Park SW, et al. J Korean Med Sci. 2024 Jan;39:e53. https://doi.org/10.3346/jkms.2024.39.e53

Background: Vancomycin pharmacokinetics are highly variable in patients with critical illnesses, and clinicians commonly use population pharmacokinetic (PPK) models based on a Bayesian approach to dose. However, these models are population-dependent, may only sometimes meet the needs of individual patients, and are only used by experienced clinicians as a reference for making treatment decisions. To assist real-world clinicians, we developed a deep learning–based decision-making system that predicts vancomycin therapeutic drug monitoring (TDM) levels in patients in intensive care unit. Objective: This study aimed to establish joint multilayer perceptron (JointMLP), a new deep-learning model for predicting vancomycin TDM levels, and compare its performance with the PPK models, extreme gradient boosting (XGBoost), and TabNet. Methods: We used a 977-case data set split into training and testing groups in a 9:1 ratio. We performed external validation of the model using 1429 cases from Kangwon National University Hospital and 2394 cases from the Medical Information Mart for Intensive Care–IV (MIMIC-IV). In addition, we performed 10-fold cross-validation on the internal training data set and calculated the 95% CIs using the metric. Finally, we evaluated the generalization ability of the JointMLP model using the MIMIC-IV data set. Results: Our JointMLP model outperformed other models in predicting vancomycin TDM levels in internal and external data sets. Compared to PPK, the JointMLP model improved predictive power by up to 31% (mean absolute error [MAE] 6.68 vs 5.11) on the internal data set and 81% (MAE 11.87 vs 6.56) on the external data set. In addition, the JointMLP model significantly outperforms XGBoost and TabNet, with a 13% (MAE 5.75 vs 5.11) and 14% (MAE 5.85 vs 5.11) improvement in predictive accuracy on the inner data set, respectively. On both the internal and external data sets, our JointMLP model performed well compared to XGBoost and TabNet, achieving prediction accuracy improvements of 34% and 14%, respectively. Additionally, our JointMLP model showed higher robustness to outlier data than the other models, as evidenced by its higher root mean squared error performance across all data sets. The mean errors and variances of the JointMLP model were close to zero and smaller than those of the PPK model in internal and external data sets. Conclusions: Our JointMLP approach can help optimize treatment outcomes in patients with critical illnesses in an intensive care unit setting, reducing side effects associated with suboptimal vancomycin administration. These include increased risk of bacterial resistance, extended hospital stays, and increased health care costs. In addition, the superior performance of our model compared to existing models highlights its potential to help real-world clinicians.

Pressure ulcers (PUs) are a prevalent skin disease affecting patients with impaired mobility and in high-risk groups. These ulcers increase patients’ suffering, medical expenses, and burden on medical staff. This study introduces a clinical decision support system and verifies it for predicting real-time PU occurrences within the intensive care unit (ICU) by using MIMIC-IV and in-house ICU data. We develop various machine learning (ML) and deep learning (DL) models for predicting PU occurrences in real time using the MIMIC-IV and validate using the MIMIC-IV and Kangwon National University Hospital (KNUH) dataset. To address the challenge of missing values in time series, we propose a novel recurrent neural network model, GRU-D++. This model outperformed other experimental models by achieving the area under the receiver operating characteristic curve (AUROC) of 0.945 for the on-time prediction and AUROC of 0.912 for 48h in-advance prediction. Furthermore, in the external validation with the KNUH dataset, the fine-tuned GRU-D++ model demonstrated superior performances, achieving an AUROC of 0.898 for on-time prediction and an AUROC of 0.897 for 48h in-advance prediction. The proposed GRU-D++, designed to consider temporal information and missing values, stands out for its predictive accuracy. Our findings suggest that this model can significantly alleviate the workload of medical staff and prevent the worsening of patient conditions by enabling timely interventions for PUs in the ICU.

Early detection of lung nodules is essential for preventing lung cancer. However, the number of radiologists who can diagnose lung nodules is limited, and considerable effort and time are required. To address this problem, researchers are investigating the automation of deep-learning-based lung nodule detection. However, deep learning requires large amounts of data, which can be difficult to collect. Therefore, data collection should be optimized to facilitate experiments at the beginning of lung nodule detection studies. We collected chest computed tomography scans from 515 patients with lung nodules from three hospitals and high-quality lung nodule annotations reviewed by radiologists. We conducted several experiments using the collected datasets and publicly available data from LUNA16. The object detection model, YOLOX was used in the lung nodule detection experiment. Similar or better performance was obtained when training the model with the collected data rather than LUNA16 with large amounts of data. We also show that weight transfer learning from pre-trained open data is very useful when it is difficult to collect large amounts of data. Good performance can otherwise be expected when reaching more than 100 patients. This study offers valuable insights for guiding data collection in lung nodules studies in the future.

This study introduces a deep-learning-based automatic sleep scoring system to detect sleep apnea using a single-lead electrocardiography (ECG) signal, focusing on accurately estimating the apnea–hypopnea index (AHI). Unlike other research, this work emphasizes AHI estimation, crucial for the diagnosis and severity evaluation of sleep apnea. The suggested model, trained on 1465 ECG recordings, combines the deep-shallow fusion network for sleep apnea detection network (DSF-SANet) and gated recurrent units (GRUs) to analyze ECG signals at 1-min intervals, capturing sleep-related respiratory disturbances. Achieving a 0.87 correlation coefficient with actual AHI values, an accuracy of 0.82, an F1 score of 0.71, and an area under the receiver operating characteristic curve of 0.88 for per-segment classification, our model was effective in identifying sleep-breathing events and estimating the AHI, offering a promising tool for medical professionals.

Background and objective: The incidence of facial fractures is on the rise globally, yet limited studies are addressing the diverse forms of facial fractures present in 3D images. In particular, due to the nature of the facial fracture, the direction in which the bone fractures vary, and there is no clear outline, it is difficult to determine the exact location of the fracture in 2D images. Thus, 3D image analysis is required to find the exact fracture area, but it needs heavy computational complexity and expensive pixel-wise labeling for supervised learning. In this study, we tackle the problem of reducing the computational burden and increasing the accuracy of fracture localization by using a weakly-supervised object localization without pixel-wise labeling in a 3D image space. Methods: We propose a Very Fast, High-Resolution Aggregation 3D Detection CAM (VFHA-CAM) model, which can detect various facial fractures. To better detect tiny fractures, our model uses high-resolution feature maps and employs Ablation CAM to find an exact fracture location without pixel-wise labeling, where we use a rough fracture image detected with 3D box-wise labeling. To this end, we extract important features and use only essential features to reduce the computational complexity in 3D image space. Results: Experimental findings demonstrate that VFHA-CAM surpasses state-of-the-art 2D detection methods by up to 20% in sensitivity/person and specificity/person, achieving sensitivity/person and specificity/person scores of 87% and 85%, respectively. In addition, Our VFHA-CAM reduces location analysis time to 76 s without performance degradation compared to a simple Ablation CAM method that takes more than 20 min. Conclusion: This study introduces a novel weakly-supervised object localization approach for bone fracture detection in 3D facial images. The proposed method employs a 3D detection model, which helps detect various forms of facial bone fractures accurately. The CAM algorithm adopted for fracture area segmentation within a 3D fracture detection box is key in quickly informing medical staff of the exact location of a facial bone fracture in a weakly-supervised object localization. In addition, we provide 3D visualization so that even non-experts unfamiliar with 3D CT images can identify the fracture status and location.