Eye tracking technology has been continuously researched for application in various fields. In the past, studies have been conducted to interpret eye movements in 3D space in order to solve the problem of not being able to find the centre of rotation of the eye. In this paper, we propose a novel pre-processing method for eye-gaze tracking by monitoring the front of the face with a camera. Our method works regardless of the distance between the eye and the camera. The proposed method includes an analysis technique that simplifies conventional three-dimensional space analysis to two dimensions. The contribution this work presents is a method to simplify gaze direction detection. The errors in our model’s estimations appear to be under 1 pixel. In addition, our approach has an execution time of less than 1 s, enabling an adaptive model that responds to user movements in real time. The proposed method was able to overcome various problems that methods in existing studies still suffer from, including accurately finding the rotational centre of the user’s eye-ball. Moreover, even when a user’s pupil can only be monitored from a distance, our approach still makes it possible to produce accurate estimations.
Eye-gaze direction-tracking technology is used in fields such as medicine, education, engineering, and gaming. Stability, accuracy, and precision of eye-gaze direction-tracking are demanded with simultaneous upgrades in response speed. In this study, a method is proposed to improve the speed with decreases in the system load and precision in the human pupil orbit model (HPOM) estimation method. The new method was proposed based on the phenomenon that the minor axis of the elliptical-deformed pupil always pointed toward the rotational center presented in various eye-gaze direction detection studies and HPOM estimation methods. Simulation experimental results confirmed that the speed was improved by at least 74 times by consuming less than 7 ms compared to the HPOM estimation. The accuracy of the eye’s ocular rotational center point showed a maximum error of approximately 0.2 pixels on the x-axis and approximately 8 pixels on the y-axis. The precision of the proposed method was 0.0 pixels when the number of estimation samples (ES) was 7 or less, which showed results consistent with those of the HPOM estimation studies. However, the proposed method was judged to work conservatively against the allowable angle error (AAE), considering that the experiment was conducted under the worst conditions and the cost used to estimate the final model. Therefore, the proposed method could estimate HPOM with high accuracy and precision through AAE adjustment according to system performance and the usage environment.
The frustrated total internal reflection (FTIR) optical fingerprint scanning method is widely used due to its cost-effectiveness. However, fingerprint image quality is highly dependent on fingertip surface conditions, with moisture generally considered a degrading factor. Interestingly, a prior study reported that excessive moisture may improve image quality, though their findings were based on qualitative observations, necessitating further quantitative analysis. Additionally, since the FTIR method relies on optical principles, image quality is also influenced by the wavelength of the light source. In this study, we conducted a preliminary clinical experiment to quantitatively analyze the impact of moisture levels on fingertips (wet, dry, and control) and light wavelengths (red, green, and blue) on FTIR fingerprint image quality. A total of 20 male and female participants with no physical impairments were involved. The results suggest that FTIR fingerprint image quality may improve under wet conditions and when illuminated with green and blue light sources compared to dry conditions and red light. Statistical evidence supports this consistent trend. However, given the limited sample size, the statistical validity and generalizability of these findings should be interpreted with caution. These insights provide a basis for optimizing fingerprint imaging conditions, potentially enhancing the reliability and accuracy of automatic fingerprint identification systems (AFIS) by reducing variations in individual fingerprint quality.
The purpose of this study was to develop an automatic bronchial aspiration system to minimize tracheal mucosal damage in ventilator patients. Operation performances of the system's suction pressure, bronchial intubation depth, suction cycle, and tube cuff pressure were tested. To check clinical results, subjects underwent endoscopy after applying the previous manual method for 24 hours. After that, they underwent endoscopy after applying the proposed suction system for 24 hours. For quantitative evaluation of test results, tracheal mucosal injury was divided into five grades: Grade 0 = normal, Grade 1 = erythema or edema, Grade 2 = erosion, Grade 3 = hemorrhage, and Grade 4 = ulcer or necrosis. In the performance test, an error of up to 12 mmHg occurred within the normal operation error range for suction pressure control. The insertion depth control had a maximum error of 7.0 mm within the normal operation error range. On the other hand, there was no error in the time control or the tube cuff pressure control. In the clinical trial, after using the proposed system for five subjects to find changes in tracheal mucosal injury by endoscopy, reduced injury or no change in injury was found. The system proposed in this study is confirmed to be able to remove sputum while minimizing tracheal mucosal injury that can occur when using previous manual suction device.
Heart rate variability (HRV)–based seizure prediction studies typically aim to identify preictal states by distinguishing them from interictal periods. However, in clinical practice, interictal and preictal boundaries are not temporally well defined, and relying on heuristic definitions may blur the physiological distinction between seizure-related and non–seizure-related dynamics. Furthermore, conventionally defined interictal periods may inadvertently include artifacts, signal instability, or latent seizure-related pathological characteristics. To address this limitation, we propose a seizure-centered one-class deep learning approach that directly models ictal HRV patterns using clinician-annotated seizure timing as the ground truth. Unlike conventional methods that detect anomalies from an interictal baseline, the proposed method learns the distribution of ictal HRV patterns and evaluates the closeness of incoming HRV windows to the learned ictal distribution in a latent space. Seizure prediction performance was evaluated under a seizure prediction horizon–seizure occurrence period scheme (SPH = 10 min, SOP = 50 min) using internal and external validation datasets. Experimental results demonstrate that the proposed seizure-centered approach consistently achieves higher performance than the conventional interictal-based anomaly detection method. Additionally, by identifying ictal-like patterns within undiagnosed interictal periods, this method may capture latent pathological events, such as preictal states, aborted seizures, or subclinical seizures that do not manifest overt clinical symptoms. Therefore, this study resolves the reliance on ambiguous definitions by mitigating interictal uncertainty.