<p>Monitoring blood glucose levels (BGL) is vital for the medical care and management, driving demand for continuous non-invasive BGL monitoring methods that operate beyond clinical settings. Photoplethysmography (PPG) is a promising approach for its accessibility, affordability, user-friendliness, and seamless integration with wearable devices. Recent innovations in smartphone-based PPG signal capturing further enhance its practicality. While artificial intelligence (AI) techniques have proven potential in estimating BGL from PPG signals, existing models often suffer from high computational complexity or inadequate integration of temporal and spatial features, limiting their clinical accuracy. To address these challenges, we propose a hybrid deep learning model aimed at improving BGL monitoring accuracy. Specifically, our model incorporates a Bi-LSTM for temporal feature learning and two CNNs–one focused on global spatial (Macro-CNN) and local spatial (Micro-CNN)–for spatial feature learning. The proposed model achieved a mean absolute error (MAE) of 13.8 mg/dL, a root mean squared error (RMSE) of 16.84 mg/dL, and a coefficient of determination (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">$\hbox {R}^{2}$</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mtext>R</mtext> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation>) of 0.52. Clinical validity was checked through Clarke’s error grid (CEG) and Parkes’ error grid (PEG) analyses, with <InlineEquation ID="IEq2"> <EquationSource Format="TEX">$92\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>92</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> of estimated BGL values residing in the clinically proper Zone A and the other <InlineEquation ID="IEq3"> <EquationSource Format="TEX">$8\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>8</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> in clinically acceptable Zone B. These results on independent subjects demonstrate a step toward practical, real-world deployment of non-invasive glucose monitoring systems.</p>

Hybrid deep learning-based blood glucose level monitoring using PPG signals

期刊论文
发表日期：

2026年5月5日

查看全文

Abdelrhman Y. Soliman,
Ahmed M. Nor,
Osama A. Omer,
Ahmed S. Mubarak

摘要

Monitoring blood glucose levels (BGL) is vital for the medical care and management, driving demand for continuous non-invasive BGL monitoring methods that operate beyond clinical settings. Photoplethysmography (PPG) is a promising approach for its accessibility, affordability, user-friendliness, and seamless integration with wearable devices. Recent innovations in smartphone-based PPG signal capturing further enhance its practicality. While artificial intelligence (AI) techniques have proven potential in estimating BGL from PPG signals, existing models often suffer from high computational complexity or inadequate integration of temporal and spatial features, limiting their clinical accuracy. To address these challenges, we propose a hybrid deep learning model aimed at improving BGL monitoring accuracy. Specifically, our model incorporates a Bi-LSTM for temporal feature learning and two CNNs–one focused on global spatial (Macro-CNN) and local spatial (Micro-CNN)–for spatial feature learning. The proposed model achieved a mean absolute error (MAE) of 13.8 mg/dL, a root mean squared error (RMSE) of 16.84 mg/dL, and a coefficient of determination ( $\hbox {R}^{2}$ $R^{2}$ ) of 0.52. Clinical validity was checked through Clarke’s error grid (CEG) and Parkes’ error grid (PEG) analyses, with $92\%$ $92 %$ of estimated BGL values residing in the clinically proper Zone A and the other $8\%$ $8 %$ in clinically acceptable Zone B. These results on independent subjects demonstrate a step toward practical, real-world deployment of non-invasive glucose monitoring systems.