New Breakthrough in PPG Non-Invasive Blood Pressure Monitoring: BP-Inference-EVOLVE and Stochastic Experiment Loop Framework

This article introduces the open-source BP-Inference-EVOLVE project by the vignankamarthi team. Using an LLM-driven "Stochastic Experiment Loop" framework, the project achieves cuffless blood pressure monitoring with only PPG signals and meets the international medical standards of AAMI/BHS on the PulseDB v2.0 dataset. Key highlights include pure PPG single-signal inference, dual modes (calibration-free/calibration-required), and an innovative scientific research automation framework, bringing significant progress to wearable health monitoring.

Traditional cuff-based blood pressure monitors are less portable and cannot perform continuous monitoring; invasive arterial monitoring is only suitable for intensive care. While PPG-based solutions are convenient, they have accuracy issues: large individual physiological differences, sensor position/pressure affecting signal quality, motion artifact interference, and the need for individual calibration. Theoretically, PPG can estimate blood pressure via pulse waveforms, but practical applications face multiple obstacles.

BP-Inference-EVOLVE achieves blood pressure estimation using pure PPG single signals, expanding application scenarios (supported by ordinary smartwatches); verified on the PulseDB v2.0 dataset to meet AAMI/BHS standards, with clinical application value; and supports dual modes: calibration-free (for universal screening) and calibration-required (higher accuracy for long-term monitoring).

This framework combines LLM with evolutionary algorithms: 1. Automated hypothesis generation (LLM proposes model architecture hypotheses); 2. Intelligent experiment design; 3. Evolutionary optimization (iterative improvement); 4. Knowledge accumulation. LLM acts as a research assistant: reading literature, analyzing results, providing suggestions, and writing code. Compared to manual design, its advantages include a large exploration space, reduced bias, continuous learning, and strong interpretability.

The project uses the PulseDB v2.0 dataset for validation, which features large-scale diversity (thousands of subjects covering different ages, genders, health statuses, various activity states, and synchronized gold-standard blood pressure measurements). The AAMI standard requires an average error ≤5mmHg and standard deviation ≤8mmHg; the BHS Class A standard requires the proportion of errors ≤5mmHg to be ≥60%, ≤10mmHg ≥85%, and ≤15mmHg ≥95%—the project meets these standards.

Application scenarios include: consumer health monitoring (continuous blood pressure tracking via smartwatches, nighttime monitoring, post-exercise assessment); chronic disease management (reducing patients' medical burden, optimizing treatment plans, timely detection of abnormalities); and scientific research value (providing methodological references for medical AI and promoting the development of AI for Science).

Current limitations: skin color affects accuracy, accuracy needs improvement during intense exercise, and limited extrapolation ability for extreme blood pressure values. Future directions: multi-modal fusion (combining accelerometers, etc.), federated learning (using more data under privacy protection), causal reasoning (understanding the causes of blood pressure changes), and large-scale clinical validation.