Clinical Validation of AI/ML Based Model for Down Syndrome Detection Through Graphical Analysis of Facial Dysmorphic Features
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background and objective
Down Syndrome is associated with high mortality in India, due to non-diagnosis/ late-diagnosis resource constraints. The objective of the present study was to test Clinical Validity of the Google Cloud AutoML Vision Image Classification Down Syndrome Detection model in real-life situations – i.e., does this model work in the hands of laymen (parents, guardians, end-user) with the same accuracy (98%+ in a controlled experiment).
Methods
This multi-label cross-sectional study was conducted in the Neonatal and Paediatric unit of a tertiary care hospital between August 2021 and September 2022. The participants consisted of 104 children aged 5 days - 18 years of two categories - those exhibiting facial dysmorphic features indicative of Down Syndrome on visual inspection, and those not exhibiting said features. Images were collected after written informed consent using the Down Syndrome Detection application. The outcomes recorded were the efficiency of the model in detection of Down Syndrome based of the images collected.
Results
The CloudML model trained with 104 images initially achieved: Sensitivity - 100%, specificity - 80%, Average Precision - 96.6%, precision - 86.67%, and Recall - 92.86% (Precision and Recall are calculated at a confidence threshold of 0.5) This Indo-specific Machine Learning model, specifically trained and tested on Indian children, shows remarkable accuracy in the diagnosis of Indian Down Syndrome positive neonates. On adjustment of software parameters (the confidence threshold of prediction), the technique can deliver highly accurate Down Syndrome diagnosis with a 100% Sensitivity, at the expense of false positives that may be ruled out through further confirmatory testing.
Conclusions
This diagnostic algorithm is a reliable preliminary postnatal screening tool and can be deployed in resource-limited settings where genetic testing is neither affordable nor readily accessible. False positives can be ruled out through subsequent testing.
