Scholarlink

<p>This study aimed to create AgeVision, a new framework that can improve age prediction based on a synergistic combination of physiological biomarkers and facial analysis. Age estimation is a convoluted task with relevant implications in the fields of medicine and forensics, where the use of a unimodal approach sometimes fails to deliver a comprehensive evaluation. Our proposal has two plans to attain the constraints in data: a stacking ensemble of unpaired public information, as well as a concatenation-based deep learning model that end-to-end learns a custom, paired data. To combine the extraction of facial features with a CNN of ResNeXt50 and the analysis of biomarkers with an XGBoost regressor, our stacking model was followed by a meta-learner to make the final predictions. This model of large-scale public data (UTKFace: above 20,000 images; cardiovascular: 70,000 records) obtained 2.55&#xa0;years of Mean Absolute Error (MAE), which is nearly two-fifths and two-thirds of the model based on the unimodal face (MAE: 3.43&#xa0;years) and biomarker (MAE: 5.41&#xa0;years), respectively. Our concatenation model, which was trained on our local paired dataset (<i>n</i> = 101 subjects), showed that it is possible to directly fuse features, with a significantly larger MAE of 8.2&#xa0;years, emphasizing the limitations of paired data. The Huber result showed that the MAE was 2.8 in the stacking-based model and 8.7 for the concatenation model. For the paired local dataset, we used a concatenation-based model where a six-layer-based CNN model and multilayer perceptron (MLP) were used. The explanation of the decision-making process of the model was confirmed through explainable AI (XAI) methods, such as Grad-CAM and SHAP which determined that the model focused on regions of the face and biomarkers that are biologically reasonable. In addition, we provide real-world utilization using two deployed web applications that serve age predictions in real time and are interpretable. We find that the multimodal system of integration provides a more robust, accurate and clinically viable system of age estimation, conclusively.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

AgeVision: Integrating Biomarkers and Facial Analysis for Accurate Age Prediction

  • 期刊论文

  • 发表日期：

    2026年2月23日

查看全文
  • Md. Azmain Asif Bhuiyan,
  • Mohammad Tawsif,
  • Aditta Chowdhury,
  • Mehdi Hasan Chowdhury

摘要

This study aimed to create AgeVision, a new framework that can improve age prediction based on a synergistic combination of physiological biomarkers and facial analysis. Age estimation is a convoluted task with relevant implications in the fields of medicine and forensics, where the use of a unimodal approach sometimes fails to deliver a comprehensive evaluation. Our proposal has two plans to attain the constraints in data: a stacking ensemble of unpaired public information, as well as a concatenation-based deep learning model that end-to-end learns a custom, paired data. To combine the extraction of facial features with a CNN of ResNeXt50 and the analysis of biomarkers with an XGBoost regressor, our stacking model was followed by a meta-learner to make the final predictions. This model of large-scale public data (UTKFace: above 20,000 images; cardiovascular: 70,000 records) obtained 2.55 years of Mean Absolute Error (MAE), which is nearly two-fifths and two-thirds of the model based on the unimodal face (MAE: 3.43 years) and biomarker (MAE: 5.41 years), respectively. Our concatenation model, which was trained on our local paired dataset (n = 101 subjects), showed that it is possible to directly fuse features, with a significantly larger MAE of 8.2 years, emphasizing the limitations of paired data. The Huber result showed that the MAE was 2.8 in the stacking-based model and 8.7 for the concatenation model. For the paired local dataset, we used a concatenation-based model where a six-layer-based CNN model and multilayer perceptron (MLP) were used. The explanation of the decision-making process of the model was confirmed through explainable AI (XAI) methods, such as Grad-CAM and SHAP which determined that the model focused on regions of the face and biomarkers that are biologically reasonable. In addition, we provide real-world utilization using two deployed web applications that serve age predictions in real time and are interpretable. We find that the multimodal system of integration provides a more robust, accurate and clinically viable system of age estimation, conclusively.