<![CDATA[Invasive methods of blood sampling can cause mild pain or anxiety for patients. Non-invasive devices developed have the potential for finger shifting or misplacement, resulting in data noise. Erick et al. conducted a study extracting features from photoplethysmograms to estimate blood cholesterol. However, this study focused solely on calculating estimated cholesterol values. This study aimed to gain insight into the potential impact of data processing on visual spectroscopy data. Preprocessing studies were conducted on the visual spectroscopy dataset using value conversion, transpose, and CTGAN as sampling-based data augmentation methods. After separate training of all four methods, it was found that dataset size, both the number of features and the number of samples, significantly impacted the deep learning model, particularly the relationship between labels and features. Although the transpose method provided the highest R2 score of 0.123 with an MAE of 1.03 in the CNN-1D model, compared to the MLP model, which produced an R2 score of 0.122, and the LSTM model, which produced an R2 coefficient of 0.026, the accuracy of uric acid prediction still did not reach clinical levels.]]>
Copyrights © 2026
