Back to search

University of West Attica

Αλγόριθμοι αυτοματοποιημένης επεξεργασίας για δομημένα, οπτικά και αισθητηριακά δεδομένα με χρήση μηχανικής μάθησης

Abstract

dc:description

Recent advances in computer vision and human behavior analysis, leveraging machine learning, have opened new avenues for the development of applications in domains such assecurity, civil protection, and healthcare. These applications include, among others, the monitoring of patients with cognitive impairments and the prevention of complications in various medical contexts. The continuous evolution of computing hardware, combined with the development of high-performance computing systems, has enabled the implementation of advanced machine learning methods across a wide range of fields. We are now at the threshold of a new era, where technology can significantly contribute to improving quality oflife. These systems, along with their current and future applications, can be monitored, optimized, and adapted to meet modern social needs. The present doctoral dissertation focuses on two main research areas. In the first part of the dissertation, an algorithm for predicting human behavior through bodyposture analysis is presented. Using the MediaPipe Holistic Model, a system was developed for detecting defensive and aggressive body postures based on machine learning. Experimental results demonstrated that the system can operate efficiently on both high-performance computers and low-power computing devices. Following this exploratory phase, during which an effective and functional methodology for body posture analysis was established, the research expanded toward developing a prototype model for the detection of aggressive behaviors. This model served as a dynamic framework for predicting aggressive episodes in patients with dementia. It incorporated a broader range of data and integrated external datasets to validate and strengthen the findings derived from the initial experimental study. Building upon the methodological knowledge acquired in the previous stage, a multimod alapproach for the detection and prediction of aggressive behavior was developed, based on human features such as facial expressions, body posture, and limb movements. Detecting and predicting aggressive behavior in patients with dementia represents a major challenge for healthcare professionals and caregivers. In this dissertation, a behavior classification model is presented, which builds upon the MediaPipe Holistic Model and utilizes data from gestures, posture, and facial expressions to detect conflict episodes. Experimental results indicated thatthe Random Forest Classifier achieved the best performance, laying the groundwork for future development of predictive models of aggression. Furthermore, the research was extended by incorporating acoustic features into the detection of aggressive behavior. Using a multimodal framework, the audio-based model was trained on vocal expression data corresponding to both aggressive and non-aggressive behaviors. Through a late fusion approach, the outputs from the visual and auditory models were merged into a meta-classifier, achieving improved accuracy in detecting aggressive incidents. Acomparison between early fusion and late fusion strategies was conducted to evaluate the suitability of each method for predicting aggressive behavior. Results showed that late fusion outperformed in terms of accuracy, sensitivity, and inference time, whereas early fusion achieved higher precision, making it more appropriate for applications where minimizing false positives is crucial. In the second part of the dissertation, the focus shifts to predicting antipsychotic dosage infirst-episode psychosis (FEP) patients using multimodal data. This phase centers on developing machine learning models capable of predicting the required antipsychotic dosage based on data from the Athens First Episode Research Study. Prediction models were developed using clinical variables, the Exposome Score, and neuropsychological data. The analysis revealed that incorporating environmental and cognitive variables improved predictive accuracy, making the model useful for personalized treatment planning. Patients receiving lower doses exhibited higher remission rates, while those with non-affective psychoses required higher doses. These findings highlight the potential clinical utility of multimodal machine learning methods in psychiatric practice. The significance of this dissertation lies in advancing the application of machine learning in improving accuracy and efficiency across critical domains related to quality of life, while opening new directions for its practical implementation. Moreover, it promotes the development of multimodal approaches for the detection and prediction of human behavior, with potential applications in safety, healthcare, and personalized treatment. The integration of visual, auditory, and clinical data provides a novel perspective for predicting aggressive behavior and tailoring therapies for mental disorders, laying the foundation for further research and clinical applications. The combination of human behavior analysis and clinical modeling underscores the scientific value of this research, while opening new pathways for extending artificial intelligence applications in real-world contexts.

Degree

thesis:*
Grantor dc:publisher
University of West Attica
Year dc:date
2025

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Galanakis, Ioannis
  • Γαλανάκης, Ιωάννης

Subjects

dc:subject × 20

Rights

Language dc:language
gre

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/61047

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Galanakis, Ioannis; Γαλανάκης, Ιωάννης. Αλγόριθμοι αυτοματοποιημένης επεξεργασίας για δομημένα, οπτικά και αισθητηριακά δεδομένα με χρήση μηχανικής μάθησης. University of West Attica, 2025. http://hdl.handle.net/10442/hedi/61047