A machine learning-based screening tool for genetic syndromes in children

This is the authors' abstract. We don't add key points for this paper.

In The Lancet Digital Health, Antonio R Porras and colleagues1Porras AR Rosenbaum K Tor-Diez C Summar M Linguraru MG Development and evaluation of a machine learning-based point-of-care screening tool for genetic syndromes in children: a multinational retrospective study.Lancet Digit Health. 2021; 3: e635-e643Summary Full Text Full Text PDF PubMed Scopus (25) Google Scholar investigated how deep convolutional neural networks can be used to accurately classify the facial gestalt of individuals with genetic syndromes as dysmorphic. Their study is an important addition to the growing body of evidence that artificial intelligence (AI)-driven facial phenotyping could accelerate diagnostics of rare syndromes.2Ferry Q Steinberg J Webber C et al.Diagnostically relevant facial gestalt information from ordinary photos.eLife. 2014; 3e02020Crossref PubMed Google Scholar, 3Liehr T Acquarola N Pyle K et al.Next generation phenotyping in Emanuel and Pallister Killian syndrome using computer-aided facial dysmorphology analysis of 2D photos.Clin Genet. 2018; 93: 378-381Crossref PubMed Scopus (39) Google Scholar, 4Gurovich Y Hanani Y Bar O et al.Identifying facial phenotypes of genetic disorders using deep learning.Nat Med. 2019; 25: 60-64Crossref PubMed Scopus (347) Google Scholar, 5Pantel JT Hajjir N Danyel M et al.Efficiency of computer-aided facial phenotyping (DeepGestalt) in individuals with and without a genetic syndrome: diagnostic accuracy study.J Med Internet Res. 2020; 22e19263Crossref PubMed Scopus (16) Google Scholar However, the study has several limitations that should be addressed. First, the data provided with the study are insufficient to reproduce results properly. We appreciate that original images cannot be released for privacy reasons. However, it remains unclear why the anonymised statistical shape models or descriptive information about each included individual could not be provided. Second, a smartphone application (app) does not appear to be readily accessible. The authors, however, report to have developed a smartphone app based on their neural network. Third, it is unclear how the threshold to calculate accuracy, sensitivity, and specificity was determined. Neither a list of scores assigned to each patient or control image nor a receiver operating characteristic curve are reported. The code provided to evaluate their model holds functions to produce these metrics though. Fourth, Porras and colleagues did not compare their findings with previous similar studies, which would have facilitated benchmarking of their results. In particular, the authors should have compared their tool's discriminatory power to that of DeepGestalt—a computer-aided facial phenotyping tool that has shown promising results for detecting potentially syndromic faces.3Liehr T Acquarola N Pyle K et al.Next generation phenotyping in Emanuel and Pallister Killian syndrome using computer-aided facial dysmorphology analysis of 2D photos.Clin Genet. 2018; 93: 378-381Crossref PubMed Scopus (39) Google Scholar, 5Pantel JT Hajjir N Danyel M et al.Efficiency of computer-aided facial phenotyping (DeepGestalt) in individuals with and without a genetic syndrome: diagnostic accuracy study.J Med Internet Res. 2020; 22e19263Crossref PubMed Scopus (16) Google Scholar Fifth, it is unclear how syndromes were selected or excluded. Lastly, although the authors discuss ethnicity, sex, and age as possible biases and analysed their data with respect to these parameters, they did not test for the specific syndrome of an individual as a confounder. This omission is especially unfortunate as five (4%) of the 129 included syndromes are highly overrepresented, accounting for 1076 (77%) of the 1400 images in the syndromic cohort. The median image count of the remaining 124 syndromes is 1 (mode 1, mean 2·61 [SD 3·45]). Notably, the degree of facial dysmorphism varies between different syndromes. The five syndromes overrepresented in the dataset used by Porras and colleagues1Porras AR Rosenbaum K Tor-Diez C Summar M Linguraru MG Development and evaluation of a machine learning-based point-of-care screening tool for genetic syndromes in children: a multinational retrospective study.Lancet Digit Health. 2021; 3: e635-e643Summary Full Text Full Text PDF PubMed Scopus (25) Google Scholar display a characteristic facial gestalt and are, therefore, rather distinctive (appendix). They should be easily identifiable as dysmorphic even for medical professionals not trained in syndromology. Thus, an estimate of their tool's accuracy beyond these syndromes is not possible, and its added diagnostic value (compared with such clinicians) remains unclear. A table specifying age, sex, ethnicity, diagnosis, and the yielded scores of each included individual would mitigate these concerns and further improve the promising work of Porras and colleagues. We declare no competing interests. Download .pdf (.3 MB) Help with pdf files Supplementary appendix Development and evaluation of a machine learning-based point-of-care screening tool for genetic syndromes in children: a multinational retrospective studyThis genetic screening technology could support early risk stratification at the point of care in global populations, which has the potential accelerate diagnosis and reduce mortality and morbidity through preventive care. Full-Text PDF Open AccessA machine learning-based screening tool for genetic syndromes in children – Authors' replyWe appreciate the letter submitted by Martin Atta Mensah and colleagues1 in response to our Article.2 We are happy that the scientific community is involved in improving the management of genetic conditions and would like to provide additional clarity.