In this work we present a machine learning approach that is able to classify 30 cardiac abnormalities from an arbitrary number of electrocardiogram (ECG) leads. Features extracted by a deep convolutional neural network are combined with hand-crafted features (demographic, morphological, and heart rate variability metrics) and fed into a multilayer perceptron. We employ an Asymmetric Loss (ASL) function, which enables the model to focus on hard, but under-represented, samples. To mitigate the issue of ground-truth mislabeling and to provide robustness, we investigate the use of a self-learning label correction method that iteratively estimates correct labels during training. Leaderboard results show our team SMS+1 achieved challenge scores of 0.57 0.58 0.57.56 0.57 for twelve, six, four, three, and two-lead, respectively. Our model maintains the same diagnostic potential on both standard twelve-lead ECGs and reduced-lead ECGs.
Objective assessment of emotions is a challenging task.In this research, a new approach for detection and quantification of four emotions (pleasure, sadness, anger, and joy) is introduced based on the 2D regression models of emotions.ECG signal of 70 healthy female volunteers was recorded while the participants were stimulated by four different colors for five minutes.Induced arousal and valence levels were identified using the Self-Assessment Manikin test.Based on extracted features from Poincare Plot of heart rate series, 1-the label of emotion is determined based on 2-D emotion models using a rulebased approach, and 2-the strength of emotion is calculated in percentages using proposed mathematical formula.The results confirm the findings of previous studies that showed that each color is related to one emotion.The mean square error for estimation of arousal and valence levels were 0.0393 and 0.0536, respectively.In emotion classification, sensitivity of 96.78%, specificity of 99.4%, and the accuracy of 98.75% were achieved.The proposed method for measuring emotions can be of great help to psychiatrists to identify mental disorders.Using this method, one can measure and express the strength of emotions individually and has a comparative criterion for emotions in different individuals.
Tachyarrhythmia detection through RR interval analysis could improve performance of monitoring devices. In this paper a Poincaré plot-based image approach is presented. Three cardiac rhythms were analyzed in this study: normal sinus rhythm (NSR), atrial fibrillation (AF) and atrial bigeminy (AB). Using different MIT-BIH databases, 27955, 3363 and 76 images were generated for NSR, AF and AB respectively using a 2-minute window with 50 % overlap. The 80 % of the data available for each rhythm was used to create a reference rhythm image atlas. The remaining 20 % was classified into one of the three categories using mutual information. The process was iterated 10 times, in which images used to construct the atlas and used to create the test set were randomly selected. AF was correctly classified 94.12 % ± 0.45, AB 72.00 % ± 11.24 and NSR 80.70 %±0.54. The results of the present study suggest that Poincaré plot-based image analysis is a promising path for classifying different rhythms using only ventricular activity.
Computational fluid dynamics represents a valuable non-invasive approach to determine and assess physically meaningful parameters in a complex fluid dynamics system represented by AF. The aim of this study was the design, development and test of different LA motion fields in AF on a patient-specific 3D anatomical model to clarify the influence of contraction models on LA hemodynamics. Since LA motion field is not available from clinical data during AF, three displacement models were designed to simulate the irregular, disorganized, very rapid and strongly reduced LA contraction, a random model, a discrete random model and a con-tinuous sinusoidal model. Blood velocity fields, kinetic energy, vortex structures and blood stasis were analyzed in both SR and AF conditions in the LA. Velocities in SR were higher than in AF, particular-ly during atrial systole. The three AF models resulted in different wash-out velocities both at the mitral valve and at the ostium of the LA appendage. Vortices were also differently distributed inside the LA showing a more organized flow in the sinusoidal model which was also characterized by the lowest blood stasis (8.6%) inside the LA appendage. Overall, different LA de-formation models in AF affect LA hemodynamics and additional studies should be performed to develop a realistic contraction model to simulate AF episodes.
Early prediction of sepsis is of utmost importance to provide optimal care at an early stage. This work aims to use machine learning for early prediction of sepsis using ratio and power-based feature transformation. The feature transformation and feature selection process is optimized by applying a genetic algorithm (GA) based approach to extract the information specific to the sepsis from the given raw patient covariates that maximizes the underlying classification performance in terms of utility score. The proposed method begins with filling the missing values in the training dataset. Then, GA is applied strategically to identify influential ratio and power-based features from the raw patient covariates. The utility score is maximized as an objective of the optimization. RusBoost is used with default settings for underlying classification during optimization. Subsequently, an optimal RusBoost model is developed with a set of 55 identified features. Independent performance evaluation of the proposed method with the 2019 PhysioNet/CinC Challenge dataset has officially achieved 19th rank with a utility score of 30.9% on the full hidden test data. This work appears as Shivpatidar on the leader-board. The proposed early warning system has potential clinical value in critical care clinics.
