A facial foundation model for multi-system biomarker and disease risk prediction with real-world mobile deployment

A facial foundation model for multi-system biomarker and disease risk prediction with real-world mobile deployment | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article A facial foundation model for multi-system biomarker and disease risk prediction with real-world mobile deployment Minxian Wang, Tingfeng Xu, Huixuan Xu, Li Lin, Fei Wang, Yaodong Ding, and 10 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8110055/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract While most biomarkers currently rely on invasive laboratory testing, which limits large-scale or repeated screening, scalable non-invasive methods could transform population screening, early disease detection and personalized health management. Facial photographs, as a ubiquitous and non-invasive data source, offer such potential but remain underexplored for clinically relevant biomarker and disease risk prediction. Here, we present FaceFound, a facial foundation model trained over 10 million images through a progressive general-to-clinical pretraining strategy and evaluated across 62 biomarkers spanning eight physiological systems, many of which are well-established indicators of cardiometabolic, renal, and systemic diseases.. FaceFound consistently outperformed baseline architectures in biomarker prediction, achieving state-of-the-art performance for 45 (73%) biomarkers and demonstrating robust results across internal and four independent external cohorts (N = 206-2,247). Notably, FaceFound displayed superior performance to genetic models for the prediction of 14 out of 26 biomarkers with genetic scores available in the PGS catalog, highlighting its complementary value for disease risk assessment beyond inherited genetic susceptibility. Moreover, Face-predicted cardiovascular biomarkers demonstrated strong associations with coronary stenosis, enabling accurate prediction of cardiovascular disease risk and outperforming models based on laboratory-measured biomarkers. FaceFound further exhibited label efficiency, retaining predictive power with as few as 400 training samples, underscoring its value in low-resource settings. Moreover, FaceFound was deployed as a smartphone application, enabling real-time biomarker estimation and individualized disease risk reporting from a single self-captured facial photograph. These findings provide that FaceFound can reproducibly predict multi-system biomarkers and clinically relevant disease risk from facial images with real-world feasibility, establishing a paradigm for population-wide digital screening, early disease risk stratification and personalized risk assessment. Health sciences/Health care/Public health/Population screening Health sciences/Biomarkers/Predictive markers Health sciences/Risk factors Biological sciences/Computational biology and bioinformatics/Machine learning Biological sciences/Computational biology and bioinformatics/Predictive medicine Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction Quantifying internal physiological states is fundamental to disease prevention, early diagnosis, and personalized health management 1 . Clinical biomarkers are essential indicators that span multiple organ systems and are widely utilized for diagnosis 2 , risk stratification 3 , and informed therapeutic decision-making 4 . Yet, most biomarkers require invasive sampling and laboratory-based testing, which limits accessibility for large-scale screening and frequent monitoring, particularly in resource-constrained settings. Approaches that enable non-invasive, widely deployable biomarker assessment and disease risk evaluation could therefore transform population health screening and personalized care 5 – 8 . Facial photographs are ubiquitous, low-cost, and non-invasive images. Facial features have been recognized as reflecting health and disease status; for example, renal dysfunction may manifest as facial edema or pallor 9 , and decompensated cirrhosis may lead to yellowish facial discoloration due to elevated serum bilirubin levels 10 . Advances in computer vision now enable the extraction of complex phenotypic features from facial images at scale 11 – 14 . Vision foundation models, trained on large and diverse unlabeled image datasets, offer a robust framework for learning transferable representations 12 , 15 and have demonstrated success in medical image analysis for various diseases 16 – 19 . Nevertheless, most current studies have focused on narrow facial analysis tasks 20 – 23 and limited health outcomes 24 – 26 . The systematic utility of facial images for predicting biomarkers routinely tested in a healthcare system for assessing disease risks across multiple systems remains unexplored. To address this gap, we developed FaceFound, a facial foundation model that uses a progressive general-to-clinical pretraining strategy to reduce the domain gap and enhance transferability. FaceFound was pretrained in three stages—on ImageNet-22K for general visual representations, VGGFace2-HQ for framing facial structures, and clinical facial images (AZ-TR unfiltered) for capturing pathological features (Fig. 1 a). Fine-tuned on 62 biomarkers, FaceFound was compared with conventional architectures (ResNet18 and SwinLarge) using 4 independent cohorts in terms of prediction accuracy, label efficiency, stability with repeated images, and risk assessment for coronary stenosis (Fig. 1 b ) . The prediction performance was also compared to polygenic risk score (PRS) models in another external cohort with genetic data available (Fig. 1 c ) . To translate these insights into a practical application, we deployed FaceFound on smartphones to enable selfie-based data collection and established an new external cohort of self-captured photos to evaluate its performance (Fig. 1 d). Collectively, this study establishes facial foundation modeling FaceFound as a scalable, non-invasive framework for biomarker prediction with broad implications for precision medicine and digital health. Results FaceFound model development We developed a facial foundation model called FaceFound using self-supervised learning (SSL) with a progressive general-to-clinical strategy. We implemented the general-to-clinical pretraining pipeline utilizing ImageNet-22k (14,197,122 images) 27 , VGGFace2-HQ (1,160,250 images) 28 , 29 , and facial photographs from the Chinese Anzhen Training Cohort (AZ-TR-unfiltered, 88,489 images), Fig. 1 a. In the first stage, the Swin-Large Transformer backbone was pretrained on ImageNet-22k with 14.2 million images to capture general-purpose visual features. In the second stage, the model was further adapted using approximately 1.16 million high-resolution facial images from the VGGFace2-HQ dataset. In the final stage, domain-specific pretraining was conducted using 88,489 clinical facial photographs from the AZ-TR dataset without linking to hospital records (AZ-TR-unfiltered) (Fig. 1 a, Extended Data Fig. 1 , and Methods ). This progressive approach allowed the model to initially learn general visual representations in depth, adapt to facial morphological features, and specialize in disease-relevant representations. To fine-tune FaceFound for predicting 62 biomarkers routinely tested in a healthcare system across eight physiological systems ( Supplementary Table 1 ), 4,698 individuals with linked hospital records from AZ-TR-unfiltered dataset (median age 60.0, IQR, 54.0–67.0, AZ-TR), and each with over three frontal facial photographs and corresponding biomarker measurements were used. Fine-tuning for these biomarkers was performed with task-specific adapter layers (see Methods for fine-tuning details and Extended Data Fig. 1 ). Evaluation of FaceFound performance for biomarker prediction in internal and external cohorts We evaluated the performance of Facefound model for prediction of 62 biomarkers covering 8 physiological systems and compared its performance with the baseline SwinLarge and ResNet18 models, in one internal validation cohort(AZ-TR, N = 4,698) and four external validation cohorts (AZ-EV1, N = 1,498; AZ-EV2, N = 2,247; AZTZ-EV, N = 206; DX-EV, N = 900) (details in Methods and Supplementary Table 2 ). We first examined the predictability of biomarkers across systems based on FaceFound. For 8 physiological systems, we found Anthropometrics (internal cohort: R 2 = 0.493 (0.361–0.666), IQR = 0.361–0.666; external cohort: R 2 = 0.492 (0.410–0.633), IQR = 0.410–0.633), and Renal (internal cohort: R 2 = 0.081 (0.038–0.161), IQR = 0.038–0.161; external cohort: R 2 = 0.100 (0.037–0.184), IQR = 0.037–0.184) has the highest predictive accuracy across internal cohort and four external cohorts. While, for Electrolytes (internal cohort: R 2 = 0.020 (0.013–0.036), IQR = 0.013–0.036; external cohort: R 2 = 0.001 (0.000–0.005), IQR = 0.000–0.005) has the weakest predictive performance (Fig. 2 a and Supplementary Table 4 ). Among the 62 biomarkers, 16 achieved an R 2 > 0.10 in the internal test cohort (median R 2 = 0.282, IQR: 0.123–0.381). Among them, 12 biomarkers reached R 2 > 0.10 in at least one external cohort, distributing in Anthropometrics (n = 4; Age, height, weight, and BMI), Hematology (n = 3; HCT, Hb, and RBC), Cardiovascular (n = 2; HDL-C and Hcy), Renal (n = 2; UA and eGFR), and Endocrine (n = 1; HbA1C), respectively (Fig. 2 b). These findings suggest that systemic health status is reflected in facial characteristics. Among these, we found HDL-C (internal cohort: R² 0.214, 95% CI 0.147–0.280; external cohort: median R² 0.185, IQR 0.165–0.192), HCT (internal cohort: R² 0.378, 95% CI 0.303–0.453; external cohort: median R² 0.260, IQR 0.202–0.307), RBC (internal cohort: R² 0.285, 95% CI 0.210–0.360; external cohort: median R² 0.228, IQR 0.201–0.244), and eGFR(CKD-EPI) (internal cohort: R² 0.289, 95% CI 0.218–0.359; external cohort: median R² 0.196, IQR 0.184–0.260) emerged as the most robust biomarkers predicted by FaceFound, showing strong agreement and consistent meta-analytic R 2 estimates between predicted and measured values across both internal and external cohorts (Fig. 2 c). These results indicate that facial photographs encode clinically meaningful signals related to lipid metabolism, erythrocyte parameters, and kidney function. Next, we compared the predictive performance of FaceFound with the two widely used baselines SwinLarge and ResNet18. We found FaceFound outperformed SwinLarge and ResNet18 both in internal and external cohorts. In the internal cohort, FaceFound (median R 2 = 0.039, IQR: 0.011–0.085) outperformed SwinLarge (median R 2 = 0.028, IQR: 0.005–0.059) and ResNet (median R 2 = 0.012, IQR: 0.002–0.044) for prediction of 62 biomarkers. Further, in the four external cohorts, FaceFound performed consistent superiority (median R 2 = 0.030, IQR: 0.007–0.071) to SwinLarge (median R 2 = 0.023, IQR: 0.004–0.061) and ResNet (median R 2 = 0.011, IQR: 0.001–0.035) (Fig. 2 d). Totally FaceFound achieved state-of-the-art (SOTA) performance in 45 of the 62 biomarkers in the internal cohort (Fig. 2 e and Supplementary Table 5 ). Collectively, these findings demonstrate that facial images capture a substantial amount of information across a broad spectrum of systemic biomarkers, particularly those with implications for cardiovascular and renal health. For sensitivity analysis, we found that 26 of the 62 predicted biomarkers were independently associated with laboratory measurements in the external AZ-EV1 cohort after adjustment for age, sex, height, weight, and BMI ( Supplementary Table 6 ). Further, to calculate the contribution of facial components and demographic factors in describing variation in biomarker prediction, we developed the multivariable model using both FaceFound and demographic factors (Age, Sex, Height, Weight and BMI). We found that facial features explained a larger proportion of variation (median 65.8%, IQR: 51.3–75.5%), indicating that facial features contain information relevant to clinical biomarkers, independent of common demographic and anthropometric factors ( Extended Data Fig. 3 ). FaceFound and genetic model (PRS) in predicting biomarkers To compare the performance of FaceFound and polygenic risk model for prediction of biomarkers, we curated 679 polygenic risk scores (PRS) for 26 biomarkers from the PGS Catalog and computed PRS in the AZ-GEV cohort (N = 111, median age 59.0, IQR: 52.0-65.5) with genetic data available. Among the 26 biomarkers, FaceFound outperformed PRS for 14 biomarkers prediction, including BMI, HDL-C, hemoglobin, and eGFR (CKD-EPI), which played important roles in cardiometabolism (Fig. 3 a, 3 b, Extended Data Fig. 4 Supplementary Table 7 ). Moreover, to compare the prediction performance between genetics and face features, we modeled a multivariable model using both the best PRS and FaceFound score as predictors to predict each of the 26 biomarkers in the AZ-GEV cohort. We found that facial features can independently explain a larger proportion of variation compared to the PRS for 15 biomarkers, whereas the genetics explained more for 11 biomarkers. Together, these findings demonstrate that FaceFound provides predictive performance comparable to or exceeding polygenic risk scores, underscoring the potential of facial images as an accessible and powerful modality for biomarker estimation. Evaluation of the clinical utility of FaceFound predicted biomarkers Next, we assessed the clinical values of FaceFound predicted biomarkers compared to laboratory measurements by examining their associations with coronary stenosis severity (≥ 50% vs. <50%), a condition strongly linked to downstream cardiovascular diseases across AZ-TR and DX-EV with available coronary angiography (CAG) data (Fig. 4 a). Across 16 cardiovascular-related biomarkers, facial predictions were consistent with laboratory measurements ( Extended Data Fig. 5 and Supplementary Table 9 ), with particularly similar patterns for six key biomarkers (HDL-C, LDL-C, TG, TC, HbA1C, hsTnI) (Fig. 4 b). For example, HDL-C, a biomarker of established clinical relevance, facial predictions (OR/SD 0.705, 95% CI: 0.579–0.858, P < 0.001) showed associations consistent with those obtained from laboratory measurements (OR/SD 0.856, 95% CI: 0.749–0.978, P < 0.001), both pointing in the same protective direction. In contrast, for LDL-C, since most participants in the studied cohorts were cardiology inpatients receiving statin-based lipid-lowering therapy in AZ-TR and DX-EV cohort, leading to a negative association for laboratory-measured LDL-C (OR/SD 0.726, 95% CI: 0.636–0.829, P < 0.001), whereas face-predicted LDL-C preserved the expected positive association (OR/SD 1.445, 95% CI: 1.252–1.668, P < 0.001), indicating that FaceFound captures underlying disease-related features independent of drug effects. Next, to evaluate the risk evaluation potential of face-predicted biomarkers, we developed and compared three coronary stenosis risk prediction models using different combinations of FaceFound predicted and laboratory-measured biomarkers. The lab-measured basic model (Model 1) incorporated demographic and anthropometric variables (age, sex, height, weight) with 4 basic laboratory-measured lipid profiles (LDL-C, HDL-C, TC, TG). The face-predicted basic model (Model 2) replaced 4 basic laboratory-measured lipids with face-predicted counterparts. The face-predicted advanced model (Model 3) expanded the FaceFound predicted factors to 16 face-predicted cardiovascular biomarkers (Fig. 4 c). The model was trained in the AZ-TR, and we evaluated the performance of classification by area under curve (AUC) in the external DX-EV cohort ( Supplementary Table 10 ). In the external DX-EV cohort, face-predicted basic model (Model 2) and face-predicted advanced model (Model 3) displayed comparable and higher performance (AUC: 0.649, 95% CI: 0.609–0.690 vs. 0.657, 95% CI: 0.618–0.696) compared with lab-measured basic model (Model 1) (0.614, 95% CI: 0.570–0.656) (Fig. 4 d). We next assessed the risk stratification values of three models in the external independent DX-EV cohort. In the low-risk group (the bottom 20% predicted risk), stenosis prevalence was 42.9%, 38.6%, and 37.1% for lab-measured basic model, face-predicted basic model, and face-predicted advanced model respectively; for the high-risk group (the top 20% risk), the rates were 71.6%, 66.7%, and 73.0%. These results indicate that FaceFound predicted biomarkers can achieve comparable risk stratification values with laboratory measurements (Fig. 4 e). Finally, feature importance analysis for FaceFound predicted biomarkers compared to lab-measured biomarkers in AZ-TR showed that the contributions of face-predicted biomarkers were consistent with laboratory-measured counterparts, with an intraclass correlation coefficient (ICC) of 0.86 (95% CI, 0.68–0.94) (Fig. 4 f and Supplementary Table 11 ). Among them, HDL-C, HbA1C, and TC were the most influential predictors in both modalities, highlighting their robust association with stenosis risk. Together, these results demonstrate that FaceFound predicted biomarkers preserve disease-associated information, perform comparably to laboratory measurements for risk evaluation, and support the development of scalable, clinically meaningful risk assessment strategies. Stability and label efficiency of FaceFound A critical requirement for clinical deployment is model stability, ensuring that predictions remain consistent across multiple measurements. To evaluate this, we performed a test-retest analysis from internal and external cohorts, using facial photographs repeatedly acquired from the same participant with minor variations in capture angle. FaceFound displayed excellent stability, achieving higher intraclass correlation coefficients (ICCs) compared to SwinLarge and ResNet18 across all cohorts. In the internal AZ-TR cohort, FaceFound achieved an ICC of 0.879 (IQR: 0.838–0.931); in the external datasets, FaceFound achieved higher ICC compared with other models, with 0.914 (IQR: 0.883–0.947) in AZ-EV1, 0.913 (IQR: 0.881–0.947) in AZ-EV2, and 0.971 (IQR: 0.958–0.980) in DX-EV ( Fig. 5 a and Supplementary Table 12 ). Across physiological systems, FaceFound retained stable performance, with median ICCs 0.911 (IQR: 0.903–0.951) in internal and external cohorts with low variance (Fig. 5 b; Supplementary Table 13 ). For example, HDL-C showed high stability in AZ-TR (ICC of 0.905, 95% CI: 0.898–0.912), AZ-EV1 (ICC of 0.945, 95% CI: 0.944–0.946), AZ-EV2 (ICC of 0.946, 95% CI: 0.944–0.947), and DX-EV (ICC of 0.977, 95% CI: 0.976–0.978), as shown in the Bland-Altman plots (Fig. 5 c-f and Supplementary Table 12 ). These findings confirm that FaceFound yields stable, reproducible biomarker predictions, supporting its robustness estimates. Label efficiency refers to the amount of training data and labels required to achieve a target performance level for a given downstream task, which indicates the annotation workload for medical experts. We investigated the label efficiency of FaceFound by systematic evaluation across training samples from 400 to 3200 samples, using 62 systematic biomarkers. Across all sample sizes, FaceFound consistently demonstrated superior label efficiency compared with SwinLarge and ResNet18 in all internal and external cohorts (Fig. 5 g and Supplementary Table 14 ). Notably, with as few as 400 samples, FaceFound (median R² = 0.016, IQR: 0.046–0.005) achieved substantially higher predictive performance than SwinLarge (median R² = 0.007, IQR: 0.001–0.027) and ResNet18 (median R² = 0.003, IQR: 0.001–0.017). To quantify data efficiency, we further estimated the proportion of training data required by FaceFound to match the performance of SwinLarge or ResNet18. For eGFR (CKD-EPI) and LYM, FaceFound reached equivalent accuracy using only 10% of labelled data. Superior efficiency was also observed for HbA1c (20%), BMI (28%), Hb (50%) and HDL-C (60%) ( Extended Data Fig. 6 ). These results demonstrate that FaceFound not only achieves stable and reproducible predictions but also confers strong label efficiency, a property critical for developing scalable prediction in new domains or for novel biomarkers where annotated data are scarce. Model interpretation To gain insights into the inner-workings of FaceFound leading to its superior performance and label efficiency in biomarkers prediction, we performed qualitative analyses of the pretext task used for self-supervised pretraining and biomarker-specific and stage-specific decisions of FaceFound. To evaluate the SSL pretext task, we curated internet-sourced facial images covering both Asian and European populations, displaying diverse abnormalities, including peripheral facial paralysis, renal disease–associated edema, and Down syndrome, ensuring these facial images were not included in the pre-training stage. We found that FaceFound was able to reconstruct pathological facial features with 75% of the image masked, across different types of facial deformities or abnormal traits, such as periorbital swelling, malocclusion, or macular lesions ( Extended Data Fig. 7a ). This demonstrates that FaceFound has learned to identify and infer the representation of health-related areas by means of SSL, which contributes to performance and label efficiency in biomarker predictions. Beyond reconstruction-based interpretation, we applied Grad-CAM to the Swin-Large backbone of FaceFound to visualize regions contributing to biomarker-specific predictions ( Extended Data Fig. 7b ). For eGFR (CKD-EPI), early stages captured global facial context, intermediate stages emphasized the eyes and cheeks, and the final stage focused on the periocular region. For Hb, later stages integrated signals from the entire face in addition to the eyes. For HDL-C, salient contributions were largely concentrated around the periocular region. These findings indicate that FaceFound relies on stage-dependent and physiologically plausible facial features when predicting diverse systemic biomarkers. These visualizations demonstrate that FaceFound learns interpretable and biologically meaningful facial representations, bridging facial phenotypes with systemic health indicators and supporting its potential for scalable biomarker prediction in diverse populations. Real-world deployment of FaceFound-based biomarker prediction via smartphone To assess the translational potential of FaceFound, we developed a smartphone-based application integrating with FaceFound (Fig. 6 a). 62 systematic biomarkers can be predicted using standardized frontal facial images by smartphone under a white background and indoor light. With a mobile-based application, users can check their face-predicted biomarkers spanning different health systems with an interpretable report (Fig. 6 b). We piloted the application in the cardiology department of Beijing Anzhen Hospital, where frontal images were taken by a physician who was blinded to the study design, using the front-facing camera of mobile phone as the external validation cohort (median age 60.0, IQR, 55.0–67.0, 26.4% female, N = 110, APP-EV) ( Methods for details and Supplementary Table 15 ). For each individual, the medical records of 58 biomarkers were collected. Overall, after a systematic evaluation, the median R² of FaceFound app was consistent with counterparts evaluated in external cohorts (Fig. 6 c and Supplementary Table 16 ), with consistency assessed by ICCs in the external cohorts AZ-TR, AZ-EV1, AZ-EV2, AZTZ-EV, and DX-EV, with a median ICC of 0.910 (IQR, 0.900–0.940). (Fig. 6 d, Supplementary Table 17 ). These results demonstrate that biomarker prediction from simple selfies is feasible in real-world clinical environments. This deployment highlights the potential of FaceFound for rapid, low-cost, and scalable integration into pre-screening workflows and digital health monitoring platforms. Discussion In this study, we demonstrate that a facial foundation model FaceFound with progressive general-to-clinical pretraining strategy have good prediction power for clinical biomarker prediction across eight physiological systems, utilizing over 70,000 facial photographs from approximately 10,000 participants across five diverse Chinese cohorts with stability, label efficiency, higher performance for most biomarkers compared to genetics and comparable predictive performance with facial images taken with a smartphone. To our knowledge, this provides the first broad evidence that facial photographs can be leveraged for predicting clinical biomarkers. Previous studies of facial images have focused mainly on non-medical domains such as person identity recognition and emotion analysis 20 – 23 , with only limited work addressing health-related outcomes, including aging 30 , syndromic facial dysmorphisms 25 , and coronary artery disease 24 , 26 . Our findings substantially extend this landscape by showing that facial photographs carry strong predictive power for biomarkers across renal, hematologic, cardiovascular, endocrine, digestive, and immune systems. These results suggest that facial images can be applied to diverse disease risk assessment and health monitoring, and further developed as an easily accessible medical imaging modality to support early screening and risk evaluation, thereby opening new avenues for integrating facial analysis into precision medicine and population health. Although both genetic and environmental factors shape biomarker variation, previous studies have largely emphasized genetic contributions 31 while neglecting the substantial role of acquired exposures. Facial morphology, as a complex phenotype influenced by genetics and environment 32 , provides a unique lens to evaluate personal health status, yet its variance explained compared with genetics remains unclear. By analyzing 679 polygenic risk scores (PRS) from 27 phenotypes, with the optimal PRS representing genetic effects, we demonstrate that in 15 phenotypes, facial signals surpassed genetic prediction. Notably, facial features more effectively captured estimated glomerular filtration rate (eGFR (CKD-EPI)), underscoring the predominant facial features for kidney function, and for the first time we identify stronger facial associations with high-density lipoprotein cholesterol (HDLC) and lipoprotein(a) [Lp(a)] than by genetics, which helps explain the predictive power of facial features for coronary artery disease. Collectively, these findings establish the human face as an accessible and dynamic source of biomarker information, more sensitive to environmental influences than static genetic data, and highlight the potential of combining facial and genetic information together to improve biomarker prediction and to reveal novel interactions between genes, facial development, and environmental exposures. Our study further underscores the clinical value of face-predicted biomarkers. As central indicators of physiological processes, biomarkers fluctuate with disease progression and therapeutic interventions 33 . We found that facially predicted biomarkers preserved these dynamic variations between health and disease states with high consistency by evaluation of stenosis with 16 cardiovascular biomarkers. Interestingly, we observed participants receiving lipid-lowering therapy exhibited reduced LDL-C levels while facially predicted LDL-C values remained elevated among individuals with coronary stenosis in the external independent cohort, possibly due to the improvement of FaceFound in capturing reversion correlation after medical treatment. Furthermore, incorporation of facially predicted cardiovascular biomarkers improved risk evaluation accuracy for stenosis beyond traditional risk factors 3 . These findings suggest that face-predicted biomarkers could be integrated into established clinical risk models to enhance disease assessment and management. A key strength of our work lies in the high performance and robustness of the proposed FaceFound model. FaceFound achieved excellent test-retest reliability across the internal and external cohorts, substantially outperforming existing models. This stability is critical for clinical deployment, ensuring reproducibility across repeated captures and variable imaging conditions. Moreover, FaceFound exhibited superior label efficiency, maintaining high predictive accuracy even when trained on as few as 400 samples. Such efficiency underscores the potential of FaceFound to accelerate biomarker discovery and disease modeling in rare conditions, emerging infectious diseases, or other settings where labeled data are scarce. By leveraging large-scale pretraining, our framework confers both robustness and adaptability, two properties essential for real-world clinical translation. Beyond methodological advances, we demonstrate the feasibility of real-world deployment by integrating FaceFound into a smartphone-based application. Remarkably, predictions derived from mobile-captured images achieved performance comparable to research-grade inputs, highlighting the translational strength of our approach. This capability enables novel scenarios for health monitoring, including at-home biomarker screening, community-level digital health interventions, and rapid triage in emergency or resource-limited settings. Furthermore, because our framework relies solely on facial photographs, it can be seamlessly embedded into widely adopted technologies such as facial recognition, thereby providing real-time, unobtrusive health surveillance at a population scale. Despite these advances, several limitations warrant consideration. First, data heterogeneity arising from differences in sampled populations, data-acquisition devices and collection protocols can impair model transferability. Expanding data collection to more diverse clinical populations or conducting center-specific fine-tuning will be crucial to improving model generalizability. Second, the current study was limited to Chinese populations; validation in other ancestries is needed to ensure global applicability. Finally, while stenosis was used as the primary clinical endpoint, future work should assess the predictive power of facial biomarkers in renal, endocrine, and hematologic diseases, where broader validation is essential. Nevertheless, these limitations represent opportunities for future research, and our results provide strong proof-of-concept evidence for the clinical value of facial biomarker prediction. Collectively, our study establishes a new paradigm for precision health: that a single facial photograph can serve as a non-invasive, multidimensional window into systemic physiology. By enabling scalable biomarker prediction, robust disease risk evaluation, and seamless smartphone deployment, our work paves the way toward automated, real-time, and population-wide health monitoring. We envision that facial foundation models, integrated into ubiquitous digital platforms, could transform preventive medicine and democratize access to personalized healthcare on a global scale. Methods Facial image acquisition To develop FaceFound, we curated a large-scale dataset comprising both public and clinical facial image datasets. The public datasets included ImageNet-22k 27 and VGGFace2-HQ 28,29 . The ImageNet-22k dataset, containing 14,197,122 real-world images, was preprocessed according to standard protocols 27 . The VGGFace2HQ dataset (1,160,250 public facial images) was aligned and normalized as standardized procedures 28 , 29 . The clinical facial datasets for pretraining were collected from patients at Beijing Anzhen Hospital between 2021 and Q2 2023 (AZ-TR unfiltered), all facial photographs were obtained by a physician who was blinded to the study design. A digital camera with a resolution of at least 20 megapixels was used under P mode, enabled with an ISO sensitivity of 1600 and burst mode. Images were captured in a quiet examination room with a plain white background and direct lighting to ensure privacy and consistency. Participants were instructed to keep their eyes open, maintain a neutral facial expression, and avoid accessories or hair covering facial features. Each participant was photographed in frontal, left and right 60° lateral, and downward-facing positions (exposing the forehead), with 3–5 photographs per angle. These multiple angles were collected during pretraining to comprehensively capture facial information. In the fine-tuning stage for predicting 62 biomarkers, we used only a single frontal photograph, which achieved comparable predictive performance and offered greater practicality and translational value for downstream clinical applications. To fine-tune FaceFound, for predicting 62 biomarkers routinely tested in a healthcare system across eight physiological systems, we collected 4,698 individuals with linked hospital records from AZ-TR-unfiltered dataset (N = 4,698, median age 61.0 years, IQR 54.0–67.0; 26.4% female). All patients in the internal cohort of AZ-TR were divided randomly into training, validation and testing datasets in ratio of 70%:20%:10%. For external validation, we collected five independent cohorts with the same protocols, in which image acquisition was performed by different physicians blinded to the study content, and only frontal facial photographs were obtained. In the second half of 2023, we enrolled 1,498 participants at Anzhen Hospital (AZ-EV1; median age 61.0 years, IQR 54.0–67.0; 24.4% female). In 2024, an additional 2,247 participants were recruited at the same hospital (AZ-EV2; median age 60.0 years, IQR 53.0–66.0; 22.2% female). Between 2021 and 2022, 900 participants were recruited at Beijing Daxing Hospital (DX-EV; median age 62.0 years, IQR 56.0–69.0; 49.9% female). In 2025, we enrolled 206 participants from the Tongzhou branch of Anzhen Hospital (AZTZ-EV; median age 61.0 years, IQR 56.0–68.0; 39.8% female), with one facial image per individual. Finally, the AZ-GEV cohort comprised 111 participants (median age 59.0 years, IQR 52.0–65.5; 30.6% female), with one facial image per individual. Biomarkers variables For the internal cohort, a total of 62 biomarkers were collected in the AZ-TR cohort, with the available sample size varying across biomarkers as determined by treating clinicians and standard institutional procedures. For blood sampling, venous blood was obtained in the morning after an overnight fast. Samples were stored at room temperature protected from light and delivered to the hospital laboratory within 30 minutes for routine assays. For the external cohorts, identical procedures were followed. For AZ-TR and DX-EV, we also collected coronary angiography (CAG) data to calculate stenosis, which further aided in developing a disease risk evaluation model. For AZ-GEV, we further collected peripheral blood samples to genotype for calculating genetics scores of biomarkers. Informed consent was obtained from all participants. The study was conducted in accordance with the principles of the Declaration of Helsinki and was approved by the Institutional Review Board (Ethical Approval No. 2022211X). Facial image preprocessing Facial photographs were first linked to the corresponding medical records by extracting patient identifiers through Optical Character Recognition (OCR) using “paddleOCR” (v3.0.1) 34 . When OCR failed, manual correction was performed to ensure accurate matching. Head pose was estimated using 6DRepNet 35 , and only frontal images were retained for the prediction of 62 biomarkers, while images from all angles were used during pretraining. The final dataset included only images successfully matched to hospital records based on participant name, patient ID, and timestamp. Blurry images were excluded. Remaining images were processed with DeepFace 36 for face detection, alignment, and cropping, and were subsequently standardized to 256 × 256 pixels for model training and evaluation. Development of FaceFound model We adopted the Swin Transformer backbone 37 for FaceFound, specifically selecting the Swin-Large variant for its scalability across input resolutions during pre-training and its efficiency in extracting features from high-resolution facial images. Swin-Large is organized into four hierarchical stages with 2, 2, 18, and 2 Swin Transformer Blocks in the respective stages. Input facial photographs were partitioned into non-overlapping 4×4 patches (initial feature size 4×4×3), which are linearly projected to 192-dimensional embeddings and propagated through successive transformer stages. Both the embedding dimensionality and the number of attention heads increased with depth. Within each stage, multi-head self-attention operates in local windows, alternating between standard window-based multi-head self-attention (W-MSA) and shifted-window multi-head self-attention (SW-MSA) to enhance cross-window information flow. On this backbone, we developed a facial foundation model called FaceFound using a General-to-Clinical self-supervised pre-training strategy based on the Uniform Masking Masked Autoencoder (UM-MAE) framework 38 . UM-MAE, which is well suited to pyramid-based vision transformers such as Swin, reconstructs uniformly sampled masked patches with secondary masking to improve locality awareness and feature robustness. Compared with alternatives such as SimMIM 15 , UM-MAE offers greater computational efficiency—reducing pre-training time and GPU memory consumption—while maintaining strong fine-tuning performance. Pre-training proceeded in three sequential stages, progressively specializing the representations from generic vision to clinically relevant facial features. Stage 1 (general visual pre-training). FaceFound was initialized with UM-MAE weights trained on ImageNet-22K, using 256×256 inputs with a 4×4 patch size and an 8-pixel window, to establish broad visual representations. Stage 2 (facial feature pre-training) . FaceFound was adapted to human facial characteristics using VGGFace2-HQ, a high-resolution dataset (512×512 pixels) comprising 9,131 identities derived from VGGFace2 via super-resolution. Inputs are 512×512 with a 4×4 patch size and a 16-pixel window, refining sensitivity to facial morphology and fine-grained details. Stage 3 (target-population pre-training). Finally, FaceFound was further specialized on a proprietary clinical dataset of 88,409 facial images from patients at Beijing Anzhen Hospital, using 256×256 inputs, to capture population-specific characteristics relevant to downstream clinical biomarker prediction. This hierarchical curriculum—from natural images (ImageNet-22K), to facial imagery (VGGFace2-HQ), to clinical patient photographs—enabled a gradual transition from general visual features to task-specific clinical representations. The consistent use of UM-MAE across stages yielded an efficient self-supervised pipeline and a robust facial foundation model tailored for predicting systemic biomarkers from facial photographs. Fine-tuning and evaluation for downstream tasks For downstream task adaptation, we used only the Swin-L encoder of the foundation model as the backbone of FaceFound, discarding the decoder of UM-MAE. The backbone extracted high-level facial representations, which were passed into a multilayer perceptron (MLP) to generate task-specific predictions. Depending on the downstream task, the outputs were defined as different biomarker levels. We conducted supervised fine-tuning on all parameters of the Swin-L backbone and the MLP for 50 epochs with a batch size of 32, using AdamW optimizer. The first 10 epochs employed a linear warm-up schedule, gradually increasing the learning rate from 0 to 5 × 10 − 3 , followed by a cosine annealing decay schedule that reduced the learning rate from 5 × 10 − 3 to 1 × 10 − 6 over the remaining 40 epochs. We stopped the training early when the validation loss does not decrease for 16 consecutive epochs. After each epoch, models were evaluated on the validation set, and the checkpoint with the lowest validation loss was retained for subsequent testing on both internal and external cohorts. Model performance was evaluated using the coefficient of determination (R 2 ) as the primary metric for regression tasks. Other deep learning model implementations We implemented our proposed General-to-Clinical pretraining strategy, leveraging the UM-MAE self-supervised pretraining framework, to pretrain the Swin-L backbone of FaceFound. We then compare this General-to-Clinical approach with a baseline strategy that involves pretraining the model solely on ImageNet-22K. To enable a fair comparison, we also pretrain a Swin-L backbone using UM-MAE on ImageNet-22K as a pretraining strategy comparison. For architectural comparison, we incorporate the well-established ResNet18 39 model, which we pretrain on ImageNet-22K. Genotyping and construction of polygenic risk scores Peripheral blood samples from AZ-GEV participants were genotyped using the Beadchip Array Asian Screening Array. Standard quality control was applied, including filters for sample and variant call rates, Hardy–Weinberg equilibrium, and minor allele frequency. Genotype imputation was performed on the Michigan Imputation Server 40 with the Haplotype Reference Consortium panel (HRC r1.1 2016), using Eagle v2.4 for phasing and Minimac4 for imputation. Variants with imputation INFO ≤ 0.8 or R² ≤0.8 were excluded, leaving 5,016,187 high-quality variants for PRS calculation. Polygenic risk scores (PRS) were constructed for each individual using summary statistics from previously published 679 genome-wide association studies (GWAS) relevant to 27 biomarker traits from PGS Catalog 41 . Scores were calculated as the weighted sum of risk alleles carried by each individual, with weights corresponding to the effect sizes reported in the GWAS data by PLINK2. Development of risk evaluation models We constructed three stenosis-classification models to evaluate the risk evaluation utility of face-predicted biomarkers compared with conventional laboratory measurements. Model 1 incorporated demographic and anthropometric variables (age, sex, height, weight) with laboratory-measured lipid profiles (LDL-C, HDL-C, TC, TG). Model 2 replaced laboratory-measured lipids with their face-predicted counterparts. Model 3 expanded the feature set to include 16 face-predicted cardiovascular biomarkers, thereby evaluating their incremental contribution in the absence of corresponding laboratory measurements in the DX-EV cohort (Fig. 3 c). All models were trained in the AZ-TR cohort and externally validated in the DX-EV cohort with a logistic model, where stenosis severity (≥ 50% vs. < 50%) was available as the clinical endpoint. Multivariable logistic regression was used for model development. To quantify the contribution of individual biomarkers, we applied the “relaimpo” package 42 (v2.2.7) with the “lmg” method to calculate the relative importance of 16 face-predicted versus laboratory-measured biomarkers in predicting stenosis. Test-retest analysis To assess the stability and reproducibility of our facial foundation model, we conducted a test–retest analysis across both internal and external cohorts. For each individual with multiple facial photographs collected, we randomly sampled two images taken on different occasions to form a test–retest pair. For each biomarker, the FaceFound model and other baseline models were applied independently to both images within each pair, and repeated predictions were obtained. Test–retest reliability was quantified using the intraclass correlation coefficient (ICC), calculated separately for repeated predicted biomarker values of the prediction model, and Cohen’s was calculated to evaluate retest reliability using the “pingouin” package (v0.5.5) Label efficiency Label efficiency pertains to the quantity of training data and annotations needed to attain a specified performance level for supervised fine-tuning in a particular downstream task, reflecting the annotation effort required from medical professionals 43 . To assess the label efficiency of FaceFound relative to baseline architectures, we systematically varied the amount of supervised fine-tuning samples in the AZ-TR cohort. Training subsets were randomly down sampled to 400, 800, and 1600 samples without replacement, while using the validation and internal testing cohorts in the same manner as the former supervised fine-tuning. Six representative blood biomarkers were selected to cover diverse physiological systems, and models were fine-tuned following the predefined supervised fine-tuning strategy. In total, 558 models were trained, including 186 FaceFound models (62 biomarkers × 3 data regimes) and 372 baseline models (Swin-Large and ResNet-18, each trained under the same conditions). Model performance was evaluated on both the internal and three independent external cohorts (AZ-EV1, AZ-EV2, and DX-EV). This design enabled a direct comparison of predictive performance under varying data regimes, thereby quantifying the label efficiency of FaceFound versus conventional models. Explanations for fine-tuned models We applied Gradient-weighted Class Activation Mapping (Grad-CAM) to visualize the predictive mechanisms of fine-tuned models for biomarker estimation. Grad-CAM is a gradient-based interpretability method that generates class-discriminative localization maps by weighting activations at a selected layer with the gradients of a target output, thereby highlighting input regions most influential for prediction. For each fine-tuned model, the first LayerNorm layer in the final block of each stage (i.e., the 2nd, 4th, 22nd, and 24th blocks) was designated as the target layer and the single continuous output was treated as the target “class.” Mobile application deployment and validation We developed a WeChat-based mobile application using native WeChat components and the Vant Weapp framework to support user interaction and facial image collection. The backend was implemented in Golang with MySQL and Redis for data management and caching. FaceFound was deployed on an Ubuntu 22.04 server with GPU, where the backend was compiled into a standalone binary and managed via Systemd. Secure communication was ensured through HTTPS encryption, JWT-based authentication, and SSL-encrypted database connections. To evaluate the feasibility of applying the model in real-world mobile settings, an independent validation cohort was established. Facial images were collected by a physician who was blinded to the study design (N = 110), using the front-facing camera of an iPhone 15 Pro Max. Other than a plain white background to ensure basic consistency, no restrictions were imposed on the shooting environment. One frontal image was collected per participant, without constraints on facial expression. This cohort was used specifically to validate model performance on smartphone-acquired images, thereby assessing its practicality and translational potential for broader deployment in clinical and community-based contexts. Statistical analysis Baseline characteristics were summarized as mean (standard deviation, SD) for normally distributed continuous variables, median (interquartile range, IQR) for non-normally distributed continuous variables, and frequency (percentage) for categorical variables. To evaluate the performance of facial prediction, the coefficient of determination (R2) was calculated from regression models comparing face-predicted values with corresponding laboratory measurements. P values were adjusted for multiple testing using the Benjamini-Hochberg false discovery rate (FDR-BH) method. Biomarkers with adjusted P ≥ 0.05 were considered to have no significant predictive ability. For association analysis of facial predicted values and laboratory measurement values on stenosis, we used logistic regression with covariates of age, sex, height, weight, and BMI. A two-sided P ≥ 0.05 indicates statistical significance. For risk evaluation models, we assessed performance using the area under the receiver operating characteristic curve (AUC) and the area under the precision-recall curve (APR), as well as accuracy, sensitivity, and specificity with optimal cut-off values were determined by the maximum Youden index. 95% CI for performance metrics was estimated using bootstrap resampling with 1,000 iterations. All statistical analyses were conducted in “Python” (v3.10.14) and “statsmodels” (v0.14.2). Figures were generated with the “matplotlib” package (v3.10.0). Data availability The predicted biomarker scores, polygenic risk scores (PRS), and corresponding labels that do not contain any personally identifiable information are available upon reasonable request to the corresponding author. Access to raw facial images is restricted to protect participant privacy and can be granted only to qualified investigators with appropriate institutional and ethical approvals and a justified scientific purpose. The PGS variant-level weights used is this study are available for download through the PGS Catalog ( https://www.pgscatalog.org/ ). Declarations Data availability The predicted biomarker scores, polygenic risk scores (PRS), and corresponding labels that do not contain any personally identifiable information are available upon reasonable request to the corresponding author. Access to raw facial images is restricted to protect participant privacy and can be granted only to qualified investigators with appropriate institutional and ethical approvals and a justified scientific purpose. The PGS variant-level weights used is this study are available for download through the PGS Catalog ( https://www.pgscatalog.org/ ). Code availability All software used in this study is publicly available. The custom code developed and used to train and fine-tune the FaceFound model for biomarker prediction, as well as the analysis pipelines used in this paper, is available at https://github.com/1511878618/FaceFound . Competing interests The authors declare no competing interests. References Mayeux R, Biomarkers (2004) Potential uses and limitations. Neurotherapeutics 1:182–188 Zhang WR, Parikh CR (2019) Biomarkers of Acute and Chronic Kidney Disease. Annu Rev Physiol 81:309–333 Hippisley-Cox J et al (2024) Development and validation of a new algorithm for improved cardiovascular risk prediction. Nat Med 30:1440–1447 Cummings JL et al (2025) Biomarker-guided decision making in clinical drug development for neurodegenerative disorders. Nat Rev Drug Discov 24:589–609 Low-cost (2022) point-of-care biomarker quantification. Curr Opin Biotechnol 76:102738 Macchia E et al (2024) Point-Of-Care Ultra-Portable Single-Molecule Bioassays for One-Health. Adv Mater 36:2309705 Rim TH et al (2020) Prediction of systemic biomarkers from retinal photographs: development and validation of deep-learning algorithms. Lancet Digit Health 2:e526–e536 Poplin R et al (2018) Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. Nat Biomed Eng 2:158–164 Singal A et al (2025) Part I. Dermatologic Manifestations in Patients with Kidney Disease. J Am Acad Dermatol S0190–9622(25):02360–02366. 10.1016/j.jaad.2025.06.035 Nixon-Hill M, Mookerjee RP, Leung TS (2023) Assessment of bilirubin levels in patients with cirrhosis via forehead, sclera and lower eyelid smartphone images. PLOS Digit Health 2:e0000357 Cao K et al (2023) Large-scale pancreatic cancer detection via non-contrast CT and deep learning. Nat Med 29:3033–3043 He K et al (2022) IEEE, New Orleans, LA, USA,. Masked Autoencoders Are Scalable Vision Learners. in 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 15979–15988 10.1109/CVPR52688.2022.01553 Azad B et al (2023) Foundational Models in Medical Imaging: A Comprehensive Survey and Future Vision. Preprint at http://arxiv.org/abs/2310.18689 Avram O et al (2024) Accurate prediction of disease-risk factors from volumetric medical scans by a deep vision model pre-trained with 2D scans. Nat Biomed Eng. https://doi.org/10.1038/s41551-024-01257-9 Xie Z et al (2022) SimMIM: a Simple Framework for Masked Image Modeling. in. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 9643–9653 (IEEE, New Orleans, LA, USA, 2022). 10.1109/CVPR52688.2022.00943 Zhou Y et al (2023) A foundation model for generalizable disease detection from retinal images. Nature 1–8. 10.1038/s41586-023-06555-x Lu MY et al (2024) A visual-language foundation model for computational pathology. Nat Med 30:863–874 Christensen M, Vukadinovic M, Yuan N, Ouyang D (2024) Vision–language foundation model for echocardiogram interpretation. Nat Med 30:1481–1488 Xiang J et al (2025) A vision–language foundation model for precision oncology. Nature 638:769–778 Ma B, Zhang W, Qiu F, Ding Y A Unified Approach to Facial Affect Analysis: the MAE-Face Visual Representation. in (2023) IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) 5924–5933 (IEEE, Vancouver, BC, Canada, 2023). 10.1109/CVPRW59228.2023.00630 Serengil S, Özpınar AA (2024) Benchmark of Facial Recognition Pipelines and Co-Usability Performances of Modules. Bilişim Teknolojileri Dergisi 17:95–107 Shahreza HO, George A, Marcel S (2025) Face Reconstruction from Face Embeddings using Adapter to a Face Foundation Model. in 5584–5593 Qin L et al (2023) SwinFace: A Multi-task Transformer for Face Recognition, Expression Recognition, Age Estimation and Attribute Estimation. IEEE Trans Circuits Syst Video Technol 1–1. 10.1109/TCSVT.2023.3304724 Lin S et al (2020) Feasibility of using deep learning to detect coronary artery disease based on facial photo. Eur Heart J 41:4400–4411 Hsieh T-C et al (2022) GestaltMatcher facilitates rare disease matching using facial phenotype descriptors. Nat Genet 54:349–357 Zhou J et al (2025) Facial Foundational Model Advances Early Warning of Coronary Artery Disease from Live Videos with DigitalShadow. 04.29.25326370 Preprint at https://doi.org/10.1101/2025.04.29.25326370 (2025) Deng J et al (2009) ImageNet: A large-scale hierarchical image database. in. IEEE Conference on Computer Vision and Pattern Recognition 248–255 (2009). 10.1109/CVPR.2009.5206848 Cao Q, Shen L, Xie W, Parkhi OM, Zisserman A (2018) VGGFace2: A dataset for recognising faces across pose and age. Preprint at. https://doi.org/10.48550/arXiv.1710.08092 Chen X et al (2024) SimSwap++: Towards Faster and High-Quality Identity Swapping. IEEE Trans Pattern Anal Mach Intell 46:576–592 Bontempi D et al (2025) FaceAge, a deep learning system to estimate biological age from face photographs to improve prognostication: a model development and validation study. Lancet Digit Health 7:100870 Verma A et al (2024) Diversity and scale: Genetic architecture of 2068 traits in the VA Million Veteran Program. Science 385:eadj1182 Zhang M et al (2022) Genetic variants underlying differences in facial morphology in East Asian and European populations. Nat Genet 54:403–411 MacDougall DE et al (2025) Lipoprotein(a) and recurrent atherosclerotic cardiovascular events: the US Family Heart Database. Eur Heart J ehaf297. 10.1093/eurheartj/ehaf297 Cui C et al (2025) PaddleOCR 3.0 Technical Report. Preprint at https://doi.org/10.48550/arXiv.2507.05595 Hempel T, Abdelrahman AA, Al-Hamadi A (2023) Towards Robust and Unconstrained Full Range of Rotation Head Pose Estimation. Preprint at https://doi.org/10.48550/arXiv.2309.07654 Serengil SI, Ozpinar A, LightFace: (2020) A Hybrid Deep Face Recognition Framework. in 2020 Innovations in Intelligent Systems and Applications Conference (ASYU) 1–5 10.1109/ASYU50717.2020.9259802 Liu Z et al (2021) Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows. in 10012–10022 Li X, Wang W, Yang L, Yang J (2022) Uniform Masking: Enabling MAE Pre-training for Pyramid-based Vision Transformers with Locality. Preprint at https://doi.org/10.48550/arXiv.2205.10063 He K, Zhang X, Ren S, Sun J (2016) Deep Residual Learning for Image Recognition. in 770–778 Das S et al (2016) Next-generation genotype imputation service and methods. Nat Genet 48:1284–1287 Lambert SA et al (2024) Enhancing the Polygenic Score Catalog with tools for score calculation and ancestry normalization. Nat Genet 56:1989–1994 Groemping U (2007) Relative Importance for Linear Regression in R: The Package relaimpo. J Stat Softw 17:1–27 Jin C, Guo Z, Lin Y, Luo L, Chen H (2025) Label-Efficient Deep Learning in Medical Image Analysis: Challenges and Future Directions. Preprint at https://doi.org/10.48550/arXiv.2303.12484 Additional Declarations There is NO Competing Interest. Supplementary Files SupplementaryTables.xlsx Supplementary Tables nrreportingsummaryNature.pdf Reporting summary ExtendedDataFigs.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8110055","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":571587406,"identity":"be17a06e-fc94-4347-a502-12fddfca17e6","order_by":0,"name":"Minxian Wang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAvUlEQVRIiWNgGAWjYPACGyDmIU1LGulaDpOgxeD42cOvbradlzdnP3uA4UcNg7w5QS1n8tKsc87cNtzZk5fA2HOMwXBnAwEtZgdyzIxzKm4nGNzgMWDgbWBIMDhASMv5N0AtBufAWhj/EqXlRo7x45yKA2AtzETZYn/jjRlzzplkww1n8hIOyxyTMNxASItkf47x59w2O3lg0B18+KbGRp6gLUDAJgFjARVL4FGIAMwfiFI2CkbBKBgFIxcAAPozQDEBsfpgAAAAAElFTkSuQmCC","orcid":"","institution":"Beijing Institute of Genomics, Chinese Academy of Sciences. (China National Center for Bioinformation)","correspondingAuthor":true,"prefix":"","firstName":"Minxian","middleName":"","lastName":"Wang","suffix":""},{"id":571587407,"identity":"c44e010c-1c5e-4615-b1a7-328b8d380d1e","order_by":1,"name":"Tingfeng Xu","email":"","orcid":"","institution":"Beijing Institute of Genomics, Chinese Academy of Sciences. (China National Center for Bioinformation)","correspondingAuthor":false,"prefix":"","firstName":"Tingfeng","middleName":"","lastName":"Xu","suffix":""},{"id":571587408,"identity":"b0616315-f159-44a8-83a5-029090633f34","order_by":2,"name":"Huixuan Xu","email":"","orcid":"","institution":"Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China.","correspondingAuthor":false,"prefix":"","firstName":"Huixuan","middleName":"","lastName":"Xu","suffix":""},{"id":571587409,"identity":"0297386b-c1a0-42ae-bc38-39ca5c2c16fb","order_by":3,"name":"Li Lin","email":"","orcid":"","institution":"Beijing Anzhen Hospital, Capital Medical University","correspondingAuthor":false,"prefix":"","firstName":"Li","middleName":"","lastName":"Lin","suffix":""},{"id":571587410,"identity":"aee12cda-6a2c-4864-9b16-4792b9358001","order_by":4,"name":"Fei Wang","email":"","orcid":"","institution":"Beijing Institute of Genomics, Chinese Academy of Sciences. (China National Center for Bioinformation)","correspondingAuthor":false,"prefix":"","firstName":"Fei","middleName":"","lastName":"Wang","suffix":""},{"id":571587411,"identity":"91f1848e-e98b-4fa4-9105-67b50a3e320e","order_by":5,"name":"Yaodong Ding","email":"","orcid":"","institution":"Beijing An Zhen Hospital: Capital Medical University Affiliated Anzhen Hospital","correspondingAuthor":false,"prefix":"","firstName":"Yaodong","middleName":"","lastName":"Ding","suffix":""},{"id":571587412,"identity":"b0b523c2-ad27-421a-ac7f-646c57b6217d","order_by":6,"name":"Shuangqiao Liao","email":"","orcid":"","institution":"Beijing Institute of Genomics, Chinese Academy of Sciences. (China National Center for Bioinformation)","correspondingAuthor":false,"prefix":"","firstName":"Shuangqiao","middleName":"","lastName":"Liao","suffix":""},{"id":571587413,"identity":"bada3d73-08fc-4659-8ef2-8ca59c927593","order_by":7,"name":"Xingyu Chen","email":"","orcid":"","institution":"Beijing Institute of Genomics, Chinese Academy of Sciences. (China National Center for Bioinformation)","correspondingAuthor":false,"prefix":"","firstName":"Xingyu","middleName":"","lastName":"Chen","suffix":""},{"id":571587414,"identity":"311bfa55-2fd3-4edc-808e-225ca4972c71","order_by":8,"name":"Jichen Wen","email":"","orcid":"","institution":"Beijing Institute of Genomics, Chinese Academy of Sciences. (China National Center for Bioinformation)","correspondingAuthor":false,"prefix":"","firstName":"Jichen","middleName":"","lastName":"Wen","suffix":""},{"id":571587415,"identity":"90004a43-7bc4-4a69-974e-f43200b0d21b","order_by":9,"name":"Hongqiang Zhao","email":"","orcid":"","institution":"Beijing Institute of Genomics, Chinese Academy of Sciences. (China National Center for Bioinformation)","correspondingAuthor":false,"prefix":"","firstName":"Hongqiang","middleName":"","lastName":"Zhao","suffix":""},{"id":571587416,"identity":"58340385-bcf0-4f1f-b490-27d96644e7dc","order_by":10,"name":"Jing Hao","email":"","orcid":"","institution":"Beijing Institute of Genomics, Chinese Academy of Sciences. (China National Center for Bioinformation)","correspondingAuthor":false,"prefix":"","firstName":"Jing","middleName":"","lastName":"Hao","suffix":""},{"id":571587417,"identity":"d7767492-74c8-4c8b-b0e0-bba468b7715f","order_by":11,"name":"Weixiang Zhao","email":"","orcid":"","institution":"Beijing Institute of Genomics, Chinese Academy of Sciences. (China National Center for Bioinformation)","correspondingAuthor":false,"prefix":"","firstName":"Weixiang","middleName":"","lastName":"Zhao","suffix":""},{"id":571587418,"identity":"84382d9d-c9e8-4c7d-acfa-6f4e2be16c29","order_by":12,"name":"Gong Zhang","email":"","orcid":"","institution":"Beijing Daxing Hospital of Capital Medical University","correspondingAuthor":false,"prefix":"","firstName":"Gong","middleName":"","lastName":"Zhang","suffix":""},{"id":571587419,"identity":"ad84338c-5a2d-4309-b483-0387e7998302","order_by":13,"name":"Shiguang Shan","email":"","orcid":"","institution":"Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China.","correspondingAuthor":false,"prefix":"","firstName":"Shiguang","middleName":"","lastName":"Shan","suffix":""},{"id":571587420,"identity":"0c38af17-e184-4cc8-914e-06d5b0667b67","order_by":14,"name":"Yong Zeng","email":"","orcid":"https://orcid.org/0000-0001-6983-0799","institution":"Capital Medical University, Beijing Anzhen Hospital, Capital Medical University, Beijing, 102600, China.","correspondingAuthor":false,"prefix":"","firstName":"Yong","middleName":"","lastName":"Zeng","suffix":""},{"id":571587421,"identity":"94fee094-c786-43de-8bd1-7cf1ad74913e","order_by":15,"name":"Hu Han","email":"","orcid":"https://orcid.org/0000-0001-6010-1792","institution":"Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China.","correspondingAuthor":false,"prefix":"","firstName":"Hu","middleName":"","lastName":"Han","suffix":""}],"badges":[],"createdAt":"2025-11-14 03:05:48","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8110055/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8110055/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":100779136,"identity":"0cc61c14-1e69-4c70-b8ab-064305c04520","added_by":"auto","created_at":"2026-01-21 11:33:52","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6521243,"visible":true,"origin":"","legend":"","description":"","filename":"Manuscript.docx","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/8f39e596df18c5711e58346a.docx"},{"id":100779242,"identity":"85e3d951-5249-4189-a96e-b3d4b9e06032","added_by":"auto","created_at":"2026-01-21 11:35:00","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":16675,"visible":true,"origin":"","legend":"","description":"","filename":"NCOMMS26000706T.json","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/3cc8b5891a485e9346dbd339.json"},{"id":100779102,"identity":"3146309e-a07b-4785-9633-e24e943351a9","added_by":"auto","created_at":"2026-01-21 11:32:55","extension":"pdf","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1854879,"visible":true,"origin":"","legend":"","description":"","filename":"nrreportingsummaryNature.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/84ff4aeff6fd154f264b11f1.pdf"},{"id":100779138,"identity":"1870132b-00ea-4fa6-b8fd-caf23d877aa9","added_by":"auto","created_at":"2026-01-21 11:33:55","extension":"xml","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":128169,"visible":true,"origin":"","legend":"","description":"","filename":"NCOMMS26000706T0enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/82ae90be12444940b4c7246d.xml"},{"id":100779298,"identity":"71e28b5f-fae9-415d-9851-411219edb693","added_by":"auto","created_at":"2026-01-21 11:35:27","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":514139,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/22039d9cd6c6769381c0cbbb.png"},{"id":100779362,"identity":"f526316c-983c-4b82-921e-e73f27e7baf3","added_by":"auto","created_at":"2026-01-21 11:35:47","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":85827,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/320b763d3b2722ed2084ffb5.png"},{"id":100779276,"identity":"d431cf6e-329e-4975-a8a4-9fa8b4968f59","added_by":"auto","created_at":"2026-01-21 11:35:10","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":355882,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage11.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/fc9037b8de0f3191d3ea03dc.png"},{"id":100779293,"identity":"65b369d3-bd17-4f51-a852-8721053ceb65","added_by":"auto","created_at":"2026-01-21 11:35:21","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":199735,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage12.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/44f483933ad083394e0ffdc5.png"},{"id":100779212,"identity":"13a98973-bf86-4160-b2a1-177a7c6747b7","added_by":"auto","created_at":"2026-01-21 11:34:49","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2535183,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage13.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/8687b9a7142980043aede4d2.png"},{"id":100779347,"identity":"28e9ba3d-4860-4ef7-91ed-5f77fd2b6bc2","added_by":"auto","created_at":"2026-01-21 11:35:45","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":495181,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/18882881bc16188d29e8a626.png"},{"id":100779103,"identity":"829f26c5-2700-4970-ac82-d81e8af0b940","added_by":"auto","created_at":"2026-01-21 11:32:55","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":180652,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/461059a145b67917df654aa2.png"},{"id":100779360,"identity":"a95e7ffd-de76-42d5-83fa-3ca5fe15fd95","added_by":"auto","created_at":"2026-01-21 11:35:46","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":340199,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/b11ffa1f7185e27bebcee17d.png"},{"id":100779567,"identity":"685be231-9b6c-41d0-ab00-8c0828074983","added_by":"auto","created_at":"2026-01-21 11:36:46","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":417469,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/0b8394e59567ec2029d86c88.png"},{"id":100779151,"identity":"504548f9-9e57-4cf4-9c6a-89d6e48d34c2","added_by":"auto","created_at":"2026-01-21 11:34:25","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":295896,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/6fc3074524324d3a0be641be.png"},{"id":100779114,"identity":"faa20cf8-2c5b-41ab-9df5-bdcd8187cbf9","added_by":"auto","created_at":"2026-01-21 11:33:17","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":795667,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/bfcfd2869e7cf90a703115db.png"},{"id":100779380,"identity":"72f0fb27-99db-4e33-a7a2-9c5fa0dd0962","added_by":"auto","created_at":"2026-01-21 11:35:59","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":103123,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/3c8980f07b4e7862e4638a32.png"},{"id":100779106,"identity":"069e7988-3c5d-4b32-83f5-a61a6cdffe6c","added_by":"auto","created_at":"2026-01-21 11:33:00","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":103029,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/322c649ea22479b25d586cf6.png"},{"id":100779418,"identity":"1f5766e6-ed93-4eb2-88e4-7d10f1f47809","added_by":"auto","created_at":"2026-01-21 11:36:16","extension":"png","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":109631,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/6a0407ac9746ca741b32f284.png"},{"id":100779281,"identity":"60b0e6dd-1be6-4f28-a817-5eeb79db230e","added_by":"auto","created_at":"2026-01-21 11:35:13","extension":"png","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":16574,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/4abf4e9fe9bf361eb98ac3b5.png"},{"id":100779105,"identity":"d271fd38-aa8d-4de0-a928-358ab41e0c44","added_by":"auto","created_at":"2026-01-21 11:32:59","extension":"png","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":104059,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage11.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/6871778d85bac4cf0eeca6f1.png"},{"id":100779366,"identity":"f2a009dc-7e40-4864-9a81-fd0de7cc02fc","added_by":"auto","created_at":"2026-01-21 11:35:48","extension":"png","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":50089,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage12.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/bec9e4490721fd2dd8f38a51.png"},{"id":100779104,"identity":"38a3f909-d93a-467c-b092-b15e538e5049","added_by":"auto","created_at":"2026-01-21 11:32:56","extension":"png","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":268667,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage13.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/a1f76c0e3e34f352e3668980.png"},{"id":100779289,"identity":"e05a1719-74f0-4a80-9af6-fa4e74e13df2","added_by":"auto","created_at":"2026-01-21 11:35:20","extension":"png","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":93872,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/3bf996a3cc523c5469989d30.png"},{"id":100804110,"identity":"67a046ec-0260-4aab-8372-88612bcb8c8a","added_by":"auto","created_at":"2026-01-21 14:37:27","extension":"png","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":27186,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/283fb8514eaaf858b75df665.png"},{"id":100779331,"identity":"ae6d6917-ce11-45d3-b7f9-593e1f7ff462","added_by":"auto","created_at":"2026-01-21 11:35:35","extension":"png","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":68529,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/5ae6c8a2aaa6e75955a28f8e.png"},{"id":100779117,"identity":"ee9cfbfd-1db4-4dc3-9982-7ae99ab8ef04","added_by":"auto","created_at":"2026-01-21 11:33:20","extension":"png","order_by":25,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":77670,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/699898d7dc3b0deae5eb5e45.png"},{"id":100779233,"identity":"9d158b96-63af-4dbc-b9a7-d35e53ddffa4","added_by":"auto","created_at":"2026-01-21 11:34:53","extension":"png","order_by":26,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":47074,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/23bbdc2e0ec5499d4d5b23ab.png"},{"id":100779097,"identity":"88832384-c941-49e8-8765-e481c233c1f8","added_by":"auto","created_at":"2026-01-21 11:32:53","extension":"png","order_by":27,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":109273,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/4c789b0ef5f3910b9495674c.png"},{"id":100779274,"identity":"6a3d8846-7f45-44fe-adc0-7e034dffb6fe","added_by":"auto","created_at":"2026-01-21 11:35:09","extension":"png","order_by":28,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":21678,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/d1cdb5b0b245ca91cf29581c.png"},{"id":100779116,"identity":"b49dbc1e-352f-4301-90c5-d464ae3b1877","added_by":"auto","created_at":"2026-01-21 11:33:19","extension":"png","order_by":29,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":22767,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/eb09f078eaad7f44a1448921.png"},{"id":100779133,"identity":"f7c3e0f8-6e47-4ef5-befe-54bd58115d27","added_by":"auto","created_at":"2026-01-21 11:33:40","extension":"xml","order_by":30,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":126035,"visible":true,"origin":"","legend":"","description":"","filename":"NCOMMS26000706T0structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/7405bbcbd8426c898d3fafce.xml"},{"id":100779153,"identity":"f34ca531-2e8f-44b5-9825-2338c115d0dc","added_by":"auto","created_at":"2026-01-21 11:34:28","extension":"html","order_by":31,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":137814,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/ca2d61ae12e04068c07dffa3.html"},{"id":100779100,"identity":"ad6854c0-c3df-4414-89cc-0cbbb52c0711","added_by":"auto","created_at":"2026-01-21 11:32:55","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":288301,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eStudy design.\u003c/strong\u003e \u003cstrong\u003ea,\u003c/strong\u003e Overview of the model pertaining and the fine-tuning process. To develop the foundation model, we utilize a progressive general-to-clinical strategy for model pretraining by integrating 15.4 million general and clinical facial photographs. After pretraining, the model was fine-tuned and evaluated on the AZ-TR datasets (N = 4,698) to predict systemic biomarkers from facial images, encompassing the anthropometric, cardiovascular, hematology, endocrine, electrolyte, digestive, and immune systems. \u003cstrong\u003eb\u003c/strong\u003e, External validation across five independent datasets. The fine-tuned models for 62 biomarkers were externally validated in four independent cohorts: AZ-EV1 (N = 1,498), AZ-EV2 (N = 2,247), AZTZ-EV1 (N = 206), DX-EV1 (N = 900) and AZ-GEV (N = 111). For each cohort, the models were applied to predict all systemic biomarkers, and performance was evaluated by R\u003csup\u003e2\u003c/sup\u003e. \u003cstrong\u003ec, \u003c/strong\u003eComparison of the performance on systematic biomarker prediction between FaceFound and genetic model of polygeneric score (PGS) in an independent cohort AZ-GEV based on R². \u003cstrong\u003ed,\u003c/strong\u003e Deployment of the FaceFound on smartphones and implementation of clinical assessment in this study to validate the real-world clinical utility of face-based biomarker prediction in an independent cohort taken by smartphones (N = 110, APP-EV)\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/3323af8e7fdbb7868b0feda9.png"},{"id":100779410,"identity":"916ab3ac-54c7-4769-a4f3-42fa12644688","added_by":"auto","created_at":"2026-01-21 11:36:08","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":262462,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFaceFound performance across internal and external validation cohorts for multi-system biomarker prediction from facial images. a, \u003c/strong\u003eFaceFound performance across multiple centers for each physiological system, with color intensity indicating R² values. \u003cstrong\u003eb,\u003c/strong\u003e Performance of biomarkers (with R² \u0026gt; 0.1 in the internal set and in at least one external validation cohort) in both internal and external validation cohorts from our model. \u003cstrong\u003ec,\u003c/strong\u003e Forest plots summarizing meta-analysis results of predictive performance (R² with 95% CI) for Hb, eGFR(CKD-EPI), RBC, and HDLC, respectively, across internal and external cohorts. \u003cstrong\u003ed, \u003c/strong\u003eComparison of prediction accuracy (R²) for 62 biomarkers across eight physiological systems (Anthropometrics, Renal, Hematology, Cardiovascular, Hepatobiliary, Digestive, Immune, Endocrine) between our model (FaceFound) and baseline models (SwinLarge and ResNet18) in the internal validation set. \u003cstrong\u003ee,\u003c/strong\u003e Number of biomarkers for which our model achieved state-of-the-art (SOTA) performance by R\u003csup\u003e2\u003c/sup\u003e compared with baseline models across different physiological systems.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/50e8bec40af3c3d70ae82c8c.png"},{"id":100779379,"identity":"f7111053-6337-4cc8-bd94-0b30bce3e3f8","added_by":"auto","created_at":"2026-01-21 11:35:58","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":98873,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eComparison of performance between FaceFound and the polygenic score model. a, \u003c/strong\u003eComparison of prediction accuracy (R²) for 26 biomarkers between FaceFound and the best genetic score model in the external AZ-GEV cohort (n=111). \u003cstrong\u003eb, \u003c/strong\u003eThe number of FaceFound outperformed the genetic model on systematic biomarkers prediction in the external AZ-GEV cohort. \u003cstrong\u003ec, \u003c/strong\u003eComparison of performance on BMI, HDLC, Hb, and eGFR(CKD-EPI) between the top 10 polygenic scores and FaceFound in the external AZ-GEV cohort. \u003cstrong\u003ed, \u003c/strong\u003ePercentage of variance explained by model FaceFound prediction and genetic model for predicting biomarkers in the external AZ-GEV cohort in a multivariable regression model.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/01d6d9998ec3f97ec0c66838.png"},{"id":100779309,"identity":"4d6776c9-f6aa-4876-b6e1-60c41c7cb2a5","added_by":"auto","created_at":"2026-01-21 11:35:35","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":146503,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eEvaluation of FaceFound predicted biomarkers for coronary stenosis detection.\u003c/strong\u003e \u003cstrong\u003ea,\u003c/strong\u003e Conceptual framework illustrating the comparison of FaceFound predicted biomarkers with laboratory measured biomarkers for detecting coronary stenosis (≥50%). \u003cstrong\u003eb\u003c/strong\u003e, Associations of six cardiovascular biomarkers (HDL-C, LDL-C, TG, TC, HbA1C%, hsTnI) with coronary stenosis, presented as odds ratios per SD (OR/SD) with 95% CI from both facial predictions and laboratory measurements with a logistic regression adjusted for age, sex, height, weight and BMI. \u003cstrong\u003ec,\u003c/strong\u003eData availability and feature composition across models and cohorts. Rows indicate cohort or model configurations (Models 1–3), and columns represent included variables. Features of the model are represented as laboratory measurements (red circles) or facial predictions (green circles). Model 1 developed with 4 basic laboratory-measured lipid profiles (LDL-C, HDL-C, TC, TG). Model 2 replaced with 4 face-predicted biomarkers. Model 3 developed with 16 face-derived cardiovascular biomarkers. \u003cstrong\u003ed,\u003c/strong\u003eArea under the receiver operating characteristic curve (AUC) for the three models in the AZ-TR and DX-EV cohorts. \u003cstrong\u003ee,\u003c/strong\u003eReceiver operating characteristic curves comparing the diagnostic performance of the three models in DX-EV cohorts. \u003cstrong\u003ee,\u003c/strong\u003ecoronary stenosis prevalence across high- (\u0026gt;80%), medium- (20%-80%), and low-risk (\u0026lt;20%) strata defined by each model in the DX-EV cohort. \u003cstrong\u003ef,\u003c/strong\u003e Relative importance of selected facially predicted biomarkers compared with their laboratory-measured counterparts with intraclass correlation (ICC) indicating concordance between modalities. For Age, Sex, Height, Weight, and BMI, laboratory-measured values were used. Relative importance was quantified using the LMG method (R² partitioned by averaging contributions over all possible predictor orderings).\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/7fcd54793707b58d9005bc46.png"},{"id":100779341,"identity":"030ccc28-86a5-47f2-ac76-c5e557e545ee","added_by":"auto","created_at":"2026-01-21 11:35:40","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":194790,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eReproducibility and generalizability of FaceFound predicted biomarkers across internal and external cohorts.\u003c/strong\u003e \u003cstrong\u003ea\u003c/strong\u003e, Intraclass correlation coefficients (ICCs) of three models on internal and three external cohorts (AZ-TR, AZ-EV1, AZ-EV2, DX-EV), showing test–retest reliability. Three models represent FaceFound and two baselines SwinLarge and ResNet18 models. \u003cstrong\u003eb\u003c/strong\u003e, FaceFound performance across eight physiological systems on internal and three external cohorts, evaluated by ICC. \u003cstrong\u003ec–f\u003c/strong\u003e, Bland–Altman plots of HDL-C predictions in four cohorts, with dashed lines indicating mean of the differences with 95% limits of agreement (mean ±1.96 SD); ICC values are shown for each cohort with 95% CIs. \u003cstrong\u003eg\u003c/strong\u003e, Prediction accuracy (R²) of 62 blood biomarkers under varying training data sample size (400 – 3200) for three models across cohorts, demonstrating the label efficiency of FaceFound with limited data.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/63baada157d68c099bd54589.png"},{"id":100779310,"identity":"4f58b0b5-e638-4f3f-aba3-ea704265e307","added_by":"auto","created_at":"2026-01-21 11:35:35","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":135523,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSmartphone deployment and validation of FaceFound. a,\u003c/strong\u003e Schematic of the FaceFound algorithm, which takes a single facial photograph as input. The model first localizes the face, then extracts quantitative features to predict 62 biomarkers. \u003cstrong\u003eb,\u003c/strong\u003e Mobile application interface showing, from left to right, the home screen, photo upload/capture screen, and visualization of the interactive report.\u003cstrong\u003e c,\u003c/strong\u003e Prediction accuracy (R²) of FaceFound in the APP-EV cohort (N = 110) using single photographs captured by smartphones. \u003cstrong\u003ed,\u003c/strong\u003e Intraclass correlation coefficients (ICCs) comparing predictions from smartphone photographs with those from internal and four external validation cohorts across 58 biomarkers with available clinical records, demonstrating comparable accuracy between smartphone- and camera-based photographs.\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/d954ce55d8c4b9d4b09112cb.png"},{"id":107706170,"identity":"cb9fd51a-0f5f-44a2-aaf9-9f659620dbd9","added_by":"auto","created_at":"2026-04-24 09:17:34","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1294706,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/047f27de-30c5-4e1f-8f28-e739e75e63b5.pdf"},{"id":100779208,"identity":"4144685a-a369-4f6f-a465-03540d491799","added_by":"auto","created_at":"2026-01-21 11:34:43","extension":"xlsx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":361547,"visible":true,"origin":"","legend":"Supplementary Tables","description":"","filename":"SupplementaryTables.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/dd46bb3728b020ce6c269a97.xlsx"},{"id":100779337,"identity":"1dbd3318-a4cc-4df5-bb2c-4f3111669e46","added_by":"auto","created_at":"2026-01-21 11:35:38","extension":"pdf","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":1854879,"visible":true,"origin":"","legend":"Reporting summary","description":"","filename":"nrreportingsummaryNature.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/41eb8639aab5f5112a01689c.pdf"},{"id":100779287,"identity":"02cd1937-79db-444e-9829-e728f6e2aaa8","added_by":"auto","created_at":"2026-01-21 11:35:19","extension":"docx","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":4196745,"visible":true,"origin":"","legend":"","description":"","filename":"ExtendedDataFigs.docx","url":"https://assets-eu.researchsquare.com/files/rs-8110055/v1/76c3257d8afdccba2fcd0956.docx"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e Competing Interest.","formattedTitle":"A facial foundation model for multi-system biomarker and disease risk prediction with real-world mobile deployment","fulltext":[{"header":"Introduction","content":"\u003cp\u003eQuantifying internal physiological states is fundamental to disease prevention, early diagnosis, and personalized health management\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. Clinical biomarkers are essential indicators that span multiple organ systems and are widely utilized for diagnosis\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e, risk stratification\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e, and informed therapeutic decision-making\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e. Yet, most biomarkers require invasive sampling and laboratory-based testing, which limits accessibility for large-scale screening and frequent monitoring, particularly in resource-constrained settings. Approaches that enable non-invasive, widely deployable biomarker assessment and disease risk evaluation could therefore transform population health screening and personalized care\u003csup\u003e\u003cspan additionalcitationids=\"CR6 CR7\" citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eFacial photographs are ubiquitous, low-cost, and non-invasive images. Facial features have been recognized as reflecting health and disease status; for example, renal dysfunction may manifest as facial edema or pallor\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e, and decompensated cirrhosis may lead to yellowish facial discoloration due to elevated serum bilirubin levels\u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e. Advances in computer vision now enable the extraction of complex phenotypic features from facial images at scale\u003csup\u003e\u003cspan additionalcitationids=\"CR12 CR13\" citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. Vision foundation models, trained on large and diverse unlabeled image datasets, offer a robust framework for learning transferable representations\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e,\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e and have demonstrated success in medical image analysis for various diseases\u003csup\u003e\u003cspan additionalcitationids=\"CR17 CR18\" citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e. Nevertheless, most current studies have focused on narrow facial analysis tasks\u003csup\u003e\u003cspan additionalcitationids=\"CR21 CR22\" citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e and limited health outcomes\u003csup\u003e\u003cspan additionalcitationids=\"CR25\" citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e. The systematic utility of facial images for predicting biomarkers routinely tested in a healthcare system for assessing disease risks across multiple systems remains unexplored.\u003c/p\u003e \u003cp\u003eTo address this gap, we developed FaceFound, a facial foundation model that uses a progressive general-to-clinical pretraining strategy to reduce the domain gap and enhance transferability. FaceFound was pretrained in three stages\u0026mdash;on ImageNet-22K for general visual representations, VGGFace2-HQ for framing facial structures, and clinical facial images (AZ-TR unfiltered) for capturing pathological features (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ea). Fine-tuned on 62 biomarkers, FaceFound was compared with conventional architectures (ResNet18 and SwinLarge) using 4 independent cohorts in terms of prediction accuracy, label efficiency, stability with repeated images, and risk assessment for coronary stenosis (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eb\u003cb\u003e)\u003c/b\u003e. The prediction performance was also compared to polygenic risk score (PRS) models in another external cohort with genetic data available (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ec\u003cb\u003e)\u003c/b\u003e. To translate these insights into a practical application, we deployed FaceFound on smartphones to enable selfie-based data collection and established an new external cohort of self-captured photos to evaluate its performance (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ed). Collectively, this study establishes facial foundation modeling FaceFound as a scalable, non-invasive framework for biomarker prediction with broad implications for precision medicine and digital health.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eFaceFound model development\u003c/h2\u003e \u003cp\u003eWe developed a facial foundation model called FaceFound using self-supervised learning (SSL) with a progressive general-to-clinical strategy. We implemented the general-to-clinical pretraining pipeline utilizing ImageNet-22k (14,197,122 images)\u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e, VGGFace2-HQ (1,160,250 images)\u003csup\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e,\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e, and facial photographs from the Chinese Anzhen Training Cohort (AZ-TR-unfiltered, 88,489 images), Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ea. In the first stage, the Swin-Large Transformer backbone was pretrained on ImageNet-22k with 14.2\u0026nbsp;million images to capture general-purpose visual features. In the second stage, the model was further adapted using approximately 1.16\u0026nbsp;million high-resolution facial images from the VGGFace2-HQ dataset. In the final stage, domain-specific pretraining was conducted using 88,489 clinical facial photographs from the AZ-TR dataset without linking to hospital records (AZ-TR-unfiltered) (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ea, \u003cb\u003eExtended Data\u003c/b\u003e Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, and \u003cb\u003eMethods\u003c/b\u003e). This progressive approach allowed the model to initially learn general visual representations in depth, adapt to facial morphological features, and specialize in disease-relevant representations.\u003c/p\u003e \u003cp\u003eTo fine-tune FaceFound for predicting 62 biomarkers routinely tested in a healthcare system across eight physiological systems (\u003cb\u003eSupplementary Table\u0026nbsp;1\u003c/b\u003e), 4,698 individuals with linked hospital records from AZ-TR-unfiltered dataset (median age 60.0, IQR, 54.0\u0026ndash;67.0, AZ-TR), and each with over three frontal facial photographs and corresponding biomarker measurements were used. Fine-tuning for these biomarkers was performed with task-specific adapter layers (see \u003cb\u003eMethods\u003c/b\u003e for fine-tuning details and \u003cb\u003eExtended Data\u003c/b\u003e Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eEvaluation of FaceFound performance for biomarker prediction in internal and external cohorts\u003c/h3\u003e\n\u003cp\u003eWe evaluated the performance of Facefound model for prediction of 62 biomarkers covering 8 physiological systems and compared its performance with the baseline SwinLarge and ResNet18 models, in one internal validation cohort(AZ-TR, N\u0026thinsp;=\u0026thinsp;4,698) and four external validation cohorts (AZ-EV1, N\u0026thinsp;=\u0026thinsp;1,498; AZ-EV2, N\u0026thinsp;=\u0026thinsp;2,247; AZTZ-EV, N\u0026thinsp;=\u0026thinsp;206; DX-EV, N\u0026thinsp;=\u0026thinsp;900) (details in \u003cb\u003eMethods\u003c/b\u003e and \u003cb\u003eSupplementary Table\u0026nbsp;2\u003c/b\u003e).\u003c/p\u003e \u003cp\u003eWe first examined the predictability of biomarkers across systems based on FaceFound. For 8 physiological systems, we found Anthropometrics (internal cohort: R\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;0.493 (0.361\u0026ndash;0.666), IQR\u0026thinsp;=\u0026thinsp;0.361\u0026ndash;0.666; external cohort: R\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;0.492 (0.410\u0026ndash;0.633), IQR\u0026thinsp;=\u0026thinsp;0.410\u0026ndash;0.633), and Renal (internal cohort: R\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;0.081 (0.038\u0026ndash;0.161), IQR\u0026thinsp;=\u0026thinsp;0.038\u0026ndash;0.161; external cohort: R\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;0.100 (0.037\u0026ndash;0.184), IQR\u0026thinsp;=\u0026thinsp;0.037\u0026ndash;0.184) has the highest predictive accuracy across internal cohort and four external cohorts. While, for Electrolytes (internal cohort: R\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;0.020 (0.013\u0026ndash;0.036), IQR\u0026thinsp;=\u0026thinsp;0.013\u0026ndash;0.036; external cohort: R\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;0.001 (0.000\u0026ndash;0.005), IQR\u0026thinsp;=\u0026thinsp;0.000\u0026ndash;0.005) has the weakest predictive performance (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea and \u003cb\u003eSupplementary Table\u0026nbsp;4\u003c/b\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAmong the 62 biomarkers, 16 achieved an R\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;\u0026gt;\u0026thinsp;0.10 in the internal test cohort (median R\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;0.282, IQR: 0.123\u0026ndash;0.381). Among them, 12 biomarkers reached R\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;\u0026gt;\u0026thinsp;0.10 in at least one external cohort, distributing in Anthropometrics (n\u0026thinsp;=\u0026thinsp;4; Age, height, weight, and BMI), Hematology (n\u0026thinsp;=\u0026thinsp;3; HCT, Hb, and RBC), Cardiovascular (n\u0026thinsp;=\u0026thinsp;2; HDL-C and Hcy), Renal (n\u0026thinsp;=\u0026thinsp;2; UA and eGFR), and Endocrine (n\u0026thinsp;=\u0026thinsp;1; HbA1C), respectively (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb). These findings suggest that systemic health status is reflected in facial characteristics.\u003c/p\u003e \u003cp\u003eAmong these, we found HDL-C (internal cohort: R\u0026sup2; 0.214, 95% CI 0.147\u0026ndash;0.280; external cohort: median R\u0026sup2; 0.185, IQR 0.165\u0026ndash;0.192), HCT (internal cohort: R\u0026sup2; 0.378, 95% CI 0.303\u0026ndash;0.453; external cohort: median R\u0026sup2; 0.260, IQR 0.202\u0026ndash;0.307), RBC (internal cohort: R\u0026sup2; 0.285, 95% CI 0.210\u0026ndash;0.360; external cohort: median R\u0026sup2; 0.228, IQR 0.201\u0026ndash;0.244), and eGFR(CKD-EPI) (internal cohort: R\u0026sup2; 0.289, 95% CI 0.218\u0026ndash;0.359; external cohort: median R\u0026sup2; 0.196, IQR 0.184\u0026ndash;0.260) emerged as the most robust biomarkers predicted by FaceFound, showing strong agreement and consistent meta-analytic R\u003csup\u003e2\u003c/sup\u003e estimates between predicted and measured values across both internal and external cohorts (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ec). These results indicate that facial photographs encode clinically meaningful signals related to lipid metabolism, erythrocyte parameters, and kidney function.\u003c/p\u003e \u003cp\u003eNext, we compared the predictive performance of FaceFound with the two widely used baselines SwinLarge and ResNet18. We found FaceFound outperformed SwinLarge and ResNet18 both in internal and external cohorts. In the internal cohort, FaceFound (median R\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;0.039, IQR: 0.011\u0026ndash;0.085) outperformed SwinLarge (median R\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;0.028, IQR: 0.005\u0026ndash;0.059) and ResNet (median R\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;0.012, IQR: 0.002\u0026ndash;0.044) for prediction of 62 biomarkers. Further, in the four external cohorts, FaceFound performed consistent superiority (median R\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;0.030, IQR: 0.007\u0026ndash;0.071) to SwinLarge (median R\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;0.023, IQR: 0.004\u0026ndash;0.061) and ResNet (median R\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;0.011, IQR: 0.001\u0026ndash;0.035) (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ed). Totally FaceFound achieved state-of-the-art (SOTA) performance in 45 of the 62 biomarkers in the internal cohort (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ee and \u003cb\u003eSupplementary Table\u0026nbsp;5\u003c/b\u003e). Collectively, these findings demonstrate that facial images capture a substantial amount of information across a broad spectrum of systemic biomarkers, particularly those with implications for cardiovascular and renal health.\u003c/p\u003e \u003cp\u003eFor sensitivity analysis, we found that 26 of the 62 predicted biomarkers were independently associated with laboratory measurements in the external AZ-EV1 cohort after adjustment for age, sex, height, weight, and BMI (\u003cb\u003eSupplementary Table\u0026nbsp;6\u003c/b\u003e). Further, to calculate the contribution of facial components and demographic factors in describing variation in biomarker prediction, we developed the multivariable model using both FaceFound and demographic factors (Age, Sex, Height, Weight and BMI). We found that facial features explained a larger proportion of variation (median 65.8%, IQR: 51.3\u0026ndash;75.5%), indicating that facial features contain information relevant to clinical biomarkers, independent of common demographic and anthropometric factors (\u003cb\u003eExtended Data\u003c/b\u003e Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\n\u003ch3\u003eFaceFound and genetic model (PRS) in predicting biomarkers\u003c/h3\u003e\n\u003cp\u003eTo compare the performance of FaceFound and polygenic risk model for prediction of biomarkers, we curated 679 polygenic risk scores (PRS) for 26 biomarkers from the PGS Catalog and computed PRS in the AZ-GEV cohort (N\u0026thinsp;=\u0026thinsp;111, median age 59.0, IQR: 52.0-65.5) with genetic data available. Among the 26 biomarkers, FaceFound outperformed PRS for 14 biomarkers prediction, including BMI, HDL-C, hemoglobin, and eGFR (CKD-EPI), which played important roles in cardiometabolism (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ea, \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb, \u003cb\u003eExtended Data\u003c/b\u003e Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e \u003cb\u003eSupplementary Table\u0026nbsp;7\u003c/b\u003e). Moreover, to compare the prediction performance between genetics and face features, we modeled a multivariable model using both the best PRS and FaceFound score as predictors to predict each of the 26 biomarkers in the AZ-GEV cohort. We found that facial features can independently explain a larger proportion of variation compared to the PRS for 15 biomarkers, whereas the genetics explained more for 11 biomarkers. Together, these findings demonstrate that FaceFound provides predictive performance comparable to or exceeding polygenic risk scores, underscoring the potential of facial images as an accessible and powerful modality for biomarker estimation.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\n\u003ch3\u003eEvaluation of the clinical utility of FaceFound predicted biomarkers\u003c/h3\u003e\n\u003cp\u003eNext, we assessed the clinical values of FaceFound predicted biomarkers compared to laboratory measurements by examining their associations with coronary stenosis severity (\u0026ge;\u0026thinsp;50% vs. \u0026lt;50%), a condition strongly linked to downstream cardiovascular diseases across AZ-TR and DX-EV with available coronary angiography (CAG) data (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ea). Across 16 cardiovascular-related biomarkers, facial predictions were consistent with laboratory measurements (\u003cb\u003eExtended Data\u003c/b\u003e Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e \u003cb\u003eand Supplementary Table\u0026nbsp;9\u003c/b\u003e), with particularly similar patterns for six key biomarkers (HDL-C, LDL-C, TG, TC, HbA1C, hsTnI) (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eb). For example, HDL-C, a biomarker of established clinical relevance, facial predictions (OR/SD 0.705, 95% CI: 0.579\u0026ndash;0.858, \u003cb\u003eP\u003c/b\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001) showed associations consistent with those obtained from laboratory measurements (OR/SD 0.856, 95% CI: 0.749\u0026ndash;0.978, \u003cb\u003eP\u003c/b\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001), both pointing in the same protective direction. In contrast, for LDL-C, since most participants in the studied cohorts were cardiology inpatients receiving statin-based lipid-lowering therapy in AZ-TR and DX-EV cohort, leading to a negative association for laboratory-measured LDL-C (OR/SD 0.726, 95% CI: 0.636\u0026ndash;0.829, \u003cb\u003eP\u003c/b\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001), whereas face-predicted LDL-C preserved the expected positive association (OR/SD 1.445, 95% CI: 1.252\u0026ndash;1.668, \u003cb\u003eP\u003c/b\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001), indicating that FaceFound captures underlying disease-related features independent of drug effects.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eNext, to evaluate the risk evaluation potential of face-predicted biomarkers, we developed and compared three coronary stenosis risk prediction models using different combinations of FaceFound predicted and laboratory-measured biomarkers. The lab-measured basic model (Model 1) incorporated demographic and anthropometric variables (age, sex, height, weight) with 4 basic laboratory-measured lipid profiles (LDL-C, HDL-C, TC, TG). The face-predicted basic model (Model 2) replaced 4 basic laboratory-measured lipids with face-predicted counterparts. The face-predicted advanced model (Model 3) expanded the FaceFound predicted factors to 16 face-predicted cardiovascular biomarkers (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ec). The model was trained in the AZ-TR, and we evaluated the performance of classification by area under curve (AUC) in the external DX-EV cohort (\u003cb\u003eSupplementary Table\u0026nbsp;10\u003c/b\u003e). In the external DX-EV cohort, face-predicted basic model (Model 2) and face-predicted advanced model (Model 3) displayed comparable and higher performance (AUC: 0.649, 95% CI: 0.609\u0026ndash;0.690 vs. 0.657, 95% CI: 0.618\u0026ndash;0.696) compared with lab-measured basic model (Model 1) (0.614, 95% CI: 0.570\u0026ndash;0.656) (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ed).\u003c/p\u003e \u003cp\u003eWe next assessed the risk stratification values of three models in the external independent DX-EV cohort. In the low-risk group (the bottom 20% predicted risk), stenosis prevalence was 42.9%, 38.6%, and 37.1% for lab-measured basic model, face-predicted basic model, and face-predicted advanced model respectively; for the high-risk group (the top 20% risk), the rates were 71.6%, 66.7%, and 73.0%. These results indicate that FaceFound predicted biomarkers can achieve comparable risk stratification values with laboratory measurements (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ee).\u003c/p\u003e \u003cp\u003eFinally, feature importance analysis for FaceFound predicted biomarkers compared to lab-measured biomarkers in AZ-TR showed that the contributions of face-predicted biomarkers were consistent with laboratory-measured counterparts, with an intraclass correlation coefficient (ICC) of 0.86 (95% CI, 0.68\u0026ndash;0.94) (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ef and \u003cb\u003eSupplementary Table\u0026nbsp;11\u003c/b\u003e). Among them, HDL-C, HbA1C, and TC were the most influential predictors in both modalities, highlighting their robust association with stenosis risk.\u003c/p\u003e \u003cp\u003eTogether, these results demonstrate that FaceFound predicted biomarkers preserve disease-associated information, perform comparably to laboratory measurements for risk evaluation, and support the development of scalable, clinically meaningful risk assessment strategies.\u003c/p\u003e\n\u003ch3\u003eStability and label efficiency of FaceFound\u003c/h3\u003e\n\u003cp\u003eA critical requirement for clinical deployment is model stability, ensuring that predictions remain consistent across multiple measurements. To evaluate this, we performed a test-retest analysis from internal and external cohorts, using facial photographs repeatedly acquired from the same participant with minor variations in capture angle. FaceFound displayed excellent stability, achieving higher intraclass correlation coefficients (ICCs) compared to SwinLarge and ResNet18 across all cohorts. In the internal AZ-TR cohort, FaceFound achieved an ICC of 0.879 (IQR: 0.838\u0026ndash;0.931); in the external datasets, FaceFound achieved higher ICC compared with other models, with 0.914 (IQR: 0.883\u0026ndash;0.947) in AZ-EV1, 0.913 (IQR: 0.881\u0026ndash;0.947) in AZ-EV2, and 0.971 (IQR: 0.958\u0026ndash;0.980) in DX-EV \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ea and \u003cb\u003eSupplementary Table\u0026nbsp;12\u003c/b\u003e). Across physiological systems, FaceFound retained stable performance, with median ICCs 0.911 (IQR: 0.903\u0026ndash;0.951) in internal and external cohorts with low variance (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eb; \u003cb\u003eSupplementary Table\u0026nbsp;13\u003c/b\u003e). For example, HDL-C showed high stability in AZ-TR (ICC of 0.905, 95% CI: 0.898\u0026ndash;0.912), AZ-EV1 (ICC of 0.945, 95% CI: 0.944\u0026ndash;0.946), AZ-EV2 (ICC of 0.946, 95% CI: 0.944\u0026ndash;0.947), and DX-EV (ICC of 0.977, 95% CI: 0.976\u0026ndash;0.978), as shown in the Bland-Altman plots (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ec-f \u003cb\u003eand Supplementary Table\u0026nbsp;12\u003c/b\u003e). These findings confirm that FaceFound yields stable, reproducible biomarker predictions, supporting its robustness estimates.\u003c/p\u003e \u003cp\u003eLabel efficiency refers to the amount of training data and labels required to achieve a target performance level for a given downstream task, which indicates the annotation workload for medical experts. We investigated the label efficiency of FaceFound by systematic evaluation across training samples from 400 to 3200 samples, using 62 systematic biomarkers. Across all sample sizes, FaceFound consistently demonstrated superior label efficiency compared with SwinLarge and ResNet18 in all internal and external cohorts (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eg and \u003cb\u003eSupplementary Table\u0026nbsp;14\u003c/b\u003e). Notably, with as few as 400 samples, FaceFound (median R\u0026sup2; = 0.016, IQR: 0.046\u0026ndash;0.005) achieved substantially higher predictive performance than SwinLarge (median R\u0026sup2; = 0.007, IQR: 0.001\u0026ndash;0.027) and ResNet18 (median R\u0026sup2; = 0.003, IQR: 0.001\u0026ndash;0.017). To quantify data efficiency, we further estimated the proportion of training data required by FaceFound to match the performance of SwinLarge or ResNet18. For eGFR (CKD-EPI) and LYM, FaceFound reached equivalent accuracy using only 10% of labelled data. Superior efficiency was also observed for HbA1c (20%), BMI (28%), Hb (50%) and HDL-C (60%) (\u003cb\u003eExtended Data\u003c/b\u003e Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThese results demonstrate that FaceFound not only achieves stable and reproducible predictions but also confers strong label efficiency, a property critical for developing scalable prediction in new domains or for novel biomarkers where annotated data are scarce.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eModel interpretation\u003c/h2\u003e \u003cp\u003eTo gain insights into the inner-workings of FaceFound leading to its superior performance and label efficiency in biomarkers prediction, we performed qualitative analyses of the pretext task used for self-supervised pretraining and biomarker-specific and stage-specific decisions of FaceFound.\u003c/p\u003e \u003cp\u003eTo evaluate the SSL pretext task, we curated internet-sourced facial images covering both Asian and European populations, displaying diverse abnormalities, including peripheral facial paralysis, renal disease\u0026ndash;associated edema, and Down syndrome, ensuring these facial images were not included in the pre-training stage. We found that FaceFound was able to reconstruct pathological facial features with 75% of the image masked, across different types of facial deformities or abnormal traits, such as periorbital swelling, malocclusion, or macular lesions (\u003cb\u003eExtended Data Fig.\u0026nbsp;7a\u003c/b\u003e). This demonstrates that FaceFound has learned to identify and infer the representation of health-related areas by means of SSL, which contributes to performance and label efficiency in biomarker predictions. Beyond reconstruction-based interpretation, we applied Grad-CAM to the Swin-Large backbone of FaceFound to visualize regions contributing to biomarker-specific predictions (\u003cb\u003eExtended Data Fig.\u0026nbsp;7b\u003c/b\u003e). For eGFR (CKD-EPI), early stages captured global facial context, intermediate stages emphasized the eyes and cheeks, and the final stage focused on the periocular region. For Hb, later stages integrated signals from the entire face in addition to the eyes. For HDL-C, salient contributions were largely concentrated around the periocular region. These findings indicate that FaceFound relies on stage-dependent and physiologically plausible facial features when predicting diverse systemic biomarkers.\u003c/p\u003e \u003cp\u003eThese visualizations demonstrate that FaceFound learns interpretable and biologically meaningful facial representations, bridging facial phenotypes with systemic health indicators and supporting its potential for scalable biomarker prediction in diverse populations.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eReal-world deployment of FaceFound-based biomarker prediction via smartphone\u003c/h3\u003e\n\u003cp\u003eTo assess the translational potential of FaceFound, we developed a smartphone-based application integrating with FaceFound (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003ea). 62 systematic biomarkers can be predicted using standardized frontal facial images by smartphone under a white background and indoor light. With a mobile-based application, users can check their face-predicted biomarkers spanning different health systems with an interpretable report (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eb).\u003c/p\u003e \u003cp\u003eWe piloted the application in the cardiology department of Beijing Anzhen Hospital, where frontal images were taken by a physician who was blinded to the study design, using the front-facing camera of mobile phone as the external validation cohort (median age 60.0, IQR, 55.0\u0026ndash;67.0, 26.4% female, N\u0026thinsp;=\u0026thinsp;110, APP-EV) (\u003cb\u003eMethods for details and Supplementary Table\u0026nbsp;15\u003c/b\u003e). For each individual, the medical records of 58 biomarkers were collected. Overall, after a systematic evaluation, the median R\u0026sup2; of FaceFound app was consistent with counterparts evaluated in external cohorts (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003ec \u003cb\u003eand Supplementary Table\u0026nbsp;16\u003c/b\u003e), with consistency assessed by ICCs in the external cohorts AZ-TR, AZ-EV1, AZ-EV2, AZTZ-EV, and DX-EV, with a median ICC of 0.910 (IQR, 0.900\u0026ndash;0.940). (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003ed, \u003cb\u003eSupplementary Table\u0026nbsp;17\u003c/b\u003e). These results demonstrate that biomarker prediction from simple selfies is feasible in real-world clinical environments. This deployment highlights the potential of FaceFound for rapid, low-cost, and scalable integration into pre-screening workflows and digital health monitoring platforms.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we demonstrate that a facial foundation model FaceFound with progressive general-to-clinical pretraining strategy have good prediction power for clinical biomarker prediction across eight physiological systems, utilizing over 70,000 facial photographs from approximately 10,000 participants across five diverse Chinese cohorts with stability, label efficiency, higher performance for most biomarkers compared to genetics and comparable predictive performance with facial images taken with a smartphone. To our knowledge, this provides the first broad evidence that facial photographs can be leveraged for predicting clinical biomarkers. Previous studies of facial images have focused mainly on non-medical domains such as person identity recognition and emotion analysis\u003csup\u003e\u003cspan additionalcitationids=\"CR21 CR22\" citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e–\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e, with only limited work addressing health-related outcomes, including aging\u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e, syndromic facial dysmorphisms\u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e, and coronary artery disease\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e,\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e. Our findings substantially extend this landscape by showing that facial photographs carry strong predictive power for biomarkers across renal, hematologic, cardiovascular, endocrine, digestive, and immune systems. These results suggest that facial images can be applied to diverse disease risk assessment and health monitoring, and further developed as an easily accessible medical imaging modality to support early screening and risk evaluation, thereby opening new avenues for integrating facial analysis into precision medicine and population health.\u003c/p\u003e \u003cp\u003eAlthough both genetic and environmental factors shape biomarker variation, previous studies have largely emphasized genetic contributions\u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e while neglecting the substantial role of acquired exposures. Facial morphology, as a complex phenotype influenced by genetics and environment\u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e, provides a unique lens to evaluate personal health status, yet its variance explained compared with genetics remains unclear. By analyzing 679 polygenic risk scores (PRS) from 27 phenotypes, with the optimal PRS representing genetic effects, we demonstrate that in 15 phenotypes, facial signals surpassed genetic prediction. Notably, facial features more effectively captured estimated glomerular filtration rate (eGFR (CKD-EPI)), underscoring the predominant facial features for kidney function, and for the first time we identify stronger facial associations with high-density lipoprotein cholesterol (HDLC) and lipoprotein(a) [Lp(a)] than by genetics, which helps explain the predictive power of facial features for coronary artery disease. Collectively, these findings establish the human face as an accessible and dynamic source of biomarker information, more sensitive to environmental influences than static genetic data, and highlight the potential of combining facial and genetic information together to improve biomarker prediction and to reveal novel interactions between genes, facial development, and environmental exposures.\u003c/p\u003e \u003cp\u003eOur study further underscores the clinical value of face-predicted biomarkers. As central indicators of physiological processes, biomarkers fluctuate with disease progression and therapeutic interventions\u003csup\u003e\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e. We found that facially predicted biomarkers preserved these dynamic variations between health and disease states with high consistency by evaluation of stenosis with 16 cardiovascular biomarkers. Interestingly, we observed participants receiving lipid-lowering therapy exhibited reduced LDL-C levels while facially predicted LDL-C values remained elevated among individuals with coronary stenosis in the external independent cohort, possibly due to the improvement of FaceFound in capturing reversion correlation after medical treatment. Furthermore, incorporation of facially predicted cardiovascular biomarkers improved risk evaluation accuracy for stenosis beyond traditional risk factors\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. These findings suggest that face-predicted biomarkers could be integrated into established clinical risk models to enhance disease assessment and management.\u003c/p\u003e \u003cp\u003eA key strength of our work lies in the high performance and robustness of the proposed FaceFound model. FaceFound achieved excellent test-retest reliability across the internal and external cohorts, substantially outperforming existing models. This stability is critical for clinical deployment, ensuring reproducibility across repeated captures and variable imaging conditions. Moreover, FaceFound exhibited superior label efficiency, maintaining high predictive accuracy even when trained on as few as 400 samples. Such efficiency underscores the potential of FaceFound to accelerate biomarker discovery and disease modeling in rare conditions, emerging infectious diseases, or other settings where labeled data are scarce. By leveraging large-scale pretraining, our framework confers both robustness and adaptability, two properties essential for real-world clinical translation.\u003c/p\u003e \u003cp\u003eBeyond methodological advances, we demonstrate the feasibility of real-world deployment by integrating FaceFound into a smartphone-based application. Remarkably, predictions derived from mobile-captured images achieved performance comparable to research-grade inputs, highlighting the translational strength of our approach. This capability enables novel scenarios for health monitoring, including at-home biomarker screening, community-level digital health interventions, and rapid triage in emergency or resource-limited settings. Furthermore, because our framework relies solely on facial photographs, it can be seamlessly embedded into widely adopted technologies such as facial recognition, thereby providing real-time, unobtrusive health surveillance at a population scale.\u003c/p\u003e \u003cp\u003eDespite these advances, several limitations warrant consideration. First, data heterogeneity arising from differences in sampled populations, data-acquisition devices and collection protocols can impair model transferability. Expanding data collection to more diverse clinical populations or conducting center-specific fine-tuning will be crucial to improving model generalizability. Second, the current study was limited to Chinese populations; validation in other ancestries is needed to ensure global applicability. Finally, while stenosis was used as the primary clinical endpoint, future work should assess the predictive power of facial biomarkers in renal, endocrine, and hematologic diseases, where broader validation is essential. Nevertheless, these limitations represent opportunities for future research, and our results provide strong proof-of-concept evidence for the clinical value of facial biomarker prediction.\u003c/p\u003e \u003cp\u003eCollectively, our study establishes a new paradigm for precision health: that a single facial photograph can serve as a non-invasive, multidimensional window into systemic physiology. By enabling scalable biomarker prediction, robust disease risk evaluation, and seamless smartphone deployment, our work paves the way toward automated, real-time, and population-wide health monitoring. We envision that facial foundation models, integrated into ubiquitous digital platforms, could transform preventive medicine and democratize access to personalized healthcare on a global scale.\u003c/p\u003e "},{"header":"Methods","content":"\u003ch2\u003eFacial image acquisition\u003c/h2\u003e\u003cp\u003eTo develop FaceFound, we curated a large-scale dataset comprising both public and clinical facial image datasets. The public datasets included ImageNet-22k\u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e and VGGFace2-HQ\u003csup\u003e28,29\u003c/sup\u003e. The ImageNet-22k dataset, containing 14,197,122 real-world images, was preprocessed according to standard protocols\u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e. The VGGFace2HQ dataset (1,160,250 public facial images) was aligned and normalized as standardized procedures\u003csup\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e,\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eThe clinical facial datasets for pretraining were collected from patients at Beijing Anzhen Hospital between 2021 and Q2 2023 (AZ-TR unfiltered), all facial photographs were obtained by a physician who was blinded to the study design. A digital camera with a resolution of at least 20 megapixels was used under P mode, enabled with an ISO sensitivity of 1600 and burst mode. Images were captured in a quiet examination room with a plain white background and direct lighting to ensure privacy and consistency. Participants were instructed to keep their eyes open, maintain a neutral facial expression, and avoid accessories or hair covering facial features. Each participant was photographed in frontal, left and right 60° lateral, and downward-facing positions (exposing the forehead), with 3–5 photographs per angle. These multiple angles were collected during pretraining to comprehensively capture facial information. In the fine-tuning stage for predicting 62 biomarkers, we used only a single frontal photograph, which achieved comparable predictive performance and offered greater practicality and translational value for downstream clinical applications.\u003c/p\u003e\u003cp\u003eTo fine-tune FaceFound, for predicting 62 biomarkers routinely tested in a healthcare system across eight physiological systems, we collected 4,698 individuals with linked hospital records from AZ-TR-unfiltered dataset (N = 4,698, median age 61.0 years, IQR 54.0–67.0; 26.4% female). All patients in the internal cohort of AZ-TR were divided randomly into training, validation and testing datasets in ratio of 70%:20%:10%.\u003c/p\u003e\u003cp\u003eFor external validation, we collected five independent cohorts with the same protocols, in which image acquisition was performed by different physicians blinded to the study content, and only frontal facial photographs were obtained. In the second half of 2023, we enrolled 1,498 participants at Anzhen Hospital (AZ-EV1; median age 61.0 years, IQR 54.0–67.0; 24.4% female). In 2024, an additional 2,247 participants were recruited at the same hospital (AZ-EV2; median age 60.0 years, IQR 53.0–66.0; 22.2% female). Between 2021 and 2022, 900 participants were recruited at Beijing Daxing Hospital (DX-EV; median age 62.0 years, IQR 56.0–69.0; 49.9% female). In 2025, we enrolled 206 participants from the Tongzhou branch of Anzhen Hospital (AZTZ-EV; median age 61.0 years, IQR 56.0–68.0; 39.8% female), with one facial image per individual. Finally, the AZ-GEV cohort comprised 111 participants (median age 59.0 years, IQR 52.0–65.5; 30.6% female), with one facial image per individual.\u003c/p\u003e\u003ch2\u003eBiomarkers variables\u003c/h2\u003e\u003cp\u003eFor the internal cohort, a total of 62 biomarkers were collected in the AZ-TR cohort, with the available sample size varying across biomarkers as determined by treating clinicians and standard institutional procedures. For blood sampling, venous blood was obtained in the morning after an overnight fast. Samples were stored at room temperature protected from light and delivered to the hospital laboratory within 30 minutes for routine assays. For the external cohorts, identical procedures were followed. For AZ-TR and DX-EV, we also collected coronary angiography (CAG) data to calculate stenosis, which further aided in developing a disease risk evaluation model. For AZ-GEV, we further collected peripheral blood samples to genotype for calculating genetics scores of biomarkers.\u003c/p\u003e\u003cp\u003e \u003cstrong\u003eInformed consent\u003c/strong\u003e \u003c/p\u003e\u003cp\u003ewas obtained from all participants. The study was conducted in accordance with the principles of the Declaration of Helsinki and was approved by the Institutional Review Board (Ethical Approval No. 2022211X).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003ch2\u003eFacial image preprocessing\u003c/h2\u003e\u003cp\u003eFacial photographs were first linked to the corresponding medical records by extracting patient identifiers through Optical Character Recognition (OCR) using “paddleOCR” (v3.0.1) \u003csup\u003e34\u003c/sup\u003e. When OCR failed, manual correction was performed to ensure accurate matching. Head pose was estimated using 6DRepNet\u003csup\u003e\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u003c/sup\u003e, and only frontal images were retained for the prediction of 62 biomarkers, while images from all angles were used during pretraining. The final dataset included only images successfully matched to hospital records based on participant name, patient ID, and timestamp. Blurry images were excluded. Remaining images were processed with DeepFace\u003csup\u003e\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u003c/sup\u003e for face detection, alignment, and cropping, and were subsequently standardized to 256 × 256 pixels for model training and evaluation.\u003c/p\u003e\u003ch2\u003eDevelopment of FaceFound model\u003c/h2\u003e\u003cp\u003eWe adopted the Swin Transformer backbone\u003csup\u003e\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u003c/sup\u003e for FaceFound, specifically selecting the Swin-Large variant for its scalability across input resolutions during pre-training and its efficiency in extracting features from high-resolution facial images. Swin-Large is organized into four hierarchical stages with 2, 2, 18, and 2 Swin Transformer Blocks in the respective stages. Input facial photographs were partitioned into non-overlapping 4×4 patches (initial feature size 4×4×3), which are linearly projected to 192-dimensional embeddings and propagated through successive transformer stages. Both the embedding dimensionality and the number of attention heads increased with depth. Within each stage, multi-head self-attention operates in local windows, alternating between standard window-based multi-head self-attention (W-MSA) and shifted-window multi-head self-attention (SW-MSA) to enhance cross-window information flow.\u003c/p\u003e\u003cp\u003eOn this backbone, we developed a facial foundation model called FaceFound using a General-to-Clinical self-supervised pre-training strategy based on the Uniform Masking Masked Autoencoder (UM-MAE) framework\u003csup\u003e\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u003c/sup\u003e. UM-MAE, which is well suited to pyramid-based vision transformers such as Swin, reconstructs uniformly sampled masked patches with secondary masking to improve locality awareness and feature robustness. Compared with alternatives such as SimMIM\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e, UM-MAE offers greater computational efficiency—reducing pre-training time and GPU memory consumption—while maintaining strong fine-tuning performance.\u003c/p\u003e\u003cp\u003ePre-training proceeded in three sequential stages, progressively specializing the representations from generic vision to clinically relevant facial features.\u003c/p\u003e\u003cp\u003e \u003cb\u003eStage 1 (general visual pre-training).\u003c/b\u003e FaceFound was initialized with UM-MAE weights trained on ImageNet-22K, using 256×256 inputs with a 4×4 patch size and an 8-pixel window, to establish broad visual representations.\u003c/p\u003e\u003cp\u003e \u003cb\u003eStage 2 (facial feature pre-training)\u003c/b\u003e. FaceFound was adapted to human facial characteristics using VGGFace2-HQ, a high-resolution dataset (512×512 pixels) comprising 9,131 identities derived from VGGFace2 via super-resolution. Inputs are 512×512 with a 4×4 patch size and a 16-pixel window, refining sensitivity to facial morphology and fine-grained details.\u003c/p\u003e\u003cp\u003e \u003cb\u003eStage 3 (target-population pre-training).\u003c/b\u003e Finally, FaceFound was further specialized on a proprietary clinical dataset of 88,409 facial images from patients at Beijing Anzhen Hospital, using 256×256 inputs, to capture population-specific characteristics relevant to downstream clinical biomarker prediction.\u003c/p\u003e\u003cp\u003eThis hierarchical curriculum—from natural images (ImageNet-22K), to facial imagery (VGGFace2-HQ), to clinical patient photographs—enabled a gradual transition from general visual features to task-specific clinical representations. The consistent use of UM-MAE across stages yielded an efficient self-supervised pipeline and a robust facial foundation model tailored for predicting systemic biomarkers from facial photographs.\u003c/p\u003e\u003ch2\u003eFine-tuning and evaluation for downstream tasks\u003c/h2\u003e\u003cp\u003eFor downstream task adaptation, we used only the Swin-L encoder of the foundation model as the backbone of FaceFound, discarding the decoder of UM-MAE. The backbone extracted high-level facial representations, which were passed into a multilayer perceptron (MLP) to generate task-specific predictions. Depending on the downstream task, the outputs were defined as different biomarker levels.\u003c/p\u003e\u003cp\u003eWe conducted supervised fine-tuning on all parameters of the Swin-L backbone and the MLP for 50 epochs with a batch size of 32, using AdamW optimizer. The first 10 epochs employed a linear warm-up schedule, gradually increasing the learning rate from 0 to 5 × 10\u003csup\u003e− 3\u003c/sup\u003e, followed by a cosine annealing decay schedule that reduced the learning rate from 5 × 10\u003csup\u003e− 3\u003c/sup\u003e to 1 × 10\u003csup\u003e− 6\u003c/sup\u003e over the remaining 40 epochs. We stopped the training early when the validation loss does not decrease for 16 consecutive epochs. After each epoch, models were evaluated on the validation set, and the checkpoint with the lowest validation loss was retained for subsequent testing on both internal and external cohorts. Model performance was evaluated using the coefficient of determination (R\u003csup\u003e2\u003c/sup\u003e) as the primary metric for regression tasks.\u003c/p\u003e\u003ch2\u003eOther deep learning model implementations\u003c/h2\u003e\u003cp\u003eWe implemented our proposed General-to-Clinical pretraining strategy, leveraging the UM-MAE self-supervised pretraining framework, to pretrain the Swin-L backbone of FaceFound. We then compare this General-to-Clinical approach with a baseline strategy that involves pretraining the model solely on ImageNet-22K. To enable a fair comparison, we also pretrain a Swin-L backbone using UM-MAE on ImageNet-22K as a pretraining strategy comparison. For architectural comparison, we incorporate the well-established ResNet18\u003csup\u003e39\u003c/sup\u003e model, which we pretrain on ImageNet-22K.\u003c/p\u003e\u003ch2\u003eGenotyping and construction of polygenic risk scores\u003c/h2\u003e\u003cp\u003ePeripheral blood samples from AZ-GEV participants were genotyped using the Beadchip Array Asian Screening Array. Standard quality control was applied, including filters for sample and variant call rates, Hardy–Weinberg equilibrium, and minor allele frequency. Genotype imputation was performed on the Michigan Imputation Server\u003csup\u003e\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u003c/sup\u003e with the Haplotype Reference Consortium panel (HRC r1.1 2016), using Eagle v2.4 for phasing and Minimac4 for imputation. Variants with imputation INFO ≤ 0.8 or R² ≤0.8 were excluded, leaving 5,016,187 high-quality variants for PRS calculation.\u003c/p\u003e\u003cp\u003ePolygenic risk scores (PRS) were constructed for each individual using summary statistics from previously published 679 genome-wide association studies (GWAS) relevant to 27 biomarker traits from PGS Catalog\u003csup\u003e\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e\u003c/sup\u003e. Scores were calculated as the weighted sum of risk alleles carried by each individual, with weights corresponding to the effect sizes reported in the GWAS data by PLINK2.\u003c/p\u003e\u003ch2\u003eDevelopment of risk evaluation models\u003c/h2\u003e\u003cp\u003eWe constructed three stenosis-classification models to evaluate the risk evaluation utility of face-predicted biomarkers compared with conventional laboratory measurements.\u003c/p\u003e\u003cp\u003eModel 1 incorporated demographic and anthropometric variables (age, sex, height, weight) with laboratory-measured lipid profiles (LDL-C, HDL-C, TC, TG). Model 2 replaced laboratory-measured lipids with their face-predicted counterparts. Model 3 expanded the feature set to include 16 face-predicted cardiovascular biomarkers, thereby evaluating their incremental contribution in the absence of corresponding laboratory measurements in the DX-EV cohort (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ec).\u003c/p\u003e\u003cp\u003eAll models were trained in the AZ-TR cohort and externally validated in the DX-EV cohort with a logistic model, where stenosis severity (≥ 50% vs. \u0026lt; 50%) was available as the clinical endpoint. Multivariable logistic regression was used for model development. To quantify the contribution of individual biomarkers, we applied the “relaimpo” package\u003csup\u003e\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e\u003c/sup\u003e (v2.2.7) with the “lmg” method to calculate the relative importance of 16 face-predicted versus laboratory-measured biomarkers in predicting stenosis.\u003c/p\u003e\u003ch2\u003eTest-retest analysis\u003c/h2\u003e\u003cp\u003eTo assess the stability and reproducibility of our facial foundation model, we conducted a test–retest analysis across both internal and external cohorts. For each individual with multiple facial photographs collected, we randomly sampled two images taken on different occasions to form a test–retest pair. For each biomarker, the FaceFound model and other baseline models were applied independently to both images within each pair, and repeated predictions were obtained. Test–retest reliability was quantified using the intraclass correlation coefficient (ICC), calculated separately for repeated predicted biomarker values of the prediction model, and Cohen’s was calculated to evaluate retest reliability using the “pingouin” package (v0.5.5)\u003c/p\u003e\u003ch2\u003eLabel efficiency\u003c/h2\u003e\u003cp\u003eLabel efficiency pertains to the quantity of training data and annotations needed to attain a specified performance level for supervised fine-tuning in a particular downstream task, reflecting the annotation effort required from medical professionals\u003csup\u003e\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e\u003c/sup\u003e. To assess the label efficiency of FaceFound relative to baseline architectures, we systematically varied the amount of supervised fine-tuning samples in the AZ-TR cohort. Training subsets were randomly down sampled to 400, 800, and 1600 samples without replacement, while using the validation and internal testing cohorts in the same manner as the former supervised fine-tuning. Six representative blood biomarkers were selected to cover diverse physiological systems, and models were fine-tuned following the predefined supervised fine-tuning strategy.\u003c/p\u003e\u003cp\u003eIn total, 558 models were trained, including 186 FaceFound models (62 biomarkers × 3 data regimes) and 372 baseline models (Swin-Large and ResNet-18, each trained under the same conditions). Model performance was evaluated on both the internal and three independent external cohorts (AZ-EV1, AZ-EV2, and DX-EV). This design enabled a direct comparison of predictive performance under varying data regimes, thereby quantifying the label efficiency of FaceFound versus conventional models.\u003c/p\u003e\u003ch2\u003eExplanations for fine-tuned models\u003c/h2\u003e\u003cp\u003eWe applied Gradient-weighted Class Activation Mapping (Grad-CAM) to visualize the predictive mechanisms of fine-tuned models for biomarker estimation. Grad-CAM is a gradient-based interpretability method that generates class-discriminative localization maps by weighting activations at a selected layer with the gradients of a target output, thereby highlighting input regions most influential for prediction. For each fine-tuned model, the first LayerNorm layer in the final block of each stage (i.e., the 2nd, 4th, 22nd, and 24th blocks) was designated as the target layer and the single continuous output was treated as the target “class.”\u003c/p\u003e\u003ch2\u003eMobile application deployment and validation\u003c/h2\u003e\u003cp\u003eWe developed a WeChat-based mobile application using native WeChat components and the Vant Weapp framework to support user interaction and facial image collection. The backend was implemented in Golang with MySQL and Redis for data management and caching. FaceFound was deployed on an Ubuntu 22.04 server with GPU, where the backend was compiled into a standalone binary and managed via Systemd. Secure communication was ensured through HTTPS encryption, JWT-based authentication, and SSL-encrypted database connections.\u003c/p\u003e\u003cp\u003eTo evaluate the feasibility of applying the model in real-world mobile settings, an independent validation cohort was established. Facial images were collected by a physician who was blinded to the study design (N = 110), using the front-facing camera of an iPhone 15 Pro Max. Other than a plain white background to ensure basic consistency, no restrictions were imposed on the shooting environment. One frontal image was collected per participant, without constraints on facial expression. This cohort was used specifically to validate model performance on smartphone-acquired images, thereby assessing its practicality and translational potential for broader deployment in clinical and community-based contexts.\u003c/p\u003e\u003ch2\u003eStatistical analysis\u003c/h2\u003e\u003cp\u003eBaseline characteristics were summarized as mean (standard deviation, SD) for normally distributed continuous variables, median (interquartile range, IQR) for non-normally distributed continuous variables, and frequency (percentage) for categorical variables. To evaluate the performance of facial prediction, the coefficient of determination (R2) was calculated from regression models comparing face-predicted values with corresponding laboratory measurements. P values were adjusted for multiple testing using the Benjamini-Hochberg false discovery rate (FDR-BH) method. Biomarkers with adjusted P ≥ 0.05 were considered to have no significant predictive ability. For association analysis of facial predicted values and laboratory measurement values on stenosis, we used logistic regression with covariates of age, sex, height, weight, and BMI. A two-sided P ≥ 0.05 indicates statistical significance. For risk evaluation models, we assessed performance using the area under the receiver operating characteristic curve (AUC) and the area under the precision-recall curve (APR), as well as accuracy, sensitivity, and specificity with optimal cut-off values were determined by the maximum Youden index. 95% CI for performance metrics was estimated using bootstrap resampling with 1,000 iterations. All statistical analyses were conducted in “Python” (v3.10.14) and “statsmodels” (v0.14.2). Figures were generated with the “matplotlib” package (v3.10.0).\u003c/p\u003e\u003ch2\u003eData availability\u003c/h2\u003e\u003cp\u003eThe predicted biomarker scores, polygenic risk scores (PRS), and corresponding labels that do not contain any personally identifiable information are available upon reasonable request to the corresponding author. Access to raw facial images is restricted to protect participant privacy and can be granted only to qualified investigators with appropriate institutional and ethical approvals and a justified scientific purpose. The PGS variant-level weights used is this study are available for download through the PGS Catalog (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.pgscatalog.org/\u003c/span\u003e\u003cspan address=\"https://www.pgscatalog.org/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e).\u003c/span\u003e\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eData availability\u003c/h2\u003e\n\u003cp\u003eThe predicted biomarker scores, polygenic risk scores (PRS), and corresponding labels that do not contain any personally identifiable information are available upon reasonable request to the corresponding author. Access to raw facial images is restricted to protect participant privacy and can be granted only to qualified investigators with appropriate institutional and ethical approvals and a justified scientific purpose. The PGS variant-level weights used is this study are available for download through the PGS Catalog (\u003cu\u003ehttps://www.pgscatalog.org/\u003c/u\u003e).\u003c/p\u003e\n\u003ch2\u003eCode availability\u003c/h2\u003e\n\u003cp\u003eAll software used in this study is publicly available. The custom code developed and used to train and fine-tune the \u003cstrong\u003eFaceFound\u003c/strong\u003e model for biomarker prediction, as well as the analysis pipelines used in this paper, is available at \u003cu\u003ehttps://github.com/1511878618/FaceFound\u003c/u\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eMayeux R, Biomarkers (2004) Potential uses and limitations. Neurotherapeutics 1:182\u0026ndash;188\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang WR, Parikh CR (2019) Biomarkers of Acute and Chronic Kidney Disease. Annu Rev Physiol 81:309\u0026ndash;333\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHippisley-Cox J et al (2024) Development and validation of a new algorithm for improved cardiovascular risk prediction. Nat Med 30:1440\u0026ndash;1447\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCummings JL et al (2025) Biomarker-guided decision making in clinical drug development for neurodegenerative disorders. Nat Rev Drug Discov 24:589\u0026ndash;609\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLow-cost (2022) point-of-care biomarker quantification. Curr Opin Biotechnol 76:102738\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMacchia E et al (2024) Point-Of-Care Ultra-Portable Single-Molecule Bioassays for One-Health. Adv Mater 36:2309705\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRim TH et al (2020) Prediction of systemic biomarkers from retinal photographs: development and validation of deep-learning algorithms. Lancet Digit Health 2:e526\u0026ndash;e536\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePoplin R et al (2018) Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. Nat Biomed Eng 2:158\u0026ndash;164\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSingal A et al (2025) Part I. Dermatologic Manifestations in Patients with Kidney Disease. J Am Acad Dermatol S0190\u0026ndash;9622(25):02360\u0026ndash;02366. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.jaad.2025.06.035\u003c/span\u003e\u003cspan address=\"10.1016/j.jaad.2025.06.035\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNixon-Hill M, Mookerjee RP, Leung TS (2023) Assessment of bilirubin levels in patients with cirrhosis via forehead, sclera and lower eyelid smartphone images. PLOS Digit Health 2:e0000357\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCao K et al (2023) Large-scale pancreatic cancer detection via non-contrast CT and deep learning. Nat Med 29:3033\u0026ndash;3043\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHe K et al (2022) IEEE, New Orleans, LA, USA,. Masked Autoencoders Are Scalable Vision Learners. in \u003cem\u003e2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)\u003c/em\u003e 15979\u0026ndash;15988 \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/CVPR52688.2022.01553\u003c/span\u003e\u003cspan address=\"10.1109/CVPR52688.2022.01553\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAzad B et al (2023) Foundational Models in Medical Imaging: A Comprehensive Survey and Future Vision. Preprint at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://arxiv.org/abs/2310.18689\u003c/span\u003e\u003cspan address=\"http://arxiv.org/abs/2310.18689\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAvram O et al (2024) Accurate prediction of disease-risk factors from volumetric medical scans by a deep vision model pre-trained with 2D scans. Nat Biomed Eng. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41551-024-01257-9\u003c/span\u003e\u003cspan address=\"10.1038/s41551-024-01257-9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXie Z et al (2022) SimMIM: a Simple Framework for Masked Image Modeling. in. \u003cem\u003eIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)\u003c/em\u003e 9643\u0026ndash;9653 (IEEE, New Orleans, LA, USA, 2022). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/CVPR52688.2022.00943\u003c/span\u003e\u003cspan address=\"10.1109/CVPR52688.2022.00943\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhou Y et al (2023) A foundation model for generalizable disease detection from retinal images. Nature 1\u0026ndash;8. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41586-023-06555-x\u003c/span\u003e\u003cspan address=\"10.1038/s41586-023-06555-x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLu MY et al (2024) A visual-language foundation model for computational pathology. Nat Med 30:863\u0026ndash;874\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChristensen M, Vukadinovic M, Yuan N, Ouyang D (2024) Vision\u0026ndash;language foundation model for echocardiogram interpretation. Nat Med 30:1481\u0026ndash;1488\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXiang J et al (2025) A vision\u0026ndash;language foundation model for precision oncology. Nature 638:769\u0026ndash;778\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMa B, Zhang W, Qiu F, Ding Y A Unified Approach to Facial Affect Analysis: the MAE-Face Visual Representation. in (2023) \u003cem\u003eIEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)\u003c/em\u003e 5924\u0026ndash;5933 (IEEE, Vancouver, BC, Canada, 2023). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/CVPRW59228.2023.00630\u003c/span\u003e\u003cspan address=\"10.1109/CVPRW59228.2023.00630\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSerengil S, \u0026Ouml;zpınar AA (2024) Benchmark of Facial Recognition Pipelines and Co-Usability Performances of Modules. Bilişim Teknolojileri Dergisi 17:95\u0026ndash;107\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShahreza HO, George A, Marcel S (2025) Face Reconstruction from Face Embeddings using Adapter to a Face Foundation Model. in 5584\u0026ndash;5593\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eQin L et al (2023) SwinFace: A Multi-task Transformer for Face Recognition, Expression Recognition, Age Estimation and Attribute Estimation. IEEE Trans Circuits Syst Video Technol 1\u0026ndash;1. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/TCSVT.2023.3304724\u003c/span\u003e\u003cspan address=\"10.1109/TCSVT.2023.3304724\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLin S et al (2020) Feasibility of using deep learning to detect coronary artery disease based on facial photo. Eur Heart J 41:4400\u0026ndash;4411\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHsieh T-C et al (2022) GestaltMatcher facilitates rare disease matching using facial phenotype descriptors. Nat Genet 54:349\u0026ndash;357\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhou J et al (2025) Facial Foundational Model Advances Early Warning of Coronary Artery Disease from Live Videos with DigitalShadow. 04.29.25326370 Preprint at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1101/2025.04.29.25326370\u003c/span\u003e\u003cspan address=\"10.1101/2025.04.29.25326370\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDeng J et al (2009) ImageNet: A large-scale hierarchical image database. in. \u003cem\u003eIEEE Conference on Computer Vision and Pattern Recognition\u003c/em\u003e 248\u0026ndash;255 (2009). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/CVPR.2009.5206848\u003c/span\u003e\u003cspan address=\"10.1109/CVPR.2009.5206848\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCao Q, Shen L, Xie W, Parkhi OM, Zisserman A (2018) VGGFace2: A dataset for recognising faces across pose and age. Preprint at. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.1710.08092\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.1710.08092\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen X et al (2024) SimSwap++: Towards Faster and High-Quality Identity Swapping. IEEE Trans Pattern Anal Mach Intell 46:576\u0026ndash;592\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBontempi D et al (2025) FaceAge, a deep learning system to estimate biological age from face photographs to improve prognostication: a model development and validation study. Lancet Digit Health 7:100870\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVerma A et al (2024) Diversity and scale: Genetic architecture of 2068 traits in the VA Million Veteran Program. Science 385:eadj1182\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang M et al (2022) Genetic variants underlying differences in facial morphology in East Asian and European populations. Nat Genet 54:403\u0026ndash;411\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMacDougall DE et al (2025) Lipoprotein(a) and recurrent atherosclerotic cardiovascular events: the US Family Heart Database. Eur Heart J ehaf297. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/eurheartj/ehaf297\u003c/span\u003e\u003cspan address=\"10.1093/eurheartj/ehaf297\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCui C et al (2025) PaddleOCR 3.0 Technical Report. Preprint at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2507.05595\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2507.05595\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHempel T, Abdelrahman AA, Al-Hamadi A (2023) Towards Robust and Unconstrained Full Range of Rotation Head Pose Estimation. Preprint at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2309.07654\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2309.07654\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSerengil SI, Ozpinar A, LightFace: (2020) A Hybrid Deep Face Recognition Framework. in \u003cem\u003e2020 Innovations in Intelligent Systems and Applications Conference (ASYU)\u003c/em\u003e 1\u0026ndash;5 \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/ASYU50717.2020.9259802\u003c/span\u003e\u003cspan address=\"10.1109/ASYU50717.2020.9259802\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu Z et al (2021) Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows. in 10012\u0026ndash;10022\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi X, Wang W, Yang L, Yang J (2022) Uniform Masking: Enabling MAE Pre-training for Pyramid-based Vision Transformers with Locality. Preprint at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2205.10063\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2205.10063\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHe K, Zhang X, Ren S, Sun J (2016) Deep Residual Learning for Image Recognition. in 770\u0026ndash;778\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDas S et al (2016) Next-generation genotype imputation service and methods. Nat Genet 48:1284\u0026ndash;1287\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLambert SA et al (2024) Enhancing the Polygenic Score Catalog with tools for score calculation and ancestry normalization. Nat Genet 56:1989\u0026ndash;1994\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGroemping U (2007) Relative Importance for Linear Regression in R: The Package relaimpo. J Stat Softw 17:1\u0026ndash;27\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJin C, Guo Z, Lin Y, Luo L, Chen H (2025) Label-Efficient Deep Learning in Medical Image Analysis: Challenges and Future Directions. Preprint at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2303.12484\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2303.12484\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-8110055/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8110055/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eWhile most biomarkers currently rely on invasive laboratory testing, which limits large-scale or repeated screening, scalable non-invasive methods could transform population screening, early disease detection and personalized health management. Facial photographs, as a ubiquitous and non-invasive data source, offer such potential but remain underexplored for clinically relevant biomarker and disease risk prediction. Here, we present FaceFound, a facial foundation model trained over 10\u0026nbsp;million images through a progressive general-to-clinical pretraining strategy and evaluated across 62 biomarkers spanning eight physiological systems, many of which are well-established indicators of cardiometabolic, renal, and systemic diseases.. FaceFound consistently outperformed baseline architectures in biomarker prediction, achieving state-of-the-art performance for 45 (73%) biomarkers and demonstrating robust results across internal and four independent external cohorts (N\u0026thinsp;=\u0026thinsp;206-2,247). Notably, FaceFound displayed superior performance to genetic models for the prediction of 14 out of 26 biomarkers with genetic scores available in the PGS catalog, highlighting its complementary value for disease risk assessment beyond inherited genetic susceptibility. Moreover, Face-predicted cardiovascular biomarkers demonstrated strong associations with coronary stenosis, enabling accurate prediction of cardiovascular disease risk and outperforming models based on laboratory-measured biomarkers. FaceFound further exhibited label efficiency, retaining predictive power with as few as 400 training samples, underscoring its value in low-resource settings. Moreover, FaceFound was deployed as a smartphone application, enabling real-time biomarker estimation and individualized disease risk reporting from a single self-captured facial photograph. These findings provide that FaceFound can reproducibly predict multi-system biomarkers and clinically relevant disease risk from facial images with real-world feasibility, establishing a paradigm for population-wide digital screening, early disease risk stratification and personalized risk assessment.\u003c/p\u003e","manuscriptTitle":"A facial foundation model for multi-system biomarker and disease risk prediction with real-world mobile deployment","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-01-21 10:06:26","doi":"10.21203/rs.3.rs-8110055/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"4748329d-1d84-4e11-9172-b6e929d9237b","owner":[],"postedDate":"January 21st, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":60826830,"name":"Health sciences/Health care/Public health/Population screening"},{"id":60826831,"name":"Health sciences/Biomarkers/Predictive markers"},{"id":60826832,"name":"Health sciences/Risk factors"},{"id":60826833,"name":"Biological sciences/Computational biology and bioinformatics/Machine learning"},{"id":60826834,"name":"Biological sciences/Computational biology and bioinformatics/Predictive medicine"}],"tags":[],"updatedAt":"2026-04-23T16:00:25+00:00","versionOfRecord":[],"versionCreatedAt":"2026-01-21 10:06:26","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8110055","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8110055","identity":"rs-8110055","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}