ABSTRACT
Machine learning system based on fuzzy sets was used for detecting predictors of new-born survival within the first 15 days after birth. The system processed real-life medical data of 566 new-borns; 528 survived and were classified as "alive", the remaining 38 did not survive and were classified as "died". The state of each new-born was described by values of 112 attributes. The system detected five of the attributes as the best predictors for survival (asphyxia, apnoea, birth weight, reanimation and gestation age). To evaluate are the selected attributes relevant for prediction, the following procedure was used: 566 new-borns were divided into two groups by random selection; 396 (70%) randomly selected new-borns formed the learning group, the remaining 170 new-borns formed the test group. The system acquired prediction knowledge by processing data of the learning group new-borns. Using the knowledge thus acquired, the system predicted survival for each new-born from the test group several times, each time using another set of attributes: once, using all 112 attributes; once, using attributes detected by the system as the best predictors; once, using remaining attributes without the best predictors. After the predictions for all new-borns from the test group had been finished, classification accuracy, sensitivity (accuracy for "alive") and specificity (accuracy for "died") were calculated as measures of prediction success with particular sets of attributes. The results have shown that the best prognostic accuracies were achieved when prediction was done using those attributes which the system detected as the best predictors for new-born survival.
