Hyeonyong Hae

Invasive fractional flow reserve (FFR) is a standard tool for identifying ischemia-producing coronary stenosis. However, in clinical practice, over 70% of treatment decisions still rely on visual estimation of angiographic stenosis, which has limited accuracy (about 60%-65%) for the prediction of FFR < 0.80. One of the reasons for the visual-functional mismatch is that myocardial ischemia can be affected by the supplied myocardial size, which is not always evident by coronary angiography. The aims of this study were to develop an angiographybased machine learning (ML) algorithm for predicting the supplied myocardial volume for a stenosis, as measured using coronary computed tomography angiography (CCTA), and then to build an angiography-based classifier for the lesions with an FFR < 0.80 versus 0.80.

Inspecting the number of rebars in each column of a reinforced concrete (RC) structure is a significant task that must be undertaken during the rebar inspection process. Conventionally, counting the rebars has relied on a manual inspection carried out by visiting inspectors. However, this approach is very time-consuming, labor-intensive, and poses a potential safety risk. Previous studies have focused on the applications of counting the rebars for a production line and/or warehouse, using vision-based methods. Therefore, this study aims to propose an innovative approach incorporating the use of an unmanned aerial vehicle (UAV) on real construction sites to count the rebars automatically. For analyzing the images, robust object detection methods based on deep learning (Faster R-CNN, R-FCN, SSD 300, SSD500, YOLOv5, and YOLOv6) were developed. A total of 384 models generated from six different methods were trained and implemented using data sets based on the original and augmented images with adjustments made for the hyperparameters. In a test, the best optimized model based on Faster R-CNN produced an accuracy of 94.61% at AP50. In addition, video testing demonstrated a coverage of up to 32 frames per second in the experimental environment, suggesting that this method has potential for real-time application.

In many countries, DR (Demand Response) has been developed for which customers are motivated to save electricity by themselves during peak time to prevent grand-scale blackouts. One of the common methods in DR, is CPP (Critical Peak Pricing). Predicting energy consumption is recognized as one of the tool for dealing with CPP. There are a variety of studies in developing the model of energy consumption, which is based on energy simulation, data-driven model or metamodelling. However, it is difficult for general users to use these models due to requirement of various sensing data and expertise. And it also takes long time to simulate the models. These limitations can be an obstacle for achieving CPP’s purpose that encourages general users to manage their energy usage by themselves. As an alternative, this research suggests to use open data and GA (Genetic Algorithm)–SVR (Support Vector Regression). The model is applied to a hospital in Korea and 34,636 data sets (1 year) are collected while 31,756 (11 months) sets are used for training and 2880 sets (1 month) are used for validation. As a result, the performance of proposed model is 14.17% in CV (RMSE), which satisfies the Korea Energy Agency’s and ASHRAE (American Society of Heating, Refrigerating and Air-Conditioning Engineers) error allowance range of ±30%, and ±20% respectively.

Purpose. Pre-treatment prediction of individual blood pressure (BP) response to anti-hypertensive medication is important to determine the specific regimen for promptly and safely achieving a target BP. This study aimed to develop supervised machine learning (ML) models for predicting patient-specific treatment effects using 24-hour ambulatory BP monitoring (ABPM) data.Materials and Methods. A total of 1,129 patients who had both baseline and follow-up ABPM data were randomly assigned into training, validation and test sets in a 3:1:1 ratio. Utilising the features including clinical and laboratory findings, initial ABPM data, and anti-hypertensive medication at baseline and at follow-up, ML models were developed to predict post-treatment individual BP response. Each case was labelled by the mean 24-hour and daytime BPs derived from the follow-up ABPM.Results. At baseline, 616 (55%) patients had been treated using mono or combination therapy with 45 anti-hypertensive drugs and the remaining 513 (45%) patients had been untreated (drug-naïve). By using CatBoost, the difference between predicted vs. measured mean 24-hour systolic BP at follow-up was 8.4 ± 7.0 mm Hg (% difference of 6.6% ± 5.7%). The difference between predicted vs. measured mean 24-hour diastolic BP was 5.3 ± 4.3 mm Hg (% difference of 6.8% ± 5.5%). There were significant correlations between the CatBoost-predicted vs. the ABPM-measured changes in the mean 24-hour Systolic (r = 0.74) and diastolic (r = 0.68) BPs from baseline to follow-up. Even in the patients with renal insufficiency or diabetes, the correlations between CatBoost-predicted vs. ABPM-measured BP changes were significant.Conclusion. ML algorithms accurately predict the post-treatment ambulatory BP levels, which may assist clinicians in personalising anti-hypertensive treatment.