AI Case Study
University of Southern California researchers determine an efficient combination of tests to accurately predict FASD in children
Researchers from several universities including University of Southern California applied machine learning to a series of tests designed to determine fetal alcohol spectrum disorder. They establish an efficient cost-effective screening protocol by investigating different combinations of tests used to predict whether a child has FASD with an acceptable level of accuracy.
Healthcare Providers And Services
The research puts forth "a machine learning framework to address the various outcome measures that are used to quantify deficits across multiple domains in children/youth with FASD, and to use these measures to differentiate the FASD group from typically developing controls. We utilized data from eye movement behaviors, psychometric test scores, and DTI of the brain to construct a new, multimodal classifier that demonstrates high performance in identifying the clinical population... A relatively simple classifier, the LR classifier, was used for our multiple assessments data. This classifier is easier to train, requiring much less computing time and resources, and can generalize better compared to other complicated classifier."
"Fetal alcohol spectrum disorder (FASD) is the most common preventable developmental disorder, resulting from prenatal alcohol exposure... Early diagnosis of FASD is important in that it can lead to early interventions that reduce the risk of developing secondary disabilities (7). Despite the high prevalence of FASD, the clinical diagnosis can be both challenging and time consuming. Therefore, an easy, objective, and effective procedure which can assess the deficits and differentiate the neurological groups is needed as a screening tool for children at risk for FASD."
"The classification accuracy based on data from all assessments under the iterative train-test procedure reached 84.78 with 52.17% as the chance level (naïve Bayes), which was an 11.4% improvement of the best single assessment accuracy... classifiers trained on combinations with psychometric tests data still performed the best (with no significant difference between combinations), followed by classifiers on AntiSac and natural viewing, and ProSac and natural viewing (no significant difference between these two). The best performance was achieved with the combination of data from four assessments (ProSac, AntiSac, Psychometric and Natural Viewing)."
"The probability of a participant being identified as having FASD predicted from each assessment was concatenated as input for training a LR classifier. To statistically compare the performance of different classifiers, we first computed the performance variances by repeating the training and testing process 20 times for each classification method with a resampled training set.
Classification analysis performed on 46 participants who completed 6 different assessment tasks.