PIS 2024 Abstracts

Area 1 - Pervasive Information Systems

Full Papers

Paper Nr:	5
Title:	An Improved Stroke Type Detection Approach: Combining Chatbot Voice, Ontology, and LSTM-GRU Methods
Authors:	Mayssa Ben Kahla, Dalel Kanzari, Sana Ben Amor and Sonia Ayachi Ghannouchi
Abstract:	Detecting stroke type is important as it ranks second globally in causing dementia after Alzheimer disease. Late recognition often leads to high mortality rates due to misinterpretation. Early identification enables prompt treatment, minimizing neurological damage and reducing long-term disabilities like paralysis or cognitive impairment. The study presents a comprehensive method for the detection of stroke types, considering both unstructured non-clinical data and structured clinical data. It is founded on three key modules: data collection, ontology-based knowledge database modeling, and stroke type detection. In our approach, we seamlessly integrated heterogeneous data, encompassing structured and unstructured formats. Based on ontology and advanced deep learning techniques, we aimed to significantly improve the accuracy and efficiency of stroke-type detection, increasing both accuracy and efficiency. Combining various intelligent algorithms for stroke-type detection leverages their complementary strengths, enhancing diagnostic and prognostic accuracy. This integration reduces misdiagnosis risks, leading to more precise and effective patient care. In the experimental study, we used structured clinical stroke data from Sahloul Hospital, supplemented by patient information from the voice chatbot technique. The use of a combination of voice chatbot, ontology, and deep learning methods for stroke type detection resulted in higher levels of accuracy compared to other existing approaches. These findings highlight the effectiveness of our approach for the detection of stroke types.

Paper Nr:	8
Title:	Hybrid System for Intelligent Context Situation Detection
Authors:	Ikhlass Mastour, Hela Zorgati, Raoudha Ben Djemaa and Layth Sliman
Abstract:	A context-aware Internet of Things system must be able to observe, interpret, and reason the dynamic situations of the environment to provide pertinent information and services to the user. This article proposes an approach structured around three contributions: semantic representation of IoT data, context situations detection, and contextual information dissemination for consumers. The main contribution of this article is represented by a hybrid system for the intelligent detection of the context situation in an IoT environment, which combines machine learning algorithms and approximate logic, more precisely, artificial neural network, case-based reasoning, and fuzzy logic. Finally, we chose a use case in the intelligent transportation sector to validate our approach.

Short Papers

Paper Nr:	6
Title:	Enhancing Multi-View ASD Diagnosis Using Structural MRI and Pretrained CNN
Authors:	Nesrine Zemzemi, Imen Hmida, Nadra Ben Romdhane and Emna Fendri
Abstract:	Autism Spectrum Disorder (ASD) remains challenging to diagnose despite its prevalence. Structural Magnetic Resonance Imaging (sMRI) has emerged as a valuable tool for such tasks. In this paper, we introduce a new multi-view deep learning-based method for ASD diagnosis based on pretrained CNN and majority voting. By combining the information from multiple views, our model can form a more complete understanding of the brain’s anatomy. The majority voting technique further boosts performance by aggregating predictions from multiple instances of the model, ensuring more reliable and accurate classifications. Our method was evaluated on the ABIDE (Autism Brain Imaging Data Exchange) dataset, achieving a notable accuracy of 98.91%.

Paper Nr:	7
Title:	Advanced Multi-View Structural MRI Analysis with Self-Attention for Alzheimer's Disease Detection
Authors:	Safa Hlawa, Nadra Ben Romdhane and Emna Fendri
Abstract:	Accurate diagnosis of Alzheimer's Disease (AD) remains a critical challenge, despite its growing prevalence. Recent advancements in neuroimaging have opened new avenues for enhancing diagnostic precision through deep learning techniques. In this study, we introduce a new method that combines diverse pretrained CNN models and attention mechanisms. We select the most suitable model for each view (axial, coronal and sagittal) after a series of experiments. This method aims to capture a broader range of features from the multi-view input sMRI images, leading to improved classification performance. The method presented here was tested on ADNI dataset. The obtained results prove the efficiency of our proposed method.

Paper Nr:	9
Title:	Pervasive Aided Screening System of Multiple Sclerosis from Retinal OCT Images
Authors:	Sabrin Ouni, Yaroub Elloumi and Raoudha Ben Djemaa
Abstract:	Multiple Sclerosis (MS) is a neurodegenerative disease. As an irreversible disease, the MS screening in early stage is highly recommended. Several works suggested to diagnose MS from the retinal Optical Coherence Tomog-raphy (OCT) images, which implies layer thickness. . However, a delay on MS screening is registered, caused by the unavailability of medical equip-ment’s and the low-rate of medical staff. We propose in this paper a novel method for MS screening from OCT imag-es captured by a pervasive devices. The main challenges is to insure a higher accurate MS screening through a processing workflow executed into smartphone device. For this purpose, we fine tune the convolutive deep neu-ral network “Inception-V4” to extract features from OCT images, which are provided to an Deep Extreme Learning Machine to deduce MS disease. A cross-validation process is conducted where 92.00% accuracy, 94.00% sen-sitivity, 90.00% specificity and 90.38% precision in average are achieved. In addition, the whole method is implemented into a mobile device, where exe-cution time is under one second whatever the OCT image is, which is ade-quate to employ screening on clinical context.