Development and evaluation of a recommendation system to suggest medical items for inpatients in Hospital da Luz Lisboa

• Autor(a) principal: Cabral, Miguel Alexandre Garcia
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10773/37347
• topic: Recommendation systems; Collaborative filtering; Health domain; Medical items; Inpatients

The internet advances have led to an increase of data and information availability. This overload of information tends to compromise the capacity to manage and filter the available data. In the health domain, the increasing digitalization in healthcare led to a substantial rise of the recorded data. Various recommendation systems (RS) have been developed to help healthcare professionals integrate all information and make efficient and effective decisions. Here, a preliminary RS based on collaborative filtering is proposed to reduce the time that healthcare professionals spend in registering medical items consumed during patients’ hospitalization. For that purpose, the RS was built to perform suggestions of the medical items and respective quantities needed in the first day of hospitalization of a patient. Data regarding the diagnostics, surgical procedures and medical item records associated to surgeries of inpatients during a period of one year in Hospital da Luz Lisboa was filtered, restructured, and analysed (N = 5088 surgeries) for the construction of the RS. A 75-25% split of the data was considered with a 4-fold cross-validation procedure applied on the train set to tune the hyperparameters settings for the algorithm. The RS was then tested and evaluated regarding its overall performance in terms of accuracy, classification performance, and coverage. The same measures were applied to assess the quality of the recommendations for each medical specialty of the hospital. Furthermore, the trust of healthcare professionals in the RS was also assessed. A moderate overall performance was achieved (precision = 0.608, recall = 0.729, F1-Measure = 0.663, RMSE = 6.901) and the quality of the algorithm’s recommendations varied between medical specialties. Additionally, the algorithm presented higher values of precision, recall and F1-Measure in the predictions of the most frequently registered medical items in the test set, which corresponded to approximately 85% of the consumptions in the first day of hospitalization. Regarding the coverage of the RS, approximately 80% of the medical items used in the test set were never recommended by the algorithm, corresponding to only 5.57% of the consumptions. Lastly, although in the point of view of hospital’s nurses there is some trust in the RS results, several suggestions were given for further improvements of the algorithm. Despite the limitations of the RS, the observed results represent a starting point for the development of a tool that can support healthcare professionals of Hospital da Luz Lisboa in registering medical items needed during inpatients’ hospitalization.