Μεταπτυχιακές εργασίες / Master Theses

Μόνιμο URI για αυτήν τη συλλογή

https://apothesis.hmu.gr/handle/20.500.12688/9246

Περιηγούμαι

Πλοήγηση Μεταπτυχιακές εργασίες / Master Theses ανά Θέμα "Adverse drug reaction"

Τώρα δείχνει 1 - 1 of 1
  • Μικρογραφία εικόνας
    Τεκμήριο
    Developing an LSTM-based framework for assessing the risk of adverse drug reactions in polypharmacy patients using multimodal data
    (ΕΛΜΕΠΑ, Σχολή Μηχανικών (ΣΜΗΧ), ΠΜΣ Μηχανικών Πληροφορικής, 2026-03-05) Droumalia, Fotini; Δρουμαλιά, Φωτεινή; Tsiknakis, Emmanouil; Τσικνάκης, Εμμανουήλ
    Adverse drug reactions (ADRs) are harmful, unexpected responses to medications at standard therapeutic doses, while adverse drug events (ADEs) more broadly include ADRs as well as harm from overdoses or prescription errors. In this study, the two terms are treated interchangeably. The prediction of ADRs is crucial in clinical practice since these events can cause serious complications and increase healthcare costs, yet they are often preventable. Advances in electronic health records and health informatics have increased access to large, real-world clinical datasets, often including both static and longitudinal patient information. Multimodal approaches, which integrate multiple data types, can extract complementary knowledge from each data modality, while deep learning models are particularly well-suited to uncover hidden patterns and complex relationships. However, even with powerful techniques, it is essential to frame the prediction problem according to its specific challenges. Additionally, harmful events such as ADRs are generally rare, which can create extreme class imbalance in datasets. This master’s thesis develops a predictive model for personalized ADR risk assessment using UK Biobank data, incorporating genetic information, drug prescriptions, diagnoses, and demographic factors that were carefully organized to ensure the problem was accurately formulated. A multimodal Long Short-Term Memory (LSTM) model integrates all data modalities separately for prediction, handling ADR prediction as a binary classification task. During the study, we addressed the class imbalance using algorithm-level techniques rather than data-level approaches, preserving the natural distribution of real-world data. The study highlights the difficulty of predicting rare ADRs, with only 0,39% of over 2 million samples being positive. Additional limitations included incomplete medication data and missing patient-level features, which affected model effectiveness. Nonetheless, the model can detect roughly 4 in 10 ADR cases without overpredicting positives, demonstrating its potential. These findings underscore the value of richer data and reframing ADR prediction as an anomaly detection task rather than simple binary classification, which better suits the identification of rare events and handling extreme imbalance, ultimately improving predictive performance and patient safety.