# An Artificial Intelligence Approach to the Craniofacial Recapitulation of Crisponi/Cold-Induced Sweating Syndrome 1 (CISS1/CISS) from Newborns to Adolescent Patients

## TL;DR

This study uses artificial intelligence to analyze facial features of patients with a rare genetic disorder to improve its diagnosis.

## Contribution

The first AI-powered image analysis of craniofacial features in Crisponi/cold-induced sweating syndrome 1 (CISS1/CISS).

## Key findings

- AI tools like DeepGestalt and GestaltMatcher successfully identified facial features of CISS1/CISS in severe cases.
- Facial D-Score confirmed consistent dysmorphic signs across patient cohorts.
- The study demonstrated the feasibility of using AI for clinical recognition of CISS1/CISS.

## Abstract

Background/Objectives: Crisponi/cold-induced sweating syndrome 1 (CISS1/CISS, MIM#272430) is a genetic disorder due to biallelic variants in CRFL1 (MIM*604237). The related phenotype is mainly characterized by abnormal thermoregulation and sweating, facial muscle contractions in response to tactile and crying-inducing stimuli at an early age, skeletal anomalies (camptodactyly of the hands, scoliosis), and craniofacial dysmorphisms, comprising full cheeks, micrognathia, high and narrow palate, low-set ears, and a depressed nasal bridge. The condition is associated with high lethality during the neonatal period and can benefit from timely symptomatic therapy. Methods: We collected frontal images of all patients with CISS1/CISS published to date, which were analyzed with Face2Gene (F2G), a machine-learning technology for the facial diagnosis of syndromic phenotypes. In total, 75 portraits were subdivided into three cohorts, based on age (Cohort 1 and 2) and the presence of the typical facial trismus (Cohort 3). These portraits were uploaded to F2G to test their suitability for facial analysis and to verify the capacity of the AI tool to correctly recognize the syndrome based on the facial features only. The photos which passed this phase (62 images) were fed to three different AI algorithms—DeepGestalt, Facial D-Score, and GestaltMatcher. Results: The DeepGestalt algorithm results, including the correct diagnosis using a frontal portrait, suggested a similar facial phenotype in the first two cohorts. Cohort 3 seemed to be highly differentiable. The results were expressed in terms of the area under the curve (AUC) of the receiver operating characteristic (ROC) curve and p Value. The Facial D-Score values indicated the presence of a consistent degree of dysmorphic signs in the three cohorts, which was also confirmed by the GestaltMatcher algorithm. Interestingly, the latter allowed us to identify overlapping genetic disorders. Conclusions: This is the first AI-powered image analysis in defining the craniofacial contour of CISS1/CISS and in determining the feasibility of training the tool used in its clinical recognition. The obtained results showed that the use of F2G can reveal valid support in the diagnostic process of CISS1/CISS, especially in more severe phenotypes, manifesting with facial contractions and potentially lethal consequences.