035, RMSE: 0.072) and LSTM (PCC: 0.819, MAE: 0.039, RMSE: 0.088). This study systematically compares the performance of RNN-based models for ECG signal reconstruction and identifies GRU as the most effective architecture for reconstructing standard 12-lead ECGs from single-lead inputs. Due to its relatively lightweight structure, the proposed model shows strong potential for real-time implementation on wearable devices and has valuable implications for mobile healthcare and remote diagnostic systems.
- 23 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 167315-167323
Open AccessEvaluating Time-Series Deep Learning Models for Accurate and Efficient Reconstruction of Clinical 12-Lead ECG Signals
With the transformation of healthcare systems, the aging population, and the impact of the COVID-19 pandemic, interest in the healthcare industry has significantly increased. In particular, demand has grown for personal electrocardiogram (ECG) monitoring using wearable devices, such as smartwatches. However, these devices usually only record single-lead ECG signals, which are not as accurate for diagnosing cardiac conditions as the standard 12-lead ECGs used in clinical settings. Deep learning models have been actively explored to reconstruct standard 12-lead ECG signals from several leads to overcome this limitation. Nevertheless, comprehensive comparative studies on time-series-based deep learning models capable of generating clinically equivalent ECG signals are lacking. In this study, we designed and evaluated three recurrent neural network (RNN) architectures LSTM, Bi-LSTM, and GRU to reconstruct 12-lead ECG signals from a limited subset of leads and compared their reconstruction performance. Each model consists of three RNN layers, batch normalization, and dropout layers. Experimental results on the PTB-XL dataset show that the GRU model achieves the best reconstruction performance, with a PCC of 0.911, MAE of 0.033 and RMSE of 0.065, outperforming Bi-LSTM (PCC: 0.893, MAE: 0.035, RMSE: 0.072) and LSTM (PCC: 0.819, MAE: 0.039, RMSE: 0.088). This study systematically compares the performance of RNN-based models for ECG signal reconstruction and identifies GRU as the most effective architecture for reconstructing standard 12-lead ECGs from single-lead inputs. Due to its relatively lightweight structure, the proposed model shows strong potential for real-time implementation on wearable devices and has valuable implications for mobile healthcare and remote diagnostic systems.
- 1 September 2023
- Published by AIP Publishing in Review of Scientific Instruments
- Vol. 94 (9)
A vital sign signal noise suppression method for wearable piezoelectric devices
This paper tackles the problem of noise suppression during vital sign signal monitoring. Physiological signal monitoring is a significant and promising medical monitoring method, and wearable medical monitoring devices based on piezoelectric polymer sensors are a trending way for their advantages of being flexible in the shape, portable to use, and comfortable to wear. However, this raises the question that the measured signal contains much more noise components. To avoid the following shortcoming of low signal to noise ratio (SNR), a noise suppression method based on improved wavelet threshold and empirical mode decomposition combined with singular value decomposition (SVD) screening the intrinsic mode function (IMF) components is proposed. A wavelet transform is first used under the combination of hard and soft thresholds to focus the target range in the low-frequency region where the energy of the physiological signal is concentrated. Then, a complete ensemble empirical mode decomposition is used to decompose the signal effectively, which can resist the influence of random noises. Meanwhile, a SVD decomposition procedure was used to filter out the lower correlated IMF components to retain the validity of the original signal. We verified the effectiveness of the proposed method through simulated and measured experiments as well as the advantages and disadvantages of the algorithm compared with other physiological signal denoising algorithms through SNR filtering results, power spectrum distribution, and other perspectives. The results proved that the proposed method could effectively remove more detailed noise and improve the SNR of the signal efficiently, which is more conducive to the demand for auxiliary medical diagnosis in the future.
- 1 September 2023
- Published by AIP Publishing in Review of Scientific Instruments
- Vol. 94 (9)
A vital sign signal noise suppression method for wearable piezoelectric devices
This paper tackles the problem of noise suppression during vital sign signal monitoring. Physiological signal monitoring is a significant and promising medical monitoring method, and wearable medical monitoring devices based on piezoelectric polymer sensors are a trending way for their advantages of being flexible in the shape, portable to use, and comfortable to wear. However, this raises the question that the measured signal contains much more noise components. To avoid the following shortcoming of low signal to noise ratio (SNR), a noise suppression method based on improved wavelet threshold and empirical mode decomposition combined with singular value decomposition (SVD) screening the intrinsic mode function (IMF) components is proposed. A wavelet transform is first used under the combination of hard and soft thresholds to focus the target range in the low-frequency region where the energy of the physiological signal is concentrated. Then, a complete ensemble empirical mode decomposition is used to decompose the signal effectively, which can resist the influence of random noises. Meanwhile, a SVD decomposition procedure was used to filter out the lower correlated IMF components to retain the validity of the original signal. We verified the effectiveness of the proposed method through simulated and measured experiments as well as the advantages and disadvantages of the algorithm compared with other physiological signal denoising algorithms through SNR filtering results, power spectrum distribution, and other perspectives. The results proved that the proposed method could effectively remove more detailed noise and improve the SNR of the signal efficiently, which is more conducive to the demand for auxiliary medical diagnosis in the future.