Machine-Learning-Derived Model for the Stratification of Cardiovascular risk in Patients with Ischemic Stroke.

Background Stratification of cardiovascular risk in patients with ischemic stroke is important as it may inform management strategies. We aimed to develop a machine-learning-derived prognostic model for the prediction of cardiovascular risk in ischemic stroke patients. MATERIALS AND METHODS: Two prospective stroke registries with consecutive acute ischemic stroke patients were used as training/validation and test datasets. The outcome assessed was major adverse cardiovascular event, defined as non-fatal stroke, non-fatal myocardial infarction, and cardiovascular death during 2-year follow-up. The variables selection was performed with the LASSO technique. The algorithms XGBoost (Extreme Gradient Boosting), Random Forest and Support Vector Machines were selected according to their performance. The evaluation of the classifier was performed by bootstrapping the dataset 1000 times and performing cross-validation by splitting in 60% for the training samples and 40% for the validation samples. RESULTS: The model included age, gender, atrial fibrillation, heart failure, peripheral artery disease, arterial hypertension, statin treatment before stroke onset, prior anticoagulant treatment (in case of atrial fibrillation), creatinine, cervical artery stenosis, anticoagulant treatment at discharge (in case of atrial fibrillation), and statin treatment at discharge. The best accuracy was measured by the XGBoost classifier. In the validation dataset, the area under the curve was 0.648 (95%CI:0.619-0.675) and the balanced accuracy was 0.58 ± 0.14. In the test dataset, the corresponding values were 0.59 and 0.576. CONCLUSIONS: We propose an externally validated machine-learning-derived model which includes readily available parameters and can be used for the estimation of cardiovascular risk in ischemic stroke patients.

On the evaluation of deep learning interpretability methods for medical images under the scope of faithfulness.

BACKGROUND AND OBJECTIVE: Evaluating the interpretability of Deep Learning models is crucial for building trust and gaining insights into their decision-making processes. In this work, we employ class activation map based attribution methods in a setting where only High-Resolution Class Activation Mapping (HiResCAM) is known to produce faithful explanations. The objective is to evaluate the quality of the attribution maps using quantitative metrics and investigate whether faithfulness aligns with the metrics results. METHODS: We fine-tune pre-trained deep learning architectures over four medical image datasets in order to calculate attribution maps. The maps are evaluated on a threefold metrics basis utilizing well-established evaluation scores. RESULTS: Our experimental findings suggest that the Area Over Perturbation Curve (AOPC) and Max-Sensitivity scores favor the HiResCAM maps. On the other hand, the Heatmap Assisted Accuracy Score (HAAS) does not provide insights to our comparison as it evaluates almost all maps as inaccurate. To this purpose we further compare our calculated values against values obtained over a diverse group of models which are trained on non-medical benchmark datasets, to eventually achieve more responsive results. CONCLUSION: This study develops a series of experiments to discuss the connection between faithfulness and quantitative metrics over medical attribution maps. HiResCAM preserves the gradient effect on a pixel level ultimately producing high-resolution, informative and resilient mappings. In turn, this is depicted in the results of AOPC and Max-Sensitivity metrics, successfully identifying the faithful algorithm. In regards to HAAS, our experiments yield that it is sensitive over complex medical patterns, commonly characterized by strong color dependency and multiple attention areas.

Speech Based Affective Analysis of Patients Embedded in Telemedicine Platforms.

Speech is a basic means of human expression, not only due to the combination of words that exits our mouth, but also because of the different way we express these words. Apart from the main objective of speech, which is the communication of information, emotions flow in human speech as various vocal characteristics (prosodic, spectral, tonal). By processing these characteristics, Speech Emotion Recognition aims to analyze and assess the human emotional status to complement medical data captured during telemedicine sessions. Driven by the latest developments in Computer Vision concerning Deep Learning techniques, EfficientNets are exploited to extract features and classify imagery representations of human speech into emotions as a web service along with an interpretation scheme. The developed web service will be consumed during video conferences between medical staff and patients for the near real-time assessment of emotional status of patients during video teleconsultations.

Creating Visual Vocabularies for The Retrieval And Classification of Histopathology Images.

State-of-the-art technologies in the fields of computer vision and machine learning led the automatic recognition of malignant structures in histopathology images. More than often, such structures are reported to be found in glands, where different morphological characteristics indicate the existence of a variety of adenocarcinomas, including prostate, breast, lung and colon cancer. Classification of images containing glandular representations in different cancer types can be performed in the whole image by the utilization of a combination of local and global features. The proposed methodology involves the exploitation of a notion utilized often in text mining called Bag of Words and employed in the service of Computer Vision with the name of Bag of Visual Words (BOVW) for the development of a retrieval and classification system for pathology images. The paper discusses the technical details of implementation, the enhancement of the BOVW technique, while some initial results using public datasets are presented.