Minseong Kim

Photoplethysmography (PPG)-based blood pressure (BP) estimation remains challenging due to inherent signal variability and the indirect nature of the measurement. However, although BP prediction is a high-stakes domain where accuracy is crucial, existing studies often overlook multimodal data that is commonly available in real-world healthcare scenarios, thereby limiting estimation accuracy. To address this gap, we introduce the Signal-driven and Demographic-Clinical Attribute dataset for BP Estimation (SABP), which integrates diverse modalities, including physiological signals, structured clinical attributes, and disease-related medical text, to comprehensively capture causally linked factors related to BP. The multimodal nature of SABP facilitates a deeper understanding of complex interactions between BP and various clinical attributes across different data modalities. However, modeling these interactions can be challenging due to unobserved variables or hidden confounders that potentially bias the underlying interaction of BP and PPG waveforms. To mitigate this issue, we propose the Causal-aware Blood Pressure Estimator (CBPE). CBPE leverages a Variational Autoencoder (VAE) to infer latent representations of hidden confounders, thereby improving the accuracy and robustness of BP estimation by effectively reducing hidden confounding effects. To complement model evaluation, we introduce the Group Loss Consistency (GLC) metric, which jointly assesses predictive performance and fairness across different BP groups. Unlike traditional error metrics, GLC explicitly penalizes performance disparities, encouraging reliable estimation across patient categories. Experimental results show that CBPE achieves competitive BP prediction accuracy while demonstrating superior robustness to subgroup bias and performance degradation under challenging conditions. Reproducible materials are available at https://github.com/MLAI-Yonsei/CBPE

Recent advancements in deep learning for predicting Arterial Blood Pressure (ABP) have prominently featured the use of photoplethysmography (PPG) signals. Notably, PPG signals exhibit significant variability due to differences in measurement environments, alongside stark disparities in the distribution of collected signal data among different labels. To address these challenges, this study introduces a Bi-Guided Diffusion Model designed to generate PPG signals with expected features of ABP within a few-shot setting for each label group. We propose a guided diffusion model architecture that can simultaneously consider both the determinant group condition and the continuous label condition for each group in a few-shot setting. To our knowledge, this is the first study to use a diffusion model for generating PPG signals with a limited dataset. Initially, we categorized them into four groups based on SBP and DBP values: Hypo, Normal, Prehyper, and Hyper2. In each group, we sample an equal number of data points according to the few-shot setting and then generate appropriate PPG signals for each group through guidance. Additionally, Our study proposes a post-processing technique to address the limitations of generative models in few-shot settings, consistently boosting performance across various methods such as training from scratch, transfer learning, and linear probing. When benchmarked, our methodology demonstrated performance improvements across all datasets, including BCG, PPGBP, and SENSORS. We confirmed data quality by comparing training, generated, and actual data and analyzed error cases, morphology features, and t-SNE distribution to highlight the role of synthetic data in enhancing performance.

Intradialytic hypotension (IDH) is a common side effect that occurs during hemodialysis and poses a great risk for dialysis patients. Many studies have been conducted so far to predict IDH, but most of these could not be applied in real-time because they used only underlying patient information or static patient disease information. In this study, we propose a multilayer perceptron (MP)-based IDH prediction model using heart rate (HR) information corresponding to time-series information and static data of patients. This study aimed to validate whether HR differences and HR slope information affect real-time IDH prediction in patients undergoing hemodialysis. Clinical data were collected from 80 hemodialysis patients from 9 September to 17 October 2020, in the artificial kidney room at Yeungnam University Medical Center (YUMC), Daegu, South Korea. The patients typically underwent hemodialysis 12 times during this period, 1 to 2 h per session. Therefore, the HR difference and HR slope information within up to 1 h before IDH occurrence were used as time-series input data for the MP model. Among the MP models using the number and data length of different hidden layers, the model using 60 min of data before the occurrence of two layers and IDH showed maximum performance, with an accuracy of 81.5%, a true positive rate of 73.8%, and positive predictive value of 87.3%. This study aimed to predict IDH in real-time by continuously supplying HR information to MP models along with static data such as age, diabetes, hypertension, and ultrafiltration. The current MP model was implemented using relatively limited parameters; however, its performance may be further improved by adding additional parameters in the future, further enabling real-time IDH prediction to play a supporting role for medical staff.