Abstract

Introduction

Genetic conditions affect between 5 and 8% of the population [1, 2], and approximately 30–40% show recognizable facial features that provide clues for the diagnosis [3]. Memorizing syndromic facial characteristics is a challenging task for clinicians because of the high number of different syndromes and the fact that various syndromes share similar facial features. In recent years, advances in artificial intelligence have allowed to develop tools like Face2Gene (FDNA, Inc., Boston, MA, USA) as an aid in recognizing the most common syndromes based on facial gestalt. Face2Gene is based on DeepGestalt technology, where gestalt refers to the information contained in the facial morphology. DeepGestalt technology, based on computer vision and deep learning algorithms, quantifies similarities of hundreds of known syndromes and provides a list of suggested syndromes [4].

Silver–Russell syndrome (SRS) and Prader–Willi syndrome (PWS) are two examples of disorders for which early recognition is of utmost importance for proper treatment and support.

SRS is an imprinting disorder for which an underlying genetic cause is found in approximately 60% of patients who are clinically diagnosed based on the Netchine–Harbison scoring system [5]. The two main genetic alterations are loss of methylation (LOM) of chromosome 11p15 (30–60% of patients) and maternal uniparental disomy of chromosome 7 (UPD(7)mat; 5–10% of patients). Patients with SRS can present with being born small for gestational age, postnatal growth failure, feeding difficulties in the first years of life, low body mass index (BMI), body asymmetry, and typical facial features: a broad forehead tapering into a small, pointed chin giving the triangular face characteristic; a slightly beaked nose with a prominent nasal bridge; a well-demarcated philtrum; a wide mouth with downturned edges; lips with a thin vermilion border, especially the upper lip; micrognathia; and posteriorly rotated and somewhat low-set ears. These facial features become less obvious with age, making the diagnosis more difficult in older patients [6, 7].

PWS is caused by the lack of expression of genes on the paternally inherited chromosome 15q11.2-q13 region. There are three main genetic subtypes in PWS: paternal 15q11-q13 deletion (50–55% of patients), maternal uniparental disomy of chromosome 15 (UPD(15)mat; 45–50% of patients), and imprinting defects of the aforementioned region (1–3% of patients) [8]. Newborns with PWS present with hypotonia, poor sucking, and subsequent failure to thrive. In early childhood, children with PWS have short stature, food seeking behavior with excessive weight gain and early obesity when food intake is not restricted, developmental delay, cognitive disability, and behavioral problems. Patients with PWS have typical facial features which include a thin upper lip, a downturned mouth, a narrow nasal bridge, almond-shaped palpebral fissures, narrow bitemporal diameter, and strabismus. Although such features can be present at birth, they may not be obvious and progress slowly over time [9–12].

The aim of our study was to assess the clinical utility of Face2Gene technology in identifying patients with clinically or genetically proven SRS or PWS.

Materials and methods

Results

An example of the Face2Gene app feedback for a patient with SRS and a patient with PWS is shown in Fig. 1.

Open in a separate window

Fig. 1

Face2Gene app feedback on an SRS and a PWS patient. On the left: SRS (above) and PWS (below) suggested with high probability in one patient for each cohort. Both patients were molecularly diagnosed. The composite photo is shown. It results from the photos of affected individuals uploaded in the database, and it is compared by the app with the case picture to score the probability that the patient is affected by SRS (above) and PWS (below) on the basis of facial features. On the right: the heat map marks the areas of resemblance between the photo of the patient (SRS picture above, PWS picture below) and the composite photo. The intensity of the color scale grades from light blue to red; the red areas are more suggestive for the analyzed syndrome

Discussion

In this study, we evaluated the usability of the Face2Gene app in identifying patients with SRS and PWS. The correct diagnosis was suggested among the top 10 in 91% of SRS patients and in 100% of PWS. Therefore, the sensitivity was in line with what was reported by Gurovich et al. [4] and comparable to studies performed in children with various syndromes such as Cornelia De Lange syndrome [14] and Angelman syndrome [15].

Face2Gene has been tested in patients affected by very rare syndromes [16, 17], with variable overall sensitivity. In the study of Mishima et al., the top 1 sensitivity ranged from 42.9 to 61.2% and the top 10 sensitivity from 60 to 85.7%. Patients affected by 48 different dysmorphic syndromes were tested; the higher sensitivities were found after excluding patients affected by syndromes not included in the Face2Gene database and for which the app was not trained. Similar results were found by Marwaha et al. (top 10 sensitivity 57%, increasing to 82% when patients with syndromes not included by Face2Gene were excluded). The app has been validated in non-Caucasian patients, showing a good ability in suggesting the diagnosis despite facial variations that occur in different ethnic groups [18–20].

For the 2 SRS patients in whom the app did not suggest SRS as a top 10 diagnosis, SRS was clinically diagnosed according to the Netchine–Harbison clinical scoring system but not confirmed after extensive genetic testing (multilocus methylation and CNV analysis and next-generation sequencing of 18 short stature-related genes including HMGA2). This SRS scoring system has a low specificity and thus could result in false-positive results, especially since in both patients, the criterion “relative macrocephaly at birth” was absent. On the other hand, the spectrum of SRS and SRS-like diagnoses is extending, and rare genetic causes in the 11p15.5 region such as CDKN1C mutations, loss-of-function of IGF2, and large copy number variants in this region have been described [21]. Disruptions in the HMGA2-PLAG1-IGF2 pathway can cause an SRS phenotype [22]. In patients with a mutation in the HMGA2 gene, microcephaly at birth has been described [23]. It might be possible that, in the coming years, the Face2Gene app can be trained to recognize facial features of patients with non-classical SRS, if it is used more extensively.

In the SRS cohort, the app performed better when a photo taken at a younger age was available. SRS was suggested as top 1 and with higher probability in almost all patients aged 5 years or younger. This can be explained by the fact that facial features suggestive of SRS are more evident in the first years of life [7].

Few studies investigated the influence of GH treatment on craniofacial dimensions in children [24]. It is suggested that facial convexity (balance between forehead and upper and lower jaw) decreases, and mandibular length and posterior facial height increase during GH treatment in patients with GH deficiency, idiopathic short stature, being born small for gestational age with persistent short stature, and various genetic disorders. Although this may influence the performance of Face2Gene, it was not feasible to determine if it might have been of influence in our study since this was a cross-sectional study.

In the PWS group, the app performance was good. PWS is usually diagnosed in the first weeks after birth because of severe hypotonia and feeding problems. However, in some cases, the neonatal phenotype can be milder or unrecognized, the choice of the molecular test inappropriate [25], or genetic tests not available. In these cases, the diagnosis could be delayed, and tools such as Face2Gene might be supportive. Although the typical facial features in PWS become more evident in older patients, in our cohort, the performance was homogeneously satisfactory throughout the tested age range (1–15 years). Further studies performed in infants could be useful to investigate the usability of Face2Gene in children with PWS in the first months of life.

Based on our study and other studies mentioned above [14–17], Face2Gene can be helpful to identify potential syndromes which can be challenging for clinicians. However, there are some potential caveats to consider. First, the app always shows a list of 30 syndromes, also in healthy subjects without any suspicion of a syndrome or disease. Secondary, the app can suggest syndromes that do not match the phenotype except for partial overlap of some facial features. Thus, the level of probability shown by the bar plot (low, medium, or high) is the most relevant result to consider, with a high level of gestalt similarity being more relevant than the top 1 suggestion with low probability. In the study of Pantel et al., the performance of Face2Gene was studied in 323 patients affected by genetic syndromes and 323 healthy controls matched for age, sex, and ethnicity. The authors reported significantly lower scores (levels of probability) for detecting a genetic syndrome in the control cohort [26].

If a syndrome is suggested with medium–high probability, the clinician must critically evaluate if the suggestion matches the phenotype of the patient. Consultation with medical specialists such as clinical geneticists is of utmost importance and should be considered early in the course of the diagnostic work-up.

Lastly, some syndromes are often suggested in the list, creating a so-called background noise. For example, Klinefelter syndrome was proposed among the top 5 suggestion in 48% of PWS patients, of whom 57% were female.

Face2Gene presents several favorable aspects that make its use convenient. It is free of charge, it can be easily downloaded on the smartphone of verified healthcare professionals, and the access is protected by a password. Photos can be taken quickly through the app, with the possibility of using the face capture (the camera automatically takes the photo when the app frames the face contour) or can be uploaded from the gallery. Original photos are encrypted and stored in a database available only to the clinician or researcher who submitted the case. Nevertheless, ethical and privacy concerns related to the use of facial recognition technology and big data have been raised [27]. A thorough and careful conversation with patients and/or their caregivers to obtain consent is, therefore, essential prior to the use of this tool.

Our study has some limitations. The relatively small number of patients in the subgroups (for example, patients with specific genetic findings) might have resulted in not being able to find statistically significant results. Due to the rarity of these syndromes, international multi-center studies would be necessary to obtain larger samples. A quality check of the photos, in order to confirm that the photo was taken from the right angle to properly analyze the face or if it was taken with appropriate light and contrast, was absent. We did not obtain photos at different ages for each patient in order to investigate more extensively the performance of the app at different ages. As the app was tested in patients that already received the diagnosis of SRS or PWS, it was not possible to calculate the positive predictive value of the tool. However, our data support the use of Face2Gene as a tool for the identification of patients affected by SRS or PWS. A delay in diagnosis could compromise the care for these patients and unnecessarily delay treatment with, for example, growth hormone.

In conclusion, our study demonstrates that the Face2Gene app can be a useful tool to integrate in the diagnostic path of patients with SRS and PWS, as support for the clinicians. Deep learning algorithms such as Face2Gene cannot substitute a clinician or a geneticist. Nevertheless, including its use in the diagnostic path of patients suspected to have a genetic syndrome can shorten the time needed to make the right diagnosis, may reduce healthcare costs by providing clues for more targeted genetic tests, and can be supportive in establishing clinical diagnoses in settings where genetic tests are not easily available [28]. Moreover, the algorithms allow improvement in the tool’s accuracy in the future, and the identification of (new) syndromes in various ethnic groups can be quickly shared among colleagues at an international level, to enhance a global and more equal access to knowledge.

Abbreviations

Authors’ contributions

SC and DCMVDK drafted the study protocol and wrote the manuscript. SC, DCMVDK, WJG, LNG, and GFK collected the data. WJG, LNG, GFK, and ACSHK critically revised the manuscript. All authors approved the final version.

Data availability

The data generated during this study are available from the corresponding author on reasonable request.

Declarations

Footnotes

References