Up to 40% of rare diseases have facial changes that allow researchers to identify some medical conditions and help establish early diagnoses. Historically, the visual assessment and use of several classic anthropometric measurements, such as head diameter, have enabled early diagnosis of rare diseases. The most sophisticated and automated techniques based on artificial intelligence (AI) have made it possible to apply more objective methods to diagnosis. However, most AI-generated algorithms have databases that include populations of European origin, ignoring the genetic and morphological diversity of human populations worldwide.
Including populations of Native American, African, Asian, and European origins in AI-generated algorithms could be a way to diagnose rare diseases, as described in a paper published in the journal Scientific Reports. is decisive in improving The study is led by Nois Martínez Abadías, Lecturer in the Department of Biology, University of Barcelona, with participation of experts from the Universidad Ramón Llull, the ICESI University of Colombia and the Center for the Research of Birth Defects and Rare Diseases. (CIACER) and the Valle del Liri Foundation of Colombia.
Rare disease, mixed race, genetic ancestry
Automated diagnostics based on artificial intelligence can reveal patterns of severe or mild morphological abnormalities that are characteristic of each syndrome, “however, the major differences that can be detected when performing quantitative analysis of facial morphology are Yes,” emphasizes biological expert Nois Martinez-Abadias, an anthropologist and member of UB’s Department of Evolutionary Biology, Ecology and Environmental Sciences.
To address this question, the research team tested four genetic disorders in a Latino population: Down (DS), Morquio (MS), Noonan (NS), and neurofibromatosis type 1 (NF1). Facial phenotypes associated with the syndrome were assessed. Heterogeneity and great variation in genetic ancestry.
To quantitatively assess the facial features associated with each syndrome, the researchers recorded the 2D Cartesian coordinates of 18 facial landmarks in a sample of 51 individuals diagnosed with these syndromes and 79 controls. Did. Facial differences were studied using Euclidean distance matrix analysis (EDMA), based on statistical comparison of salient anatomical distances.
“Furthermore, we tested the diagnostic accuracy of an AI algorithm known as Face2Gene, which is used in clinical practice to identify these disorders through the analysis of facial morphometric features. In the case of Morquio syndrome, we were able to compare diagnostic results between Colombian and European samples,” added Martinez-Abadias.
Algorithms not representative of all human populations
The results showed that people diagnosed with DS and MS had the most severe facial dysmorphia, with 58.2% and 65.4% of people diagnosed with these diseases having significantly different facial features compared to controls. was The phenotype was mild in NS (47.7%) and less pronounced in NF1 (11.4%). The diagnostic accuracy of the deep learning automated algorithms used in this study was very high for DS but very low (less than 10%) for MS and NF1.
“Each syndrome exhibits a characteristic facial pattern, which supports the potential capacity of facial biomarkers as diagnostic tools. However, for each syndrome, we detected features unique to Colombians,” said Luis Miguel, Ph.D. student in biomedicine at UB and lead author of the paper. Echevverry says.
In this study, compared with the European sample, the average facial similarity between people diagnosed with DS and the automated algorithmic model was 100%, even though the diagnostic accuracy for Down syndrome was 100% in both populations. Variation was found to be significantly greater in the Colombian sample. For Noonan syndrome, the accuracy was significantly lower, from 66.7% in Colombian samples to 100% in European samples. Moreover, it was observed in all syndromes, with mixed race individuals being exactly those with the lowest facial similarity.
For Noonan syndrome, the accuracy was significantly lower, from 66.7% in Colombian samples to 100% in European samples. Moreover, in all syndromes, mixed-race individuals were observed to be precisely those with the lowest facial similarity.
Therefore, although AI-based automated diagnostic algorithms are optimized for European populations, they do not perform with the same accuracy in mixed populations of different genetic origins. “Developing unbiased predictive models is critical to supporting physician decision-making and providing accessible, universal, and effective technology for all humanity,” the research team points out. .
“By better understanding the facial malformations and population variability specific to each syndrome, we can improve diagnosis rates, reduce the pain of individuals and families in finding a diagnosis, and design early treatments for people. It is affected by a rare minority of medical conditions. This is especially true in resource-poor countries where it is more difficult to implement other diagnostic tests based on much more expensive genetic and molecular technologies,” the experts conclude.
