Objectives To establish a robust and clinically applicable approach for integrating heterogeneous multisource biomedical data, particularly continuous glucose monitoring (CGM) profiles and structured electronic health records (EHRs), in order to enhance the diagnostic accuracy and clinical utility of diabetic retinopathy (DR) detection. Methods This study proposed a deep hierarchical attention network (DHAN) for multisource biomedical data fusion. First, to address the heterogeneous forms of different data sources, two specific subencoders were designed, a hybrid architecture for time-series CGM sensors and a structured encoder for EHRs. Second, an entity-embedding mechanism was added to the EHR subencoder to fuse heterogeneous feature types within EHRs. Finally, a deep hierarchical attention mechanism was proposed to dynamically capture inner-source saliency and inter-source correlations. Results Using the dataset provided by Shanghai Sixth People's Hospital, 559 patients were included, comprising 157 with DR and 402 without. DHAN achieved the best performance across multiple experiments, with a diagnostic accuracy of 0.89. Its comprehensive performance, including an F 1-score of 0.80 and a G -mean of 0.89, further demonstrates its robustness. Conclusions The results indicate that DHAN is a viable approach for diagnosing DR in patients with type 2 diabetes. By effectively fusing multisource heterogeneous data, DHAN can be embedded within CGM sensors to enable remote concurrent diagnosis of DR. Moreover, it provides a generalizable paradigm for multisensor systems requiring fusion of data from multiple sources.

With the gradual rise and application of wearable devices, automated arrhythmia detection has been successfully applied to the long-term monitoring of individual hearts. However, electrocardiogram (ECG) signals vary from person to person. Different individuals and environments also change in ECG signals. General detection methods cannot address the problem of model generalization due to individual specificity. In this paper, a novel customized framework with cross-group interaction deep learning was proposed for the first time to address the problem of individual specificity in ECG signals. Firstly, unsupervised clustering was performed on the entire population to generate different subgroups. Signals with similar waveforms were grouped into a subgroup. We used the Convolutional Neural Networks, Bidirectional Long Short-Term Memory, and Attention mechanism to establish different classification frameworks to extract deep learning features. Different groups have different convolution layers and parameters. Afterward, we designed the cross-group interaction module to complement the advantageous information between subgroups. In particular, in subgroup II we proposed the cross-attention mechanism for feature concatenation to complement the information focused on in the two groups. This method was implemented on the MIT-BIH database to classify five different classes of signals. The classification results achieved an average accuracy of 99.01%, precision of 97.68%, sensitivity of 96.00%, and specificity of 99.35%. The results show that the method proposed can eliminate the problems caused by individual variability at the data source, improve the model generalization ability, and achieve better recognition results.

Effective clustering of medical data is essential for the progress of precision medicine and data-driven healthcare solutions. Modern medical datasets are increasingly complex, often integrating multi-source data with both categorical and numerical features of varying importance. However, existing algorithms like the k-prototype could not handle ordered features and fail to account for the differing importance of multi-source data in the clustering process. This study introduces a proposed k-prototype algorithm that incorporates weights to cluster medical data from multi-sources. In particular, we initially present an enhanced cost function to consistently evaluate the distance between ordered, categorical, and numerical features. Then, weights are allocated to each feature using a multi-objective optimization algorithm, and the optimal weighting operator is identified through an optimization search within the evolutionary computation framework. Experiments on simulated and real medical datasets show the algorithm's superior performance. It offers a robust solution for multi-source medical data analysis. This improvement enhances insights and supports better decision-making in medical data analysis.

For patients with type 2 diabetes (T2D), accurate prediction of blood glucose variations is essential for maintaining glycemic control, decreasing the occurrence of hypoglycemic and hyperglycemic events, and preventing diabetes complications. However, this is difficult to achieve due to insufficient early glucose data and the complexity of glucose dynamics. Additionally, the high variability among individuals poses challenges for data transfer between patients. In this work, a clustering-based domain adaptation method is proposed for personalized glucose prediction of T2D with insufficient data. Firstly, the multi-level clustering method is used to subtype the heterogeneous group of patients with T2D into multiple homogenous subgroups to deal with the high inter-individual variability. Then, a domain adaptation prediction network is designed to overcome the challenges caused by insufficient historical data of the target patient through cross-patient knowledge transfer and obtain a personalized deep prediction model suitable for the target patient. The effectiveness of the proposed method was evaluated in a clinical dataset containing continuous glucose monitoring (CGM) measurement records from 908 patients with T2D, each with only a small amount of data. The 30-minute prediction horizon achieved an average root mean square error of 14.96 mg/dL, with over 94 % of predictions clinically accurate. In addition, we evaluated the long-term transferability of the proposed method on the publicly available ShanghaiT2DM Dataset and compared it with the state-of-the-art (SOTA) methods. The results demonstrate that the proposed personalized method can achieve accurate glucose prediction for patients with T2D, even with only one day of historical CGM records available.

The current diagnosis of diabetic retinopathy is based on fundus images and clinical experience. However, considering the ineffectiveness and non-portability of medical devices, we aimed to develop a diagnostic model for diabetic retinopathy based on glucose series data from the wearable continuous glucose monitoring system. Therefore, this study developed a novel method, i.e., double deep latent autoencoder, for exploring glycemic variability influence from multi-day glucose data for diabetic retinopathy. Specifically, the model proposed in this research could encode continuous glucose sensor data with non-continuous and variable length via the integration of a data reorganization module and a novel encoding module with fragmented-missing-wise objective function. Additionally, the model implements a double deep autoencoder, which integrated convolutional neural network, long short-term memory, to jointly capturing the inter-day and intra-day glucose latent features from glucose series. The effectiveness of the proposed model is evaluated through a cross-validation method to clinical datasets of 765 type 2 diabetes patients. The proposed method achieves the highest accuracy value (0.89), precision value (0.88), and F1 score (0.73). The results suggest that our model can be used to remotely diagnose and screen for diabetic retinopathy by learning potential features of glucose series data collected by wearable continuous glucose monitoring systems.

AIM:The wealth of data generated by continuous glucose monitoring (CGM) provides new opportunities for revealing heterogeneities in patients with type 2 diabetes mellitus (T2DM). We aimed to develop a method using CGM data to discover T2DM subtypes and investigate their relationship with clinical phenotypes and microvascular complications. METHODS:The data from 3119 patients with T2DM who wore blinded CGM at an academic medical centre was collected, and a glucose symbolic pattern (GSP) metric was created that combined knowledge-based temporal abstraction with numerical vectorization. The k-means clustering was applied to GSP to obtain subgroups of patients with T2DM. Clinical characteristics and the presence of diabetic retinopathy and albuminuria were compared among the subgroups. The findings were validated in an independent population comprising 773 patients with T2DM. RESULTS:By using GSP, four subgroups were identified with distinct features in CGM profiles and parameters. Moreover, the clustered subgroups differed significantly in clinical phenotypes, including indices of pancreatic β-cell function and insulin resistance (all p < .001). After adjusting for confounders, group C (the most insulin resistant) had a significantly higher risk of albuminuria (odds ratio = 1.24, 95% confidence interval: 1.03-1.39) relative to group D, which had the best glucose control. These findings were confirmed in the validation set. CONCLUSION:Subtyping patients with T2DM using CGM data may help identify high-risk patients for microvascular complications and provide insights into the underlying pathophysiology. This method may help refine clinically meaningful stratification of patients with T2DM and inform personalized diabetes care.

Continuous glucose monitoring (CGM) data analysis will provide a new perspective to analyze factors related to diabetic retinopathy (DR). However, the problem of visualizing CGM data and automatically predicting the incidence of DR from CGM is still controversial. Here, we explored the feasibility of using CGM profiles to predict DR in type 2 diabetes (T2D) by deep learning approach. This study fused deep learning with a regularized nomogram to construct a novel deep learning nomogram from CGM profiles to identify patients at high risk of DR. Specifically, a deep learning network was employed to mine the nonlinear relationship between CGM profiles and DR. Moreover, a novel nomogram combining CGM deep factors with basic information was established to score the patients’ DR risk. This dataset consists of 788 patients belonging to two cohorts: 494 in the training cohort and 294 in the testing cohort. The area under the curve (AUC) values of our deep learning nomogram were 0.82 and 0.80 in the training cohort and testing cohort, respectively. By incorporating basic clinical factors, the deep learning nomogram achieved an AUC of 0.86 in the training cohort and 0.85 in the testing cohort. The calibration plot and decision curve showed that the deep learning nomogram had the potential for clinical application. This analysis method of CGM profiles can be extended to other diabetic complications by further investigation.

The accurate, timely, and personalized prediction for future blood glucose (BG) levels is undoubtedly needed for further advancement of diabetes management technologies. Human inherent circadian rhythm and regular lifestyle resulting in similarity of daily glycemic dynamics play a positive role in the prediction of blood glucose. Inspired by the iterative learning control (ILC) method in the field of automatic control, a 2-dimensional (2-D) model framework is constructed to predict the future blood glucose levels by taking both the short-range information within a day (intra-day) and long-range information between days (inter-day) into account. In this framework, the radial basis function neural network was applied to capture nonlinear relationships in glycemic metabolism, that is, short-range temporal dependence and long-range contemporaneous dependence on previous days. We build models for each patient, and the models were tested on the in silico datasets at various prediction horizons (PHs). The learning model developed in the 2-D framework successfully increases the accuracy and reduces the delay of predictions. This modeling framework provides a new point of view for BG level prediction and contributes to the development of personalized glucose management, such as hypoglycemia warning and glycemic control.

Most information used to evaluate diabetic statuses is collected at a special time-point, such as taking fasting plasma glucose test and providing a limited view of individual’s health and disease risk. As a new parameter for continuously evaluating personal clinical statuses, the newly developed technique “continuous glucose monitoring” (CGM) can characterize glucose dynamics. By calculating the complexity of glucose time series index (CGI) with refined composite multi-scale entropy analysis of the CGM data, the study showed for the first time that the complexity of glucose time series in subjects decreased gradually from normal glucose tolerance to impaired glucose regulation and then to type 2 diabetes ( P for trend < 0.01). Furthermore, CGI was significantly associated with various parameters such as insulin sensitivity/secretion (all P < 0.01), and multiple linear stepwise regression showed that the disposition index, which reflects β-cell function after adjusting for insulin sensitivity, was the only independent factor correlated with CGI ( P < 0.01). Our findings indicate that the CGI derived from the CGM data may serve as a novel marker to evaluate glucose homeostasis.

Lesion segmentation of ultrasound medical images based on deep learning techniques is a widely used method for diagnosing diseases. Although there is a large amount of ultrasound image data in medical centers and other places, labeled ultrasound datasets are a scarce resource, and it is likely that no datasets are available for new tissues/organs. Transfer learning provides the possibility to solve this problem, but there are too many features in natural images that are not related to the target domain. As a source domain, redundant features that are not conducive to the task will be extracted. Migration between ultrasound images can avoid this problem, but there are few types of public datasets, and it is difficult to find sufficiently similar source domains. Compared with natural images, ultrasound images have less information, and there are fewer transferable features between different ultrasound images, which may cause negative transfer. To this end, a multi-source adversarial transfer learning network for ultrasound image segmentation is proposed. Specifically, to address the lack of annotations, the idea of adversarial transfer learning is used to adaptively extract common features between a certain pair of source and target domains, which provides the possibility to utilize unlabeled ultrasound data. To alleviate the lack of knowledge in a single source domain, multi-source transfer learning is adopted to fuse knowledge from multiple source domains. In order to ensure the effectiveness of the fusion and maximize the use of precious data, a multi-source domain independent strategy is also proposed to improve the estimation of the target domain data distribution, which further increases the learning ability of the multi-source adversarial migration learning network in multiple domains.

Heart diseases represent a serious medical condition that can be fatal. Therefore, it is critical to investigate the measures of its early prevention. The Mel-scale frequency cepstral coefficients (MFCC) feature has been widely used in the early diagnosis of heart abnormity and achieved promising results. During feature extraction, the Mel-scale triangular overlapping filter set is applied, which makes the frequency response more in line with the human auditory property. However, the frequency of the heart sound signals has no specific relationship with the human auditory system, which may not be suitable for processing of heart sound signals. To overcome this issue and obtain a more objective feature that can better adapt to practical use, in this work, we propose an equal scale frequency cepstral coefficients (EFCC) feature based on replacing the Mel-scale filter set with a set of equally spaced triangular overlapping filters. We further designed classifiers combining convolutional neural network (CNN), recurrent neural network (RNN) and random forest (RF) layers, which can extract both the spatial and temporal information of the input features. We evaluated the proposed algorithm on our database and the PhysioNet Computational Cardiology (CinC) 2016 Challenge Database. Results from ten-fold cross-validation reveal that the EFCC-based features show considerably better performance and robustness than the MFCC-based features on the task of classifying heart sounds from novel patients. Our algorithm can be further used in wearable medical devices to monitor the heart status of patients in real time with high precision, which is of great clinical importance.

Cluster analysis discovers natural structures from different perspectives of data objects and has become an effective method in data mining. The emergence of semi-supervised clustering techniques has improved the performance of unsupervised clustering algorithms. Clustering with guidance information is a clustering method variant that uses pairwise constraints based on background knowledge. This method increases the interpretability of the results through a knowledge-guided perspective but simultaneously suffers from the problem of constraint conflict. This paper designs knowledge augmentation-based soft constraints as a new pairwise constraint representation and proposes a Soft Constraints Kmeans (SCop-Kmeans) method to resolve constraint conflicts. By describing constraint knowledge from multiple perspectives, the association strength of pairwise constraints is calculated to obtain the assignment basis of objects. SCop-Kmeans can solve the sample allocation conflict problem caused by the contradiction between different constraints and improve the clustering stability. Finally, experiments are performed using UCI public standard datasets. The proposed method further improves the accuracy of clustering and performs well in experiments with different numbers of constraints, which shows that the proposed method has advantages in using constraint information to guide clustering.

Background and objective: For patients with type 2 diabetes (T2D), accurate prediction of blood glucose variations is essential for maintaining glycemic control, decreasing the occurrence of hypo/hyperglycemic events, and preventing diabetes complications. However, this is difficult to achieve due to high inter-individual variability, insufficient glucose data, and the complexity of glucose dynamics. To address these issues, inter-individual variability can be overcome by using domain adaptation to leverage multiple patient data for personalized deep modeling of glucose dynamics.Methods: In this work, a clustering-based domain adaptation method is proposed for personalized glucose prediction of T2D with insufficient data. Firstly, the multi-level clustering method is used to subtyping the heterogeneous group of patients with T2D into four homogenous subgroups to deal with the high inter-individual variability. Then, a domain adaptation prediction network is designed to overcome the challenges caused by insufficient historical data of the target patient through cross-patient knowledge transfer and obtain a personalized deep prediction model suitable for the target patient.Results: The effectiveness of the proposed method was evaluated in a clinical dataset containing continuous glucose monitoring (CGM) measurement records from 908 patients with T2D, each with only a small amount of data. The average prediction results showed that the root mean square error of the 30-minute prediction horizon was 14.961 mg/dL, and more than 94% of the predicted values were ‘clinically accurate’.Conclusions: The proposed personalized method can achieve an accurate glucose prediction for patients with T2D even if the target patient has only one-day historical CGM records.

Phonocardiogram (PCG) auscultation is one of the most commonly used methods for coronary artery disease (CAD) detection. However, its detection accuracy is influenced by significant interpersonal variations. In this paper, a novel customized framework was proposed for the first time to address the individual specificity in PCG signals. To eliminate individual differences at source, a clustering method based on age information and PCG time-frequency features was developed to partition the subjects into subgroups. Then we designed different classification models for different subgroups to achieve an individually tailored diagnostic strategy, thus further attenuating the effect of individual specificity. In the classification stage, feature fusion was employed to overcome the shortage of single features. Mel-frequency cepstral coefficients and PCG signal fragments were sent to the corresponding convolutional neural networks to obtain two-dimensional time-frequency features and one-dimensional spatial-temporal features. Ultimately, the two multimodal features were fused and fed into a random forest for classification. The experiments demonstrated that the customized framework can effectively solve the problem of individual specificity in PCG detection with an average accuracy of 96.05%, which is an improvement in the accuracy by up to 6.51% over the general method. A comparison with existing research indicates that the proposed method is a robust and effective noninvasive technique for CAD detection, and it is a feasible solution for the problem of individual specificity in PCG classification. In addition, the framework can be extended to other similar biomedical signal applications.

目的·建立小鼠持续葡萄糖监测(continuous glucose monitoring,CGM)技术,并对其血糖时间序列进行多尺度熵(multiscale entropy,MSE)分析.方法·选取饮食诱导肥胖型(diet-induced obesity,DIO)小鼠(n=3,DIO组)及对照组小鼠(n=3)为研究对象,利用全植入式血糖遥测系统分别收集2组小鼠的血糖及体温数据,取术后第10-14日数据进行分析,并统计系统的记录时间.利用MATLABR2019b软件对2组小鼠术后第11?17日的血糖时间序列进行MSE分析,计算每个时间尺度上对应的熵值.结果·成功建立了以全植入式血糖遥测系统为基础的小鼠CGM技术.6只小鼠的平均记录时间为(27.3±9.3)d,共获得232 887个血糖数值.DIO组小鼠平均血糖水平为(7.04±0.71)mmol/L,平均体温为(33.34±0.18)°C.与对照组相比,DIO组小鼠血糖时间序列复杂度较低,但差异无统计学意义.结论.成功建立了小鼠CGM技术;MSE分析发现,DIO型小鼠血糖时间序列复杂度降低,可能县S早期糖代谢异常的表现之一?

Introduction Mining knowledge from continuous glucose monitoring (CGM) data to classify highly heterogeneous patients with type 2 diabetes according to their characteristics remains unaddressed. A refined clustering method that retrieves hidden information from CGM data could provide a viable method to identify patients with different degrees of dysglycemia and clinical phenotypes. Research design and methods From Shanghai Jiao Tong University Affiliated Sixth People’s Hospital, we selected 908 patients with type 2 diabetes (18–83 years) who wore blinded CGM sensors (iPro2, Medtronic, California, USA). Participants were clustered based on CGM data during a 24-hour period by our method. The first level extracted the knowledge-based and statistics-based features to describe CGM signals from multiple perspectives. The Fisher score and variables cluster analysis were applied to fuse features into low dimensions at the second level. The third level divided subjects into subgroups with different clinical phenotypes. The four subgroups of patients were determined by clinical phenotypes. Results Four subgroups of patients with type 2 diabetes with significantly different statistical features and clinical phenotypes were identified by our method. In particular, individuals in cluster 1 were characterized by the lowest glucose level factor and glucose fluctuation factor, and the highest negative glucose factor and C peptide index. By contrast, cluster 2 had the highest glucose level factor and the lowest C peptide index. Cluster 4 was characterized by the greatest degree of glucose fluctuation factor, was the most insulin-sensitive, and had the lowest insulin resistance. Cluster 3 ranked in the middle concerning the CGM-derived metrics and clinical phenotypes compared with those of the other three groups. Conclusion A novel multilevel clustering approach for knowledge mining from CGM data in type 2 diabetes is presented. The results demonstrate that subgroups are adequately distinguished with notable statistical and clinical differences.

Blood glucose (BG) prediction plays an important role in daily BG control. Accurate prediction of short-term glucose concentration can provide early warning for hyperglycemia and hypoglycemia events. This paper proposed a novel framework that combined an online prediction model with a residual compensation network. The autoregressive moving average (ARMA) model was used for online blood glucose prediction and the neural network was applied for compensation of prediction error. The advantages of this combined framework are: (1) the online ARMA model is efficient and robust to capture time-varying glucose dynamics, (2) the residual compensation network is capable to estimate errors from the online prediction model. The performance of this method was evaluated by the root mean squared error (RMSE) and the mean absolute error (MAE) in the dataset of OhioT1DM.The results were shown in detail that the mean values of the best RMSE of six patients at 30-min and 60-min horizon were 20.03 and 34.89 respectively, and the best MAE at 30-min and 60-min horizon were 14.52 and 24.61. Compared with the ARMA model, the combined predictor with a residual compensation network shows better prediction accuracy. Thus, we concluded that the proposed framework was an available approach for online blood glucose level prediction (BGLP).