The current study investigates the robustness of deep learning models for accurate medical diagnosis systems with a specific focus on their ability to maintain performance in the presence of adversarial or noisy inputs. We examine factors that may influence model reliability, including model complexity, training data quality, and hyperparameters; we also examine security concerns related to adversarial attacks that aim to deceive models along with privacy attacks that seek to extract sensitive information. Researchers have discussed various defenses to these attacks to enhance model robustness, such as adversarial training and input preprocessing, along with mechanisms like data augmentation and uncertainty estimation. Tools and packages that extend the reliability features of deep learning frameworks such as TensorFlow and PyTorch are also being explored and evaluated. Existing evaluation metrics for robustness are additionally being discussed and evaluated. This paper concludes by discussing limitations in the existing literature and possible future research directions to continue enhancing the status of this research topic, particularly in the medical domain, with the aim of ensuring that AI systems are trustworthy, reliable, and stable.

Smartphones have become crucial for our daily life activities and are increasingly loaded with our personal information to perform several sensitive tasks including mobile banking, communication, and are used for storing private photos and files. Therefore, there is a high demand for applying usable authentication techniques that prevent unauthorized access to sensitive information. In this work, we propose , a deep learning-based active authentication approach that exploits sensors in consumer-grade smartphones to authenticate a user. Unlike conventional approaches, is based on deep learning to identify user distinct behavior from the embedded sensors with and without the user’s interaction with the smartphone. We investigate different deep learning architectures in modeling and capturing users’ behavioral patterns for the purpose of authentication. Moreover, we explore the sufficiency of sensory data required to accurately authenticate users. We evaluate on a real-world dataset that includes sensors data of 84 participants’ smartphones collected using our designed data-collection application. The experiments show that operates accurately using readings of only three sensors (accelerometer, gyroscope, and magnetometer) with a high authentication frequency, e.g., one authentication attempt every 0.5 seconds. Using sensory data of one second enables an authentication F1-score of approximately 98%, FAR of 0.95%, FRR of 6.67%, and EER of 0.41%. While using sensory data of half a second enables an authentication F1-score of 97.52%, FAR of 0.96%, FRR of 8.08%, and EER of 0.09%. Moreover, we investigate the effects of using different sensory data at variable sampling periods on the performance of the authentication models under various settings and learning architectures.

The intensive care unit (ICU) admits the most seriously ill patients requiring extensive monitoring. Early ICU mortality prediction is crucial for identifying patients who are at great risk of dying and for providing suitable interventions to save their lives. Accordingly, early prediction of patients at high mortality risk will enable their provision of appropriate and timely medical services. Although various severity scores and machine-learning models have recently been developed for early mortality prediction, such prediction remains challenging. This paper proposes a novel stacking ensemble approach to predict the mortality of ICU patients. Our approach is more accurate and medically intuitive compared to the literature work. Data were prepared and feature selection was processed under the supervision of the ICU domain expert. The data were split into six modalities based on the expert’s decisions. For the prediction process, a separate classifier was selected for each modality based on the performance of the classifiers. We utilized the most popular and diverse classifiers in the literature, including linear discriminant analysis, decision tree (DT), multilayer perceptron, k-nearest neighbor, and logistic regression (LR). Then, a stacking ensemble classifier was constructed and optimized based on the fusion of these five classifier decisions. The framework was evaluated using 10,664 patients from the medical information mart for intensive care (MIMIC III) benchmark dataset. To predict patient mortality, extensive experiments were conducted using the patients’ time series data of different lengths. For each patient, the first 6, 12, and 24 hours of the first stay were tested. The results indicate that our model outperformed the state-of-the-art approaches in terms of accuracy (94.4%), F1 score (93.7%), precision (96.4%), recall (91.1%), and area under the receiver operator characteristic (ROC) curve (93.3%). These results demonstrate the ability and efficiency of our approach to predict ICU mortality.

Multi-label classification (MLC) is considered an essential research subject in the computer vision field, principally in medical image analysis. For this merit, we derive benefits from MLC to diagnose multiple grades of diabetic retinopathy (DR) from various colored fundus images, especially from multi-label (ML) datasets. Therefore, ophthalmologists can detect early signs of DR as well as various grades to initiate appropriate treatment and avoid DR complications. In this paper, we propose a comprehensive ML computer-aided diagnosis (CAD) system based on deep learning technique. The proposed system's main contribution is to detect and analyze various pathological changes accompanying DR development in the retina without injecting the patient with dye or making expensive scans. The proposed ML-CAD system visualizes the different pathological changes and diagnoses the DR grades for the ophthalmologists. First, we eliminate noise, enhance quality, and standardize the sizes of the retinal images. Second, we differentiated between the healthy and DR cases by calculating the gray level run length matrix average in four different directions. The system automatically extracts the four changes: exudates, microaneurysms, hemorrhages, and blood vessels by utilizing a deep learning technique (U-Net). Next, we extract six features, which are the gray level co-occurrence matrix, areas of the four segmenting pathology variations, and the bifurcation points count of the blood vessels. Finally, the resulting features were afforded to an ML support vector machine (SVM) based on a classifier chain to differentiate the various DR grades. We utilized eight benchmark datasets (four of them are considered ML) and six different performance evaluation metrics to evaluate the proposed system's performance. It achieved 95.1%, 91.9%, 86.1%, 86.8%, 84.7%, 86.2% for accuracy, area under the curve, sensitivity, specificity, positive predictive value, and dice similarity coefficient, respectively. The experiments show encouraging results as compared with other systems.