Noninvasive ML-Based BUN Prediction from Ten-Second Fingertip Video Using Generative AI

The measurement of Blood Urea Nitrogen (BUN) is an important test for examining the kidney condition. BUN is a waste product that is eliminated from the body by the kidneys. Increased BUN level often indicates renal diseases. The invasive blood test is used as the standard BUN testing method. In this paper, we have demonstrated an effective and innovative method for noninvasive BUN prediction from Ten-second fingertip video. We use the Ten-second fingertip video of human subjects to create photo plethysmography (PPG) features through signal processing and feature engineering techniques. We utilize Generative Adversarial Networks (GAN), a type of Generative AI, as a data augmentation method to create synthetic data based on our original data. Then, we use the PPG features and Gold standard BUN level of subj ects to develop four regression models before and after the implementation of GAN. We compare all models' performance in terms of MAE, MSE, and RMSE. Before applying GAN, ANN outperforms other models with an MAE of 1.17, MSE of 2.62, and RMSE of 1.62. We find that GAN implementation significantly improves every model's performance and provides lower MAE, MSE, and RMSE. After applying GAN, ANN performs best among all other models and shows the lowest MAE of 0.69, MSE of 1.03, and RMSE of 1.01. With these outstanding results, our noninvasive method of BUN prediction has the potential to offer enhanced patient comfort, early detection, continuous and remote monitoring of renal health.