OBJECTIVE: To develop a hybrid neural network-based blood donation prediction method, via this predictive model, we can obtain the best estimate of whole blood in Beijing Tongzhou District Central Blood Station and help managers smoothly solve the allocation problem under fluctuating hospital demand and limited resources. METHOD: Inspired by the practical problems faced by blood stations providing transfusion services to several hospitals, a hybrid model based on a time-series prediction method and neural network, SARIMAX-TCN-LSTM is proposed for the prediction of daily whole blood donations. The experiment was performed at the central blood station in Tongzhou district, where we used whole blood donations from January 1, 2015, to November 14, 2021, as the subject, supplemented by meteorological and epidemic factors affecting blood donation, to predict daily blood donations for the next two weeks. RESULT: The hybrid model significantly outperformed the traditional time series forecasting method on multiple regression metrics, with twice as effective fitting as the baseline and a 33% reduction in Root Mean Squared Error (RMSE). Results indicate that the proposed model can improve the prediction accuracy of daily blood donations, and the co-validity of the structure was evidenced in an ablation experiment. CONCLUSION: Development and evaluation of a hybrid neural network-based model structure improve the prediction of daily blood donations. This intelligent forecasting method can help managers to overcome the challenges of sudden blood demand and contribute to the optimization of resource allocation tasks.

Automatic radiology report generation has the potential to alert inexperienced radiologists to misdiagnoses or missed diagnoses and improve healthcare delivery efficiency by reducing the documentation workload of radiologists. Motivated by the continuous development of automatic image captioning, more and more deep learning methods have been proposed for automatic radiology report generation. However, the visual and textual data bias problem still face many challenges in the medical domain. Additionally, do not integrate medical knowledge, ignoring the mutual influences between medical findings, and abundant unlabeled medical images influence the accuracy of generating report. In this paper, we propose a Medical Knowledge with Contrastive Learning model (MKCL) to enhance radiology report generation. The proposed model MKCL uses IU Medical Knowledge Graph (IU-MKG) to mine the relationship among medical findings and improve the accuracy of identifying positive diseases findings from radiologic medical images. In particular, we design Knowledge Enhanced Attention (KEA), which integrates the IU-MKG and the extracted chest radiological visual features to alleviate textual data bias. Meanwhile, this paper leverages supervised contrastive learning to relieve radiographic medical images which have not been labeled, and identify abnormalities from images. Experimental results on the public dataset IU X-ray show that our proposed model MKCL outperforms other state-of-the-art report generation methods. Ablation studies also demonstrate that IU medical knowledge graph module and supervised contrastive learning module enhance the ability of the model to detect the abnormal parts and accurately describe the abnormal findings. The source code is available at : https://github.com/Eleanorhxd/MKCL.

OBJECTIVE: Blood glucose measurements in the intensive care unit (ICU) are typically made at irregular intervals. This presents a challenge in choice of forecasting model. This article gives an overview of continuous time autoregressive recurrent neural networks (CTRNNs) and evaluates how they compare to autoregressive gradient boosted trees (GBT) in forecasting blood glucose in the ICU. METHODS: Continuous time autoregressive recurrent neural networks (CTRNNs) are a deep learning model that account for irregular observations through incorporating continuous evolution of the hidden states between observations. This is achieved using a neural ordinary differential equation (ODE) or neural flow layer. In this manuscript, we give an overview of these models, including the varying architectures that have been proposed to account for issues such as ongoing medical interventions. Further, we demonstrate the application of these models to probabilistic forecasting of blood glucose in a critical care setting using electronic medical record and simulated data and compare with GBT and linear models. RESULTS: The experiments confirm that addition of a neural ODE or neural flow layer generally improves the performance of autoregressive recurrent neural networks in the irregular measurement setting. However, several CTRNN architecture are outperformed by a GBT model (Catboost), with only a long short-term memory (LSTM) and neural ODE based architecture (ODE-LSTM) achieving comparable performance on probabilistic forecasting metrics such as the continuous ranked probability score (ODE-LSTM: 0.118±0.001; Catboost: 0.118±0.001), ignorance score (0.152±0.008; 0.149±0.002) and interval score (175±1; 176±1). CONCLUSION: The application of deep learning methods for forecasting in situations with irregularly measured time series such as blood glucose shows promise. However, appropriate benchmarking by methods such as GBT approaches (plus feature transformation) are key in highlighting whether novel methodologies are truly state of the art in tabular data settings.

OBJECTIVE: We propose FedScore, a privacy-preserving federated learning framework for scoring system generation across multiple sites to facilitate cross-institutional collaborations. MATERIALS AND METHODS: The FedScore framework includes five modules: federated variable ranking, federated variable transformation, federated score derivation, federated model selection and federated model evaluation. To illustrate usage and assess FedScore's performance, we built a hypothetical global scoring system for mortality prediction within 30 days after a visit to an emergency department using 10 simulated sites divided from a tertiary hospital in Singapore. We employed a pre-existing score generator to construct 10 local scoring systems independently at each site and we also developed a scoring system using centralized data for comparison. RESULTS: We compared the acquired FedScore model's performance with that of other scoring models using the receiver operating characteristic (ROC) analysis. The FedScore model achieved an average area under the curve (AUC) value of 0.763 across all sites, with a standard deviation (SD) of 0.020. We also calculated the average AUC values and SDs for each local model, and the FedScore model showed promising accuracy and stability with a high average AUC value which was closest to the one of the pooled model and SD which was lower than that of most local models. CONCLUSION: This study demonstrates that FedScore is a privacy-preserving scoring system generator with potentially good generalizability.

Autism spectrum disorder (ASD) is a pervasive developmental disorder, and the earlier detection and timely intervention for treatment positively affect the prognosis of patients. Deep learning algorithms based on graph structure have achieved good results in autism prediction in recent years. However, there are problems with standardized operations in extracting features and combining neighborhood node features with the structure of the graph dependent, which limits the generalization ability of the trained model to other graph structures. In this paper, we propose a graph fusion autism prediction model based on attentional mechanisms(AGF) to address the above problems. The AGF model represents the overall population (patients or healthy controls) as a sparse graph, where nodes are subjects, and non-imaging features are integrated as edge weights. Different weights can be defined for different nodes in the neighborhood through the attention mechanism without relying on prior knowledge of the graph structure. The model is also able to dynamically fuse multiple sparse graphs obtained from different non-imaging features by way of training weight assignment. Its performance is also compared with several other models (e.g., S-AGF, GCN, etc.), and the results show that it has superior prediction accuracy compared to the baseline model. The results show that this improvement of graph fusion works better on the ABIDE databases, and the classification accuracy can reach 73.9%. The datasets and source code are freely available at https://github.com/chengyu-github1012/Graph-Fusion.git. Strengths and limitations of this study: graph fusion; disease prediction; noise.