Percorrer por autor "Carvalho, Rafael Reis de"
A mostrar 1 - 1 de 1
Resultados por página
Opções de ordenação
- A topic modelling-based recommender system for drugs using user experience reviews [TopicDrugRec]Publication . Carvalho, Rafael Reis de; Pato, Matilde Pós-de-Mina; Datia, Nuno Miguel SoaresAbstract The increasing volume of patient-reported data, alongside the rise of personalised medicine has made it challenging for healthcare professionals to incorporate patient experiences into their clinical decision-making due to information overload and demanding working shifts. Much of this data is available in the form of numerical drug ratings, which often fail to capture the complexity of user experiences by lacking contextual information and response bias. To address this, this dissertation proposes TopicDrugRec, a drug recommender system based on topic modelling and trained on the UCI ML Drug Review dataset, designed to support clinicians in providing safer and more personalised drug prescriptions. It follows a six step methodology: first, exploratory data analysis, data cleaning, followed by sentiment analysis to mitigate rating bias, topic modelling to extract latent themes from patient reports in the form of free text, integration of medical knowledge (drugdrug interactions, side effects and contraindications) to enhance patient safety, and the implementation of a web application and performance evaluation. The recommendation algorithm was designed to incorporate topic similarity, user sentiment, and perceived usefulness, allowing for tunable hyperparameters to generate the recommendations. Three topic modelling approaches were evaluated: Latent Dirichlet Allocation, Non-negative Matrix Factorization, and BERTopic. The evaluation showed semantic similarity, derived from topic modelling, to be the most influential factor in recommendation quality. Additionally, grouping medical conditions into ICD-11 categories mitigated dataset imbalanced and improved coverage, with the NMF-based model achieving the best performance on this setup, with a Precision@10 of 0.513 and Mean Reciprocal Rank @10 of 0.676. Despite being a proof-of-concept, these findings demonstrate TopicDrugRec’s potential in reducing information overload, enhancing medic-patient interaction and integrating patient feedback into data-driven decision-making. Additionally, it lays foundation for future work, including real world validation, curating more complex datasets with patient information, and providing explainable recommendations.