Computer-aided polyp detection in gastric gastroscopy has been the subject of research over the past few decades. However, despite significant advances, automatic polyp detection in real time is still an unsolved problem. In this paper, we report on a convolutional neural network (CNN) for polyp detection that is constructed based on Single Shot MultiBox Detector (SSD) architecture and which we call SSD for Gastric Polyps (SSD-GPNet). To take full advantages of feature maps’ information from the feature pyramid and to acquire higher accuracy, we re-use information that is abandoned by Max-Pooling layers. In other words, we reuse the lost data from the pooling layers and concatenate that data as extra feature maps to contribute to classification and detection. Meanwhile, in the feature pyramid, we concatenate feature maps of the lower layers and feature maps that are deconvolved from upper layers to make explicit relationships between layers and to effectively increase the number of channels. The results show that our enhanced SSD for gastric polyp detection can realize real-time polyp detection with 50 frames per second (FPS) and can improve the mean average precision (mAP) from 88.5% to 90.4%, with only a little loss in time-performance. And the further experiment shows that SSD-GPNet has excellent performance in improving polyp detection recalls over 10% (p = 0.00053), especially in small polyp detection. This can help endoscopic physicians more easily find missed polyps and decrease the gastric polyp miss rate. It may be applicable in daily clinical practice to reduce the burden on physicians.

Objective: Acute coronary syndrome (ACS), as a common and severe cardiovascular disease, is a leading cause of death and the principal cause of serious long-term disability globally. Clinical risk prediction of ACS is important for early intervention and treatment. Existing ACS risk scoring models are based mainly on a small set of hand-picked risk factors and often dichotomize predictive variables to simplify the score calculation. Methods: This study develops a regularized stacked denoising auto-encoder (SDAE) model to stratify clinical risks of ACS patients from a large volume of electronic health records (EHR). To capture characteristics of patients at similar risk levels, and preserve the discriminating information across different risk levels, two constraints are added on SDAE to make the reconstructed feature representations contain more risk information of patients, which contribute to a better clinical risk prediction result. Results: We validate our approach on a real clinical dataset consisting of 3464 ACS patient samples. The performance of our approach for predicting ACS risk remains robust and reaches 0.868 and 0.73 in terms of both AUC and accuracy, respectively. Conclusions: The obtained results show that the proposed approach achieves a competitive performance compared to state-of-the-art models in dealing with the clinical risk prediction problem. In addition, our approach can extract informative risk factors of ACS via a reconstructive learning strategy. Some of these extracted risk factors are not only consistent with existing medical domain knowledge, but also contain suggestive hypotheses that could be validated by further investigations in the medical domain.

Background: The interpretability of results predicted by the machine learning models is vital, especially in the critical fields like healthcare. With the increasingly adoption of electronic healthcare records (EHR) by the medical organizations in the last decade, which accumulated abundant electronic patient data, neural networks or deep learning techniques are gradually being applied to clinical tasks by utilizing the huge potential of EHR data. However, typical deep learning models are black-boxes, which are not transparent and the prediction outcomes of which are difficult to interpret. Methods: To remedy this limitation, we propose an attention neural network model for interpretable clinical prediction. In detail, the proposed model employs an attention mechanism to capture critical/essential features with their attention signals on the prediction results, such that the predictions generated by the neural network model can be interpretable. Results: We evaluate our proposed model on a real-world clinical dataset consisting of 736 samples to predict readmissions for heart failure patients. The performance of the proposed model achieved 66.7 and 69.1% in terms of accuracy and AUC, respectively, and outperformed the baseline models. Besides, we displayed patient-specific attention weights, which can not only help clinicians understand the prediction outcomes, but also assist them to select individualized treatment strategies or intervention plans. Conclusions: The experimental results demonstrate that the proposed model can improve both the prediction performance and interpretability by equipping the model with an attention mechanism.

Objective: To compare the performance, clinical feasibility, and reliability of statistical and machine learning (ML) models in predicting heart failure (HF) events. Background: Although ML models have been proposed to revolutionize medicine, their promise in predicting HF events has not been investigated in detail. Methods: A systematic search was performed on Medline, Web of Science, and IEEE Xplore for studies published between January 1, 2011 to July 14, 2021 that developed or validated at least one statistical or ML model that could predict all-cause mortality or all-cause readmission of HF patients. Prediction Model Risk of Bias Assessment Tool was used to assess the risk of bias, and random effect model was used to evaluate the pooled c-statistics of included models. Result: Two-hundred and two statistical model studies and 78 ML model studies were included from the retrieved papers. The pooled c-index of statistical models in predicting all-cause mortality, ML models in predicting all-cause mortality, statistical models in predicting all-cause readmission, ML models in predicting all-cause readmission were 0.733 (95% confidence interval 0.724–0.742), 0.777 (0.752–0.803), 0.678 (0.651–0.706), and 0.660 (0.633–0.686), respectively, indicating that ML models did not show consistent superiority compared to statistical models. The head-to-head comparison revealed similar results. Meanwhile, the immoderate use of predictors limited the feasibility of ML models. The risk of bias analysis indicated that ML models' technical pitfalls were more serious than statistical models'. Furthermore, the efficacy of ML models among different HF subgroups is still unclear. Conclusions: ML models did not achieve a significant advantage in predicting events, and their clinical feasibility and reliability were worse.