Predição de risco de atraso no desenvolvimento neuropsicomotor nos primeiros mil dias de vida utilizando aprendizado de máquina: coorte brisa

Abstract

Introduction: The first thousand days of life represent a critical window of brain plasticity, in which neuropsychomotor development (NPMD) is sensitive to biological and psychosocial influences. Early prediction of risks during this period is fundamental for timely interventions, although it constitutes a complex challenge in community-based populations. Objective: To develop and evaluate Machine Learning (ML) models to predict the risk of NPMD delay in children during the first thousand days of life. Methods: Prospective study with data from 972 mother-child pairs from the BRISA cohort (São Luís – MA). The outcome was the risk of delay assessed in the cognitive, communication and motor domains evaluated by the Bayley III scale. The algorithms Penalized Logistic Regression (PLR), Random Forest (RF), Support Vector Machines (SVM) with radial kernel and Artificial Neural Networks (ANN), and the null classifier were tested. Preprocessing included imputation of missing data, encoding of categorical variables by one-shot encoding and standardization of numerical variables. For the evaluation of the generated ML models, K-fold cross-validation (k = 5) with 10 repetitions was used. As performances metrics, the area under the AUC-ROC curve (AUC), sensitivity, specificity, accuracy, precision and F1-score were used. The interpretability of the models was analyzed using the SHapley Additive Explanations (SHAP) technique. Results: The highest prevalence of risk of delay occurred in expressive communication (44.1%). The models showed moderate performance with AUC ranging between 0.533 and 0.603, with PLR being the most stable algorithm, reaching an AUC of 0.603 in the fine motor domain. SHAP analysis identified as main predictors: family history of epilepsy, child's sex, maternal depressive symptoms, psychological violence and social support, maternal education and family income. Conclusion: ML models demonstrated moderate discrimination ability in heterogeneous population samples, showing that increased algorithmic complexity did not surpass logistic regression. The findings reinforce the multifactorial nature of development, highlighting the impact of psychosocial and biological determinants, contributing to the development of screening tools applicable to Primary Health Care.