Automated Diagnoses

A new machine learning algorithm known as DeepGestalt has been developed that can identify genetic syndromes better than doctors, simply by analyzing a person’s face. By adding this automated capability to a clinician’s tool kit, earlier diagnosis and treatment could lead to improved quality of life for many children.

The DeepGestalt algorithm was “trained” on a dataset of more than 150,000 patients. The first step it takes is to identify facial landmarks such as eyes, nose, and mouth. The patient’s image is then cropped into small segments of 100 by 100 pixels, each of which is analyzed using “deep convolutional neural networks” – a well-known model for automated image classification. A probability for potential syndromes is assigned to each segment and the data from the entire image is then compiled to return a prediction. In one test on 600 images of patients with Cornelia de Lange syndrome, the algorithm was 97 percent accurate as compared with a cohort of 65 experts who achieved only 75 percent accuracy.

About 6 percent of children worldwide are born with severe genetic syndromes, but many of these syndromes are rare. In addition, with hundreds of possible diagnoses, proper detection often depends on whether a doctor has encountered a specific condition before. For that reason alone, this is truly a field where artificial intelligence (AI) could move practice forward in ways that would otherwise be impossible.