Readmissão hospitalar potencialmente evitável em população pediátrica : modelos de predição

• Autor(a) principal: Silva, Nayara Cristina da
• Data de Publicação: 2023
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: Repositório Institucional da UFU
• Texto Completo: https://repositorio.ufu.br/handle/123456789/37762; http://doi.org/10.14393/ufu.te.2023.8026
• topic: Aprendizado de Máquina; Readmissão Hospitalar; Regras de Predição Clínica; Pediatria; Fatores de Risco; Machine Learning; Hospital Readmission; Clinical Decision Rules; Pediatrics; Risk factors; CNPQ::CIENCIAS DA SAUDE

Background: Potentially avoidable hospital readmissions (PAHR) are complex events with a negative impact on both the patient and the health system. In the pediatric population, a readmission episode can be even worse, and may impact the development of the child or adolescent, negatively affecting motor, cognitive, emotional and psychosocial development in the short, medium and especially long term. Thus, the development of prediction models, especially using machine learning, has been promising to minimize this outcome. However, studies are still scarce, especially those that can be interpreted with potencial practical application. Objective: To develop 30-day PAHR prediction models for children and adolescents admitted to a tertiary hospital. Methods: A retrospective cohort was performed with data from all pediatric patients (0 to 18 years old) admitted to a tertiary hospital between January 2014 and December 2018 (n: 9,080). Admissions that resulted in death, hospital discharges against medical advice or planned visits/readmissions were excluded. Demographic, clinical, nutritional and biochemical data were collected. For the first manuscript, a prediction model based on a scoring system (HOSPITAL score) was estimated and, subsequently, patients were classified into low, intermediate and high risk groups. In the second manuscript, we used the J48 algorithm and applied leave-one-out cross-validation to develop interpretable decision trees. For the third manuscript, models based on machine learning were built, several algorithms were investigated: classification and regression tree - CART, random forest – RF, gradient boosting machine – GBM, extreme gradient boosting - XGBoost, decision tree - DC e logistic regression - LR. To compare the performance of the models, we computed the area under the receiver operating curve (AUC). Other performance measures were also calculated, such as sensitivity, specificity, Youden’s J index and accuracy, when relevant. Results: The frequency of PAHR in 30 days ranged from 9.5 to 11.70%. The HOSPITAL score showed good discriminatory ability (AUC of 0.80 95% CI 0.77-0.83) for PAHR in a pediatric population (Manuscript 1). To improve the estimates, we used machine learning techniques, aiming to build a predictive model, interpretable and easy to apply in clinical practice. The decision tree (J48 algorithm, applied to 63.6% of new cases) showed that changes in C-reactive protein, hemoglobin and sodium levels and lack of nutritional monitoring were the attributes that contributed to PAHR (Manuscript 2). Finally, the XGBoost algorithm presented the best Youden’s J index. The predictors (XGBoost) were: cancer diagnosis, age, levels of red blood cells, leukocytes, red cell distribution width and sodium, elective admission and multimorbidity (Manuscript 3). Conclusion: The HOSPITAL score can be used for the pediatric population, the XGBoost showed good discrimination and the decision tree, although with less discriminatory potential for PAHR, identified important attributes for clinical practice, especially the lack of nutritional monitoring. Besides, the three prediction models developed offer advantages, since clinical attributes were found that are part of the care routine, easy to get and low cost, allowing to obtain the readmission probability result in real time if implemented in the admission system hospital. Finally, we highlight the importance of improving the quality and specificity of hospital data records, such as nutritional data, to build hospital readmission prediction models.