Abstract
Background
Down syndrome (DS) is associated with high mortality in India, due to nondiagnosis/late-diagnosis caused by unavailability of qualified doctors and/or lack of access to expensive medical/diagnostic facilities, especially in rural India. Using artificial intelligence/machine learning graphical pattern recognition tools, relevant facial points can be extracted from children’s photographs, facial anomalies can be identified, and probability of DS affliction can be predicted.
Methods:
Trained Google’s Cloud Vision AutoML Image Classification model was employed with ~2,000 photographs of DS positive children and ~3,000 photographs of DS negative children. A subset of 300 images, 100 each of Asian, Caucasian, and Other-Race children, was used to train and test 3 race-specific models. These results were compared against a unified model trained and tested with same 300 images.
Results:
The CloudML model trained with ~5,000 images initially achieved: Sensitivity—94.6%, specificity—96.9%, and accuracy—96.0%. Upon optimizing confidence threshold to 0.1, model maximized sensitivity at 99.6%, specificity dropped to 93.8%, and accuracy maintained at 96.0%. Each of the race-specific models trained with 100 images each, after optimization, yielded perfect scores on sensitivity, specificity, and accuracy of 100% each. Against this, the unified model with 300 images yielded overall accuracy of 98% (100% sensitivity, 83% specificity for Caucasian children, and 100% sensitivity, 100% specificity for Asian/Other children).
Conclusions:
Post optimization, this model can be used as an effective postnatal screening tool for DS detection. Preliminary results indicate that race-specific models can achieve even higher accuracy, sensitivity, and specificity.
Keywords
Get full access to this article
View all access options for this article.