Prediction of Low Birth Weight in Infants via Artificial Intelligence Techniques without Using Sonographic Measurements

• Main Authors: Mahtab Farahbakhsh, Hamid Reza Marateb, Marjan Mansourian, Masoomeh Goodarzi-Khoigani
• Format: Article
• Language: fas
• Published: Vesnu Publications 2015-05-01
• Series: مجله دانشکده پزشکی اصفهان
• Subjects: Computer-aided medical diagnosis; Leave-one-out cross validation; Low birth weight; Sequential feature selection
• Online Access: http://jims.mui.ac.ir/index.php/jims/article/view/4556

Background: Birth weight is probably the most important factor affecting neonatal mortality and morbidity. Compared with normal weight infants, low-birth-weight (LBW) infants may be more at risk for many health problems. The prediction of low birth weight is important as it may cause mental and physical health problems in childhood and adulthood. We assessed a computer-aided diagnosis system to classify infants to low or normal birth weight categories. Methods: In the present study, the association between the low birth weight and the intake of about 40 types of macro- and micronutrients during the first (1st Tr), second (2nd Tr.) and third (3rd Tr.) trimesters was assessed based on demographic and reproductive characteristics, physical activity and nutrients intake in pregnant women. The dataset used in this study contained 526 pregnant women with 95 input features. The used classifiers were k-Nearest Neighbors (kNN), Probabilistic Neural Network (PNN), and two Adaptive Neuro-Fuzzy Classifiers (ANFC-SCG: Scaled Conjugate Gradient, ANFC-LHs: Linguistic Hedges). Also, sequential feature selection (FS) was applied on the low birth weight risk factors to reduce the feature space. Findings: The accuracy of the classifiers kNN, PNN, ANFC-SCG and ANFC-LHs were 48%, 50%, 50% and 50% without feature selection and 93%, 83%, 80% and 83% with feature selection, respectively. Conclusion: Among the tested classifiers, the statistical power and type I error (α) of the best configuration (FS-kNN; k = 3) were 96% and 0.10 in the Leave-One-Out validation framework, showing that the proposed diagnosis system is clinically reliable. Also, using Leave-One-Out cross-validation, the guarding against Type III error was granted.