Clin Res Cardiol (2023).

Predicting Cardiovascular Risk Factors from Facial and Full Body Photography using Deep Learning
M. Knorr1, M. Neyazi1, J. P. Bremer1, J. Brederecke1, F. M. Ojeda Echevarria1, F. Ohm2, M. Augustin2, S. Blankenberg1, N. Kirsten2, R. B. Schnabel1, für die Studiengruppe: HCHS
1Klinik und Poliklinik für Kardiologie, Universitäres Herz- und Gefäßzentrum Hamburg, Hamburg; 2Institut für Versorgungsforschung in der Dermatologie und bei Pflegeberufen (IVDP), Universitätsklinikum Hamburg-Eppendorf, Hamburg;

Introduction: The early and simple detection of pathological cardiovascular phenotypes can lead to an early medical intervention and thus slow or limit the development of cardiovascular diseases. As full body photographs are easily obtainable without the need of medical expertise, their use holds the potential to be viable for screening of populations.

Purpose: Utilizing data from the Hamburg City Health Study (HCHS), we examined the possibility to detect cardiovascular risk factors from total body photographs using deep learning.

Methods: The HCHS is a population-based cohort study. The first data release includes facial and full body photographs in dermatologic standard poses of 5744 participants (median age 63.0 years, 50.0% men) as well as information on cardiovascular risk factors. Here, we focus on the most prevalent ones: former or current smoking status (prevalence: 23.5%), hypertension (63.3%) and diabetes (8.1%). We apply a 2D-Convolutional Neural Network for predicting the risk factors. It receives the facial image and up to two full body images as input. We compare it to a logistic regression model based on variables of the SCORE2 cardiovascular disease risk prediction algorithm (sex, age, smoking status, systolic blood pressure, diabetes, cholesterol levels and HDL-cholesterol levels), except for the respective target variable. Discrimination of the logistic regression and Neural Network models were derived using a 5-fold validation scheme and differences between the methods were tested for statistical significance using t-tests.

Results: The Neural Network provided a good discrimination for detection of arterial hypertension (AUC 0.76, CI 0.75 – 0.78), that was statistically significantly higher than the respective logistic regression’s ability to discriminate (AUC 0.73, CI 0.710 – 0.740, p < 0.001). The Neural Network was also able to detect a person’s smoking status (AUC 0.73, CI 0.72 – 0.75) with significantly higher discrimination than a logistic regression model (AUC 0.62, CI 0.59 – 0.65, p < 0.001). Lastly, the Neural Network showed a high discriminatory ability regarding the detection of  diabetes (AUC 0.81, CI 0.79 – 0.83). While the logistic regression also showed high levels of discrimination regarding diabetes (AUC 0.75, CI 0.71 – 0.779) the Neural Network still performed significantly better (p < 0.01).

Conclusion: The results suggest that the presence of cardiovascular risk factors like arterial hypertension, smoking status, and diabetes can be detected from total body photographs. As total body photographs can be easily obtained with a majority of digital cameras, including smartphones, the derived Neural Network model represents a potentially widely applicable diagnostic tool to easily screen large parts of the population for relevant cardiovascular risk factors, allowing an early medical intervention.