Contactless hemoglobin estimation from facial video using edge computing and sliding-window analysisUç bilişim ve kayan pencere analizi kullanarak yüz videosundan temassız hemoglobin tahmini
Abstract
Hemoglobin is a key biomarker for anemia screening and clinical decision support. Yet, standard measurement is invasive and can be difficult to deploy for rapid, repeatable assessment in field and disaster scenarios. This motivates contactless estimation approaches that can run reliably on resource-constrained edge devices. Objective—This study investigates contactless peripheral hemoglobin (SpHb) estimation from facial videos with an explicit edge-computing viewpoint. The central aim is to quantify how sliding-window length acts as a system-level design parameter that jointly governs (i) prediction accuracy, (ii) temporal stability of sequential outputs, and (iii) power/energy cost on an embedded platform. Method—The dataset contains 279 facial videos, each paired with a single SpHb label. Each video is divided into five segments, and training samples are generated using sliding-window configurations of 30, 60, 90, and 120 frames. Evaluation follows a subject-independent split to prevent data leakage. Algorithmically, 3D CNN baselines are compared with hybrid spatiotemporal architectures (CNN–VAN–Transformer). Performance is assessed using regression metrics (RMSE/MAE) and complemented by time-series inspection on representative videos to discuss output stability under different window lengths. Results—On 60-frame segments, the 3D CNN achieves an RMSE of 0.4808, while the CNN–VAN–Transformer yields an RMSE of 0.7128. On 90-frame segments, the CNN–VAN–Transformer provides the best accuracy with an RMSE of 0.6946, compared with 0.8894 for the 3D CNN; V1-3DCNN degrades to 1.5828, indicating that longer context does not guarantee improvement and that window–architecture interaction is substantial. Video-level MAE varies by content and target range: for example, the best MAE occurs at 120 frames for HEM022 (0.350) and HEM188 (0.150), at 60 frames for HEM086 (0.161), and at 30 frames for HEM028 (0.512), highlighting the absence of a universal optimum window length across subjects/videos. Conclusion—Edge measurements on a Raspberry Pi 5 using an ACS711 current sensor and microcontroller-triggered logging show a monotonic energy increase with window size: mean power rises from 2.481 W (30-frame) to 3.951 W (120-frame), and mean energy from 1.029 J to 5.047 J. The maximum measured energy is 7.14 J (HEM022, 120-frame), while the minimum is 0.48 J (HEM028, 30-frame); peak power reaches 6.80 W in the 120-frame setting. These findings demonstrate that sliding-window length is not merely a tuning knob but a primary design decision that directly balances accuracy, output stability, and energy budget in practical, deployable SpHb estimation systems.
Keywords
Destekleyen Kurum
Proje Numarası
Etik Beyan
Kaynakça
- [1] E. J. Wang, W. Li, D. Hawkins, T. Gernsheimer, C. Norby-Slycord, S. N. Patel, “HemaApp: noninvasive blood screening of hemoglobin using smartphone cameras”, ACM International Joint Conference on Pervasive and Ubiquitous Computing (UbiComp), Heidelberg, Germany, 12-16 September 2016, 593–604.
- [2] R. G. Mannino, D. R. Myers, E. A. Tyburski, C. Caruso, J. Boudreaux, T. Leong, G. D. Clifford, W. A. Lam, “Smartphone app for non-invasive detection of anemia using only patient-sourced photos,” Nature Communications, 9, (2018), 4924.
- [3] A. X. Zhang, J. J. Lou, Z. J. Pan, J. Q. Luo, X. M. Zhang, H. Zhang, J. P. Li, L. L. Wang, X. Cui, B. Ji, L. Chen, “Prediction of anemia using facial images and deep learning technology in the emergency department”, Frontiers in Public Health, 10, (2022), 964385.
- [4] Y. W. Chen, X. Y. Hu, Y. Z. T. Zhu, X. Liu, B. Yi, “Real-time non-invasive hemoglobin prediction using deep learning-enabled smartphone imaging”, BMC Medical Informatics and Decision Making, 24(1), (2024), 187.
- [5] D. Botina-Monsalve, Y. Benezeth, J. Miteran, “RTrPPG: An ultra light 3DCNN for real-time remote photoplethysmography”, IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), New Orleans, LA, USA, 18-24 June 2022, 2145–2153.
- [6] U. Bal, F. E. Oguz, K. M. Sunnetci, A. Alkan, A. Bal, E. Akkus, H. Erol, A. Ç. Seçkin, “Estimation of Total Hemoglobin (SpHb) from Facial Videos Using 3D Convolutional Neural Network-Based Regression”, Biosensors-Basel, 15(8), (2025), 485.
- [7] P. P. Kumar, A. Pal, K. Kant, “Resource Efficient Edge Computing Infrastructure for Video Surveillance”, IEEE Transactions on Sustainable Computing, 7(4), (2022), 774–785.
- [8] M. A. Khan, R. Hamila, A. Erbad, M. Gabbouj, “Distributed Inference in Resource-Constrained IoT for Real-Time Video Surveillance”, IEEE Systems Journal, 17(1), (2023), 1512–1523.
Ayrıntılar
Birincil Dil
İngilizce
Konular
Siberfizik Sistemleri ve Nesnelerin İnterneti , Yapay Zeka (Diğer) , Elektronik , Gömülü Sistemler
Bölüm
Araştırma Makalesi
Yazarlar
Ufuk Bal
0000-0003-0345-6989
Türkiye
Abdulbaki Akgün
0009-0009-0802-9207
Türkiye
Sahra Nur Pakel
0009-0004-8549-2987
Türkiye
Hüseyin Sezerol
0009-0006-5274-0205
Türkiye
Alkan Bal
0000-0002-7884-1251
Türkiye
Erken Görünüm Tarihi
25 Nisan 2026
Yayımlanma Tarihi
-
Gönderilme Tarihi
2 Şubat 2026
Kabul Tarihi
1 Nisan 2026
Yayımlandığı Sayı
Yıl 2026 Sayı: Advanced Online Publication