Probabilistic Determination Of Down's Syndrome Using Machine Learning Techniques

Down's Syndrome (DS) is a genetic birth disorder that occurs in fetuses, which is typically associated with physical growth delays, mild to moderate mental disability, characteristic facial features and a short life expectancy rate. The paper focuses on extracting markers from the Prenatal Screening (first trimester test) reports of patients. Few of the attributes are crown rump length, nuchal translucency, age, smoking habits, history of previous pregnancies with Trisomy 21 and presence of the nasal bone, extracted using Optical Character Recognition. The collection of reports is subjected to various clustering techniques like K-means, K-medoids, DBSCAN (Density-based spatial clustering of applications with noise), Hierarchical clustering in order to deal with the problem of lack of labels of training data-set since it is bound by confidentiality and ethical issues. The paper is pivoted around the novel technique of handling skewed data-set using ADASYN (Adaptive Synthetic Sampling Approach) over-sampling approach to overcome class bias because the data set in hand has only rare cases of `High Risk' DS. The reports are further subjected to ML (Machine Learning) ensemble of supervised learning methodologies like Naive Bayes, Random Forest and ANN (Artificial Neural Network) to determine the posterior probability of the foetus suffering from Down's Syndrome. The paper envisions building an architecture that can be used as a tool to aid gynecologists by providing numerical probability measure that is representative of the Down Syndrome risk in the foetus and on the basis of this value, the gynecologist can recommend whether the patient should take up further invasive [amniocentesis and CVS (Chorionic Villus Sampling)] tests that are clear indicators of DS. The paper utilizes a live Indian data set and hence is reflective of the current scenario of DS in India. The paper aims to provide a groundbreaking idea in the field of prenatal diagnostics using ML methodologies ensemble.