Activation functions are essential in deep learning, and the rectified linear unit (ReLU) is the most widely used activation function to solve the vanishing gradient problem. However, owing to the dying ReLU problem and bias shift effect, deep learning models using ReLU cannot exploit the potential benefits of negative values. Numerous ReLU variants have been proposed to address this issue. In this study, we propose Dynamic Parametric ReLU (DPReLU), which can dynamically control the overall functional shape of ReLU with four learnable parameters. The parameters of DPReLU are determined by training rather than by humans, thereby making the formulation more suitable and flexible for each model and dataset. Furthermore, we propose an appropriate and robust weight initialization method for DPReLU. To evaluate DPReLU and its weight initialization method, we performed two experiments on various image datasets: one using an autoencoder for image generation and the other using the ResNet50 for image classification. The results show that DPReLU and our weight initialization method provide faster convergence and better accuracy than the original ReLU and the previous ReLU variants.

With the global trend toward an aging population, the increasing number of dementia patients and elderly living alone has emerged as a serious social issue in South Korea. The assessment of activities of daily living (ADL) is essential for diagnosing dementia. However, since the assessment is based on the ADL questionnaire, it relies on subjective judgment and lacks objectivity. Seven healthy seniors and six with early-stage dementia participated in the study to obtain ADL data. The derived ADL features were generated by smart home sensors. Statistical methods and machine learning techniques were employed to develop a model for auto-classifying the normal controls and early-stage dementia patients. The proposed approach verified the developed model as an objective ADL evaluation tool for the diagnosis of dementia. A random forest algorithm was used to compare a personalized model and a non-personalized model. The comparison result verified that the accuracy (91.20%) of the personalized model was higher than that (84.54%) of the non-personalized model. This indicates that the cognitive ability-based personalization showed encouraging performance in the classification of normal control and early-stage dementia and it is expected that the findings of this study will serve as important basic data for the objective diagnosis of dementia.

To design an efficient deep learning model that can be used in the real-world, it is important to detect out-of-distribution (OOD) data well. Various studies have been conducted to solve the OOD problem. The current state-of-the-art approach uses a confidence score based on the Mahalanobis distance in a feature space. Although it outperformed the previous approaches, the results were sensitive to the quality of the trained model and the dataset complexity. Herein, we propose a novel OOD detection method that can train more efficient feature space for OOD detection. The proposed method uses an ensemble of the features trained using the softmax-based classifier and the network based on distance metric learning (DML). Through the complementary interaction of these two networks, the trained feature space has a more clumped distribution and can fit well on the Gaussian distribution by class. Therefore, OOD data can be efficiently detected by setting a threshold in the trained feature space. To evaluate the proposed method, we applied our method to various combinations of image datasets. The results show that the overall performance of the proposed approach is superior to those of other methods, including the state-of-the-art approach, on any combination of datasets.

Background: Sepsis is diagnosed in millions of people every year, resulting in a high mortality rate. Although patients with sepsis present multimorbid conditions, including cancer, sepsis predictions have mainly focused on patients with severe injuries. Objective: In this paper, we present a machine learning–based approach to identify the risk of sepsis in patients with cancer using electronic health records (EHRs). Methods: We utilized deidentified anonymized EHRs of 8580 patients with cancer from the Samsung Medical Center in Korea in a longitudinal manner between 2014 and 2019. To build a prediction model based on physical status that would differ between sepsis and nonsepsis patients, we analyzed 2462 laboratory test results and 2266 medication prescriptions using graph network and statistical analyses. The medication relationships and lab test results from each analysis were used as additional learning features to train our predictive model. Results: Patients with sepsis showed differential medication trajectories and physical status. For example, in the network-based analysis, narcotic analgesics were prescribed more often in the sepsis group, along with other drugs. Likewise, 35 types of lab tests, including albumin, globulin, and prothrombin time, showed significantly different distributions between sepsis and nonsepsis patients (P<.001). Our model outperformed the model trained using only common EHRs, showing an improved accuracy, area under the receiver operating characteristic (AUROC), and F1 score by 11.9%, 11.3%, and 13.6%, respectively. For the random forest–based model, the accuracy, AUROC, and F1 score were 0.692, 0.753, and 0.602, respectively. Conclusions: We showed that lab tests and medication relationships can be used as efficient features for predicting sepsis in patients with cancer. Consequently, identifying the risk of sepsis in patients with cancer using EHRs and machine learning is feasible.

High-resolution medical images are critical for preserving intricate anatomical details essential for accurate diagnosis, effective surgical planning, and creating precise digital twins. However, acquiring such images often requires expensive equipment, specialized personnel, considerable time, and significant financial investment. This research addresses these challenges by proposing an AI-driven specialized super-resolution (SR) framework tailored for medical imaging. Our approach leverages transfer learning by fine-tuning high-performing general-domain SR models (BSRGAN, DPSR, HAT, RealESRGAN, and SwinIR) using approximately 190,000 images from KISTI’s Digital Korean dataset. On average, the fine-tuned SR models exhibited a 3.28% improvement in PSNR and a 0.6% increase in SSIM compared to their zero-shot counterparts, underscoring the effectiveness of transfer learning in enhancing both image quality and structural fidelity for medical applications. Notably, the fine-tuned HAT model achieved the highest performance, recording a Peak Signal-to-Noise Ratio (PSNR) of 39.95 and a Structural Similarity Index Map (SSIM) of 0.9461. The DPSR model also demonstrated robust detail preservation, with a PSNR of 39.60 and an SSIM of 0.9421. Qualitative evaluations by anatomy and medical imaging specialists confirmed the enhanced quality of the high-resolution outputs. These advancements highlight the transformative potential of specialized SR models to overcome the limitations of general-domain approaches in medical imaging. By laying the groundwork for developing high-resolution medical datasets, the proposed framework supports more precise diagnostics and advances the creation of accurate 3D human models and digital twins. Overall, this work underscores the vital role of AI-driven innovation in shaping the future of medical imaging and precision healthcare.