Facing an ongoing organ shortage in transplant medicine, strategies to increase the use of organs from marginal donors by objective organ assessment are being fostered. In this context, normothermic machine perfusion provides a platform for ex vivo organ evaluation during preservation. Consequently, analytical tools are emerging to determine organ quality. In this study, hyperspectral imaging (HSI) in the wavelength range of 550 nm–995 nm was applied. Classification of 26 kidneys based on HSI was established using KidneyResNet, a convolutional neural network (CNN) based on the ResNet-18 architecture, to predict inulin clearance behavior. HSI preprocessing steps were implemented, including automated region of interest (ROI) selection, before executing the KidneyResNet algorithm. Training parameters and augmentation methods were investigated concerning their influence on the prediction. When classifying individual ROIs, the optimized KidneyResNet model achieved 84% and 62% accuracy in the validation and test set, respectively. With a majority decision on all ROIs of a kidney, the accuracy increased to 96% (validation set) and 100% (test set). These results demonstrate the feasibility of HSI in combination with KidneyResNet for non-invasive prediction of ex vivo kidney function. This knowledge of preoperative renal quality may support the organ acceptance decision.
Modern patient care aims for individualized solutions. Current machine learning techniques, in general and in the medical domain, typically incorporate big amounts of data. In fact, more data contributes to the generalizability of said techniques. However, it might conflict with the desire for individualized solutions. Our works aim at the implementation of individual solutions based on machine learning techniques. Within this contribution, we investigate the potential benefit of individualized classifiers in the context of automatic sleep staging using cardiorespiratory features.To that end, we performed sleep stage classification using 237 records of the Sleep Heart Health Study. For each patient, we trained an ensemble classifier that is based on a subset of the available patients. Such subsets of varying size were chosen by a modified version of sequential forward floating selection. Our results show that the individualized classifier improves classification compared to a classifier that uses all available patients by 30% (improvement in Cohen's kappa coefficient (κ) of 0.15 from 0.46 to 0.61). On average the subset used for training thereby includes five patients.The presented contribution clearly depicts the potential of an individualized classification approach. Based on the current results, future works will try to establish metrics that can identify the most appropriate training subset in an unsupervised way.
Early detection and treatment of sepsis is of utmost importance concerning sepsis outcome and costs. However, revealing patterns in vital signs and laboratory measurements which facilitate reliable prediction of sepsis onset remains challenging. Especially exploiting the time series characteristic of those measurements is expected to play a major role concerning successful sepsis prediction. Within this work, we propose a stacked combination of a recurrent neuronal network (RNN) and a light gradient boosted machine (LGBM) to target the objective of sepsis onset prediction. Here, 8 vital signs, 26 laboratory measurements and 3 demographic parameters are included as input to our classification model. Our last running model achieved a utility score on full test set of 0.114 (TU Dresden-IBMT).
This work aims to classify sleep stages based on tachograms using Convolutional Neural Networks (CNNs) and investigate advantages of specialized classifiers. The tachograms of 5422 patients were extracted from the Sleep Heart Health Study. A CNN was trained to classify each 30 s epoch into four distinct sleep stages. The patients were divided into four subgroups by Apnoe-Hypopnoe-Index (AHI). From each subgroup, 20 % of patients were held out as test data. One general model was trained on all training patients and four narrowed models were each trained on one subgroup. Furthermore, the general model was retrained on the subgroups, yielding four additional transfer learning models. Our general model gained an average Cohen's Kappa score of 0.53. The general model outperformed the narrowed models on each test subset. From the narrowed models, training on the subgroup with AHI 5-15 achieved best overall performance. However, a correlation exists between the size of train sets and classification quality. Transfer learning did not improve the results. CNN models are capable of learning features from tachograms with very good classification performance compared to other works using heart rate only. However, the pursued strategies for specializing classifiers did not yield any advantages over our general model.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.