Science and Tech

The big mistake of many artificial intelligence algorithms to diagnose rare diseases

[Img #68455]

Up to 40% of rare diseases present in those who suffer from them facial alterations that make it possible to differentiate some pathologies from each other and that even help to establish a first diagnosis. Traditionally, visual assessment and some classic anthropometric measurements (such as head diameter, among others) have facilitated a first clinical diagnosis of rare disorders. Now, with the most sophisticated and automatic techniques (based on artificial intelligence) it is possible to apply more objective methods in diagnosis.

However, most of the algorithms generated by artificial intelligence start from databases of populations of European origin and ignore the genetic and morphological diversity of human populations around the world.

This has been verified in a recent study carried out by the team led by Luis M. Echeverry-Quiceno, from the University of Barcelona (UB).

Incorporating populations with Amerindian, African, Asian and European ancestry into algorithms generated by artificial intelligence is vital to improving the methods used to diagnose rare disorders, according to the results of the aforementioned study.

Members of the Ramon Llull University, the Icesi University in Colombia and its Research Center on Congenital Anomalies and Diseases have participated in the research, led by Professor Neus Martínez-Abadías, from the Faculty of Biology of the University of Barcelona. Raras (CIACER) and the Valle del Lili Foundation of Colombia.

Incorporating populations with Amerindian, African, Asian and European ancestry into automated algorithms is essential to improve the diagnosis of rare diseases. (Image: UB)

Rare diseases, interbreeding and genetic ancestry

Automatic diagnosis based on artificial intelligence can reveal patterns of severe or mild dysmorphology that are characteristic of each syndrome “but with significant differences that can be detected when a quantitative analysis of facial morphology is performed,” explains Professor Neus Martínez-Abadías, an expert in biological anthropology and member of the Department of Evolutionary Biology, Ecology and Environmental Sciences of the UB.

The team evaluated facial phenotypes associated with four genetic syndromes (Down syndrome, Morquio syndrome, Noonan syndrome, and neurofibromatosis type 1) in a Latin American population with individuals exhibiting a wide variety of genetic ancestry and miscegenation.

To quantitatively assess the facial features associated with each syndrome, the Cartesian coordinates of 18 facial landmarks on 2D frontal images were recorded in a sample of 51 people diagnosed with these syndromes and 79 people without such diseases (control group). Facial differences were studied using the geometric morphometry analysis methodology (Euclidian Distance Matrix Analysis or EDMA), based on the statistical comparison of prominent anatomical distances.

“In addition, we tested the diagnostic accuracy of an artificial intelligence algorithm —known as Face2Gene— that is used in clinical practice to identify these types of disorders by analyzing the morphometric characteristics of the face. In the case of Down syndrome and Morquio syndrome, we were able to compare the diagnostic results between the Colombian sample and a European one”, adds Martínez-Abadías.

Algorithms that do not represent all human populations

According to the results, people diagnosed with Down syndrome and Morquio syndrome presented the most severe facial dysmorphologies, with 58.2% and 65.4% of facial features significantly different in people compared to the population of the group of control. The phenotype was milder in Noonan syndrome (47.7%) and not significant in neurofibromatosis type 1 (11.4%). The diagnostic accuracy of the deep learning machine algorithm used in the study was very high in the case of Down syndrome, moderate in Noonan syndrome, and very low (less than 10%) in Morquio syndrome and neurofibromatosis type 1.

“Each syndrome presented a characteristic facial pattern, supporting the potential ability of facial biomarkers as diagnostic tools. In general, the observed traits were consistent with traits described in the medical literature based on European populations. However, we detected specific features of the Colombian population for each syndrome”, points out Luis Miguel Echeverry, a PhD student in Biomedicine at the UB and co-author of the study.

Compared with a European sample, the study reveals that although the diagnostic accuracy for Down syndrome was 100% in both populations, the variation in mean facial similarity between people diagnosed with Down syndrome and the automatic algorithm model was significantly higher. higher in the Colombian sample. In the case of Noonan syndrome, the precision was significantly lower, from 66.7% in the Colombian sample to 100% in the European one. In addition, it was observed that, for all the syndromes, mixed-race individuals were precisely those with the lowest facial similarities.

Thus, automatic diagnostic algorithms based on artificial intelligence are optimized in European populations, but do not work with the same precision in mixed-race populations of different genetic origin. “Developing unbiased predictive models is crucial to support physicians in their decision-making and provide accessible, universal, and effective technology for all human populations,” the study authors emphasize.

“With a greater understanding of the specific facial dysmorphologies of each syndrome and the diversity of the population, it is possible to improve diagnosis rates, try to reduce the personal and family odyssey to find a diagnosis and thus be able to design earlier treatments for people affected by rare minority pathologies. This is especially relevant in countries with scarce resources and more difficulties in performing other diagnostic tests based on genetic and molecular techniques that are much more expensive,” the experts conclude.

​​​​​​​

The study is titled “Population-specific facial traits and diagnosis accuracy of genetic and rare diseases in an admitted Colombian population.” And it has been published in the academic journal Scientific Reports, from the Nature publishing group. (Source: UB)

Source link