Application of machine learning techniques for a recommendation system in pharmacy

• Autor(a) principal: Torres, Beatriz Freitas
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10400.22/24708
• topic: Recommendation system; Community pharmacy; Non-prescription products; Products clustering; Products interactions

Community Pharmacy (CP) plays a crucial role in the population, improving patients’ quality of life and minimising medication risks. In Portugal, CPs dispense prescription and non-prescription products. Pharmacy professionals have an added responsibility when advising non-prescription products and should pay attention to self-medication and possible interactions. Therefore, a product recommendation system that incorporates relevant information about the products supports a more informed recommendation by the professional. Although there are a few studies in the area of medication RS, they are still scarce, and to the best of our knowledge, no medication RS is applied in community pharmacies in Portugal. This work aims to develop a conceptual pharmaceutical product recommendation framework and identify relevant groups of products according to their characteristics and experts’ opinions. The specific objectives consist of describing recommendation systems in pharmacy, defining and comparing distance functions capable of creating groups of similar and clinically relevant products for pharmaceutical counselling, applying machine learning techniques and comparing them, and communicating the results. For this purpose, the background of pharmaceutical products counselling without a prescription was analysed. Public databases were selected to be included in the conceptual framework, and the data obtained was processed. Therefore, a database was obtained with 1426 products (over-the-counter medication, homoeopathic medication, and dermocosmetics) and their clinical and scientific information. In order to identify relevant groups of products, seven hierarchical (single linkage, complete linkage, average linkage, median linkage, centroid linkage, and ward linkage) and non-hierarchical (K-means) clustering techniques were applied and evaluated. Dendrograms, the Calinski-Harabasz score, silhouette score, Davies-Bouldin score and the inflexion point method were used to determine the ideal number of clusters for each technique and evaluate its validity. An experts consultation was performed to define a distance function aligned with pharmaceutical counselling. This consultation allowed the identification of the importance of the variables in the distance function definition. The resultant data was analysed in Microsoft Excel, SPSS and Python with the libraries Pandas, Natural Language Toolkit (NLTK), Unidecode, Plotly, Matplotlib, NumPy, SciPy, and Scikit-learn, using Spyder IDE. Twenty-two groups of similar products were formed with K-means, the most effective clustering approach for forming pharmacologically homogeneous groups. However, the obtained clusters did not present enough clinical relevance to support professionals during counselling. Consequently, a new distance function was defined, enhancing the importance of the pharmacotherapeutic group of the products and aligned with the results obtained in the experts’ consultation. Twenty-four groups of similar products were formed with K-means, which was once again the technique that presented pharmacologically homogeneous groups, based mainly on safe use during pregnancy and breastfeeding and pharmacotherapeutic group. The remaining clustering techniques, non-hierarchical techniques, did not present pharmacologically homogeneous groups with any of the distance functions.