23/11/2021 - 09:00 - 10:10
COC 01 - NOVOS MÉTODOS DE EXPLORAÇÃO DE DADOS COM BASE EM INOVAÇÕES TECNOLÓGICAS |
33529 - PREDICTING ATTRITION IN A COHORT OF VERY PRETERM INFANTS USING CLASSIFICATION ALGORITHMS
RAQUEL MARA TEIXEIRA - EPIUNIT – INSTITUTO DE SAÚDE PÚBLICA, UNIVERSIDADE DO PORTO, RUA DAS TAIPAS, Nº 135, 4050-600 PORTO, PORTUGAL, CARINA RODRIGUES - EPIUNIT – INSTITUTO DE SAÚDE PÚBLICA, UNIVERSIDADE DO PORTO, RUA DAS TAIPAS, Nº 135, 4050-600 PORTO, PORTUGAL, HENRIQUE BARROS - EPIUNIT – INSTITUTO DE SAÚDE PÚBLICA, UNIVERSIDADE DO PORTO, RUA DAS TAIPAS, Nº 135, 4050-600 PORTO, PORTUGAL, RUI CAMACHO - FACULDADE DE ENGENHARIA DA UNIVERSIDADE DO PORTO, RUA DR. ROBERTO FRIAS,4200-465 PORTO, PORTUGAL
Objective: To develop predictive models of attrition in a population-based cohort applying different classification algorithms.
Methods: The study sample consisted of 542 very preterm (<32 gestational weeks) infants born in two regions of Portugal and enrolled during 2011-2012 as part of the European Effective Perinatal Intensive Care in Europe (EPICE) population-based cohort. Perinatal data were abstracted from medical charts to predict attrition (non-participation) at follow-ups performed at 1, 2, 3, and 4 years old. A set of sociodemographic and clinical characteristics were selected as possible predictors. Eight classification algorithms were used to build the models: Ada Boost, Artificial Neural Networks, Functional Trees, Consolidate Decision Trees (J48), Decision Trees (J48), K-Nearest Neighbours, Logistic Regression and Random Forest (RF). Performance was compared using Accuracy, Recall (sensitivity), F- measure, and AUC- PR (area under Precision-Recall curve).
Results: Attrition in the four Portuguese follow-ups was, respectively: 16%, 25%, 13%, and 17%. The AUC-PR metrics ranged from 70.9 to 94.1 among all the methods. RF presented the best performance for predicting attrition in all of the four follow-ups [AUC-PR1: 94.14 (2.0); AUC-PR2:92.12 (1.89); AUC-PR3: 92.9 (2.17); AUC-PR4: 93.4 (2.6)]. Logistic regression presented only a fair performance [AUC-PR1: 78.82 (3.44); AUC-PR2: 78.92 (1.89); AUC-PR3: 81.14 (2.02); AUC-PR4: 80.6 (3.78)]. The most relevant predictors, identified by RF, were common for all four follow-ups: birth weight, gestational age, maternal age, and days of hospitalization after birth.
Conclusion: Random forest, a machine learning algorithm, presented the highest capacity to distinguish classes and predict attrition, surpassing usual logistic regression applied in epidemiological research.
|
