Background: Despite excellent prediction performance, noninterpretability has undermined the value of applying deep-learning algorithms in clinical practice. To overcome this limitation, attention mechanism has been introduced to clinical research as an explanatory modeling method. However, potential limitations of using this attractive method have not been clarified to clinical researchers. Furthermore, there has been a lack of introductory information explaining attention mechanisms to clinical researchers.Objective: The aim of this study was to introduce the basic concepts and design approaches of attention mechanisms. In addition, we aimed to empirically assess the potential limitations of current attention mechanisms in terms of prediction and interpretability performance.Methods: First, the basic concepts and several key considerations regarding attention mechanisms were identified. Second, four approaches to attention mechanisms were suggested according to a two-dimensional framework based on the degrees of freedom and uncertainty awareness. Third, the prediction performance, probability reliability, concentration of variable importance, consistency of attention results, and generalizability of attention results to conventional statistics were assessed in the diabetic classification modeling setting. Fourth, the potential limitations of attention mechanisms were considered.Results: Prediction performance was very high for all models. Probability reliability was high in models with uncertainty awareness. Variable importance was concentrated in several variables when uncertainty awareness was not considered. The consistency of attention results was high when uncertainty awareness was considered. The generalizability of attention results to conventional statistics was poor regardless of the modeling approach.Conclusions: The attention mechanism is an attractive technique with potential to be very promising in the future. However, it may not yet be desirable to rely on this method to assess variable importance in clinical settings. Therefore, along with theoretical studies enhancing attention mechanisms, more empirical studies investigating potential limitations should be encouraged.

Background: Postoperative length of stay is a key indicator in the management of medical resources and an indirect predictor of the incidence of surgical complications and the degree of recovery of the patient after cancer surgery. Recently, machine learning has been used to predict complex medical outcomes, such as prolonged length of hospital stay, using extensive medical information. Objective: The objective of this study was to develop a prediction model for prolonged length of stay after cancer surgery using a machine learning approach. Methods: In our retrospective study, electronic health records (EHRs) from 42,751 patients who underwent primary surgery for 17 types of cancer between January 1, 2000, and December 31, 2017, were sourced from a single cancer center. The EHRs included numerous variables such as surgical factors, cancer factors, underlying diseases, functional laboratory assessments, general assessments, medications, and social factors. To predict prolonged length of stay after cancer surgery, we employed extreme gradient boosting classifier, multilayer perceptron, and logistic regression models. Prolonged postoperative length of stay for cancer was defined as bed-days of the group of patients who accounted for the top 50% of the distribution of bed-days by cancer type. Results: In the prediction of prolonged length of stay after cancer surgery, extreme gradient boosting classifier models demonstrated excellent performance for kidney and bladder cancer surgeries (area under the receiver operating characteristic curve [AUC] >0.85). A moderate performance (AUC 0.70-0.85) was observed for stomach, breast, colon, thyroid, prostate, cervix uteri, corpus uteri, and oral cancers. For stomach, breast, colon, thyroid, and lung cancers, with more than 4000 cases each, the extreme gradient boosting classifier model showed slightly better performance than the logistic regression model, although the logistic regression model also performed adequately. We identified risk variables for the prediction of prolonged postoperative length of stay for each type of cancer, and the importance of the variables differed depending on the cancer type. After we added operative time to the models trained on preoperative factors, the models generally outperformed the corresponding models using only preoperative variables. Conclusions: A machine learning approach using EHRs may improve the prediction of prolonged length of hospital stay after primary cancer surgery. This algorithm may help to provide a more effective allocation of medical resources in cancer surgery.

Data warehousing is the most important technology to address recent advances in precision medicine. However, a generic clinical data warehouse does not address unstructured and insufficient data. In precision medicine, it is essential to develop a platform that can collect and utilize data. Data were collected from electronic medical records, genomic sequences, tumor biopsy specimens, and national cancer control initiative databases in the National Cancer Center (NCC), Korea. Data were de-identified and stored in a safe and independent space. Unstructured clinical data were standardized and incorporated into cancer registries and linked to cancer genome sequences and tumor biopsy specimens. Finally, national cancer control initiative data from the public domain were independently organized and linked to cancer registries. We constructed a system for integrating and providing various cancer data called the Korea Cancer Big Data Platform (K-CBP). Although the K-CBP could be used for cancer research, the legal and regulatory aspects of data distribution and usage need to be addressed first. Nonetheless, the system will continue collecting data from cancer-related resources that will hopefully facilitate precision-based research.

Image compression is used in several clinical organizations to help address the overhead associated with medical imaging. These methods reduce file size by using a compact representation of the original image. This study aimed to analyze the impact of image compression on the performance of deep learning-based models in classifying mammograms as “malignant”—cases that lead to a cancer diagnosis and treatment—or “normal” and “benign,” non-malignant cases that do not require immediate medical intervention. In this retrospective study, 9111 unique mammograms–5672 normal, 1686 benign, and 1754 malignant cases were collected from the National Cancer Center in the Republic of Korea. Image compression was applied to mammograms with compression ratios (CRs) ranging from 15 to 11 K. Convolutional neural networks (CNNs) with three convolutional layers and three fully-connected layers were trained using these images to classify a mammogram as malignant or not malignant across a range of CRs using five-fold cross-validation. Models trained on images with maximum CRs of 5 K had an average area under the receiver operating characteristic curve (AUROC) of 0.87 and area under the precision-recall curve (AUPRC) of 0.75 across the five folds and compression ratios. For images compressed with CRs of 10 K and 11 K, model performance decreased (average 0.79 in AUROC and 0.49 in AUPRC). Upon generating saliency maps that visualize the areas each model views as significant for prediction, models trained on less compressed (CR < = 5 K) images had maps encapsulating a radiologist’s label, while models trained on images with higher amounts of compression had maps that missed the ground truth completely. In addition, base ResNet18 models pre-trained on ImageNet and trained using compressed mammograms did not show performance improvements over our CNN model, with AUROC and AUPRC values ranging from 0.77 to 0.87 and 0.52 to 0.71 respectively when trained and tested on images with maximum CRs of 5 K. This paper finds that while training models on images with increased the robustness of the models when tested on compressed data, moderate image compression did not substantially impact the classification performance of DL-based models.

A screening model for undiagnosed diabetes mellitus (DM) is important for early medical care. Insufficient research has been carried out developing a screening model for undiagnosed DM using machine learning techniques. Thus, the primary objective of this study was to develop a screening model for patients with undiagnosed DM using a deep neural network. We conducted a cross-sectional study using data from the Korean National Health and Nutrition Examination Survey (KNHANES) 2013–2016. A total of 11,456 participants were selected, excluding those with diagnosed DM, an age < 20 years, or missing data. KNHANES 2013–2015 was used as a training dataset and analyzed to develop a deep learning model (DLM) for undiagnosed DM. The DLM was evaluated with 4444 participants who were surveyed in the 2016 KNHANES. The DLM was constructed using seven non-invasive variables (NIV): age, waist circumference, body mass index, gender, smoking status, hypertension, and family history of diabetes. The model showed an appropriate performance (area under curve (AUC): 80.11) compared with existing previous screening models. The DLM developed in this study for patients with undiagnosed diabetes could contribute to early medical care.