Predição do não-comparecimento de pacientes em uma clínica de diagnóstico por imagem usando aprendizado de máquina
The objective of this work is to apply and analyze the performance of Machine Learning models for predicting patient no-shows at a diagnostic imaging clinic, using data from 2015 to 2023 from two units of Clínica Radiológica de Anápolis (CRA), in Anápolis, Goiás, Brazil. The relevance of this study...
| Autor: | |
|---|---|
| Tipo de documento: | dissertação |
| Estado: | Versão publicada |
| Data de publicação: | 2025 |
| País: | Brasil |
| Recursos: | Universidade Federal de Goiás (UFG) |
| Repositório: | Repositório Institucional da UFG |
| Idioma: | português |
| OAI Identifier: | oai:repositorio.bc.ufg.br:tede/14632 |
| Acesso em linha: | https://repositorio.bc.ufg.br/tede/handle/tede/14632 |
| Access Level: | Acceso aberto |
| Palavra-chave: | SHAP (SHapley Additive exPlanations) Aprendizado de máquina Clínica Não-comparecimento Otimização Machine learning Clinic No-show Optimization ENGENHARIAS::ENGENHARIA ELETRICA |
| Resumo: | The objective of this work is to apply and analyze the performance of Machine Learning models for predicting patient no-shows at a diagnostic imaging clinic, using data from 2015 to 2023 from two units of Clínica Radiológica de Anápolis (CRA), in Anápolis, Goiás, Brazil. The relevance of this study is based on the possibility of building a final application and on the recurrence and negative impact of patient no-shows in health centers, requiring methods to optimize the use of clinical resources and reduce financial and efficiency losses. The procedure modalities considered in this work were Magnetic Resonance Imaging, Computed Tomography, consultations, and Ultrasound. The collected data included patient age, patient gender, patient no-show history, scheduling details (date and time), procedure type, distance from the patient’s registered address to the clinic, among others. The tested models, Logistic Regression, Multilayer Perceptron, XGBoost, LightGBM, and CatBoost, underwent hyperparameter tuning and probability threshold adjustment based on the Precision-Recall curve area and a customized "Cost" metric. The SHAP framework was used for interpreting the predictions. Comparisons with the literature indicated the agreement of the obtained results and the potential of the methods in this work to serve as a no-show prediction solution for optimizing tasks such as overbooking. The analysis using the SHAP framework, in turn, was able to highlight the most influential variables in the probability of no-show for different procedure modalities, reinforcing the utility of this method for identifying actionable variables. |
|---|