12 August 2020
10:00 Master's Defense Fully distance
Theme
An Interpretable Machine Learning Approach for Retrospective Assessment of the Impact of Phlebotomy-Derived Blood Loss on Inpatients
Student
Flávia Érika Almeida Giló Azevedo
Advisor / Teacher
Andrà © Santanchè
Brief summary
In medicine, the procedure for drawing blood from a patient for laboratory tests is called phlebotomy. Inpatients are submitted to phlebotomy protocols that include successive requests for exams, at rates that can reach daily collections, in order to diagnose or monitor the evolution of their clinical conditions. However, frequent collections expose the patient to a high loss of blood in a short period of time, which can trigger problems for his health, since the time between collections can be insufficient for the adequate replacement of the lost blood components. With the objective of evaluating the hypothesis of phlebotomy blood loss to be a relevant factor for the development or worsening of anemia and for an increase in the length of hospital stay, the present research conducted analyzes guided by retrospective clinical, laboratory and demographic data of 28.312 hospitalizations that occurred at the Hospital de Clínicas da Unicamp (HC / Unicamp) between the years 2012 and 2017. In an initial phase, exploratory investigations were conducted to capture the main characteristics inherent to the data set. Then, an interpretable Machine Learning (AM) approach was developed, using the techniques based on Gradient Boosting Machines and Random Forests Decision Trees, for solving regression tasks, binary classification and multiclass classification. In order for the computational solutions produced to offer a convenient degree of interpretability for use in the medical-hospital setting, methods were used to extract interpretability from the models of BF, in order to express the relationship between the different variables used, as well as to list the degree of contribution of them for the results of the models. The results suggest a relationship between the amount of blood samples taken during hospitalization and the length of hospital stay, as well as the decline in hemoglobin levels in the blood of patients, a factor directly related to the development of anemia. The study conducted with unpublished data, as far as is known, is the first to jointly investigate the relationships between phlebotomy, anemia and length of hospital stay using BF. In addition, it deepens the existing knowledge of the problem by expanding related investigations found in the literature, since it takes into account variables adjacent to the possibility of developing anemia in the course of hospitalization, such as blood transfusions and surgeries performed. It also offers HC / Unicamp subsidies for evaluating its phlebotomy processes and grounds for decision making.
Examination Board
Headlines:
| André Santanchè | IC / UNICAMP |
| Hélio Pedrini | IC / UNICAMP |
| César Alex de Oliveira Galoro | Sabin Group Diagnostic Medicine |
Substitutes:
| Julio Cesar dos Reis | IC / UNICAMP |
| Paula Dornhofer Paro Costa | FEEC / UNICAMP |