Background: Continuous blood pressure monitoring is vital for patient safety during surgery. However, clinicians face a dilemma: invasive methods carry risks, while noninvasive cuff measurements are intermittent and can miss dangerous periods of trans...
Background: Continuous blood pressure monitoring is vital for patient safety during surgery. However, clinicians face a dilemma: invasive methods carry risks, while noninvasive cuff measurements are intermittent and can miss dangerous periods of transient hypotension. This research investigates the use of a Large Language Model (LLM) to provide a novel solution for noninvasive, near real-time, and continuous mean arterial pressure (MAP) estimation.
Objective: The goal of this study was to develop a lightweight, parameter-efficient framework for continuous MAP estimation using noninvasive biosignals (ECG, PPG, CO₂). The core challenge was to effectively adapt a frozen, pre-trained LLM for a complex physiological time-series regression task, ensuring computational efficiency for potential real-time clinical use.
Methods: Using the large-scale intraoperative VitalDB dataset, we developed a model featuring a multi-scale 1D-CNN front-end that extracts relevant features from raw waveforms. This extractor employs learnable Sinc-Convolution filters and an attention mechanism to fuse multimodal signals. To bridge the domain gap between numerical data and the LLM, we implemented key techniques such as patch reprogramming and structured natural language prompting. The model was trained using a weighted Huber loss function to penalize errors more heavily in the clinically critical hypotensive range, thereby prioritizing patient safety.
Results: The final model, utilizing a LLaMA backbone, achieved a Mean Absolute Error (MAE) of 7.72 mmHg. Ablation studies confirmed that the specialized feature extractor and prompt-based guidance were crucial for this performance. While this accuracy does not yet meet the stringent AAMI standard for medical devices (mean error ≤ 5 mmHg), it is highly promising for application in clinical early warning systems and patient triage.
Conclusion: This study successfully demonstrates that a frozen LLM can be effectively adapted for accurate, continuous MAP estimation from noninvasive biosignals. The proposed framework offers a parameter-efficient pathway to broaden the application of continuous hemodynamic monitoring, potentially enabling earlier intervention for hypotension and improving surgical outcomes.