TRANSFORMER-BASED MODELLING OF ARTERIAL BLOOD PRESSURE WAVEFORMS FOR ICU MORTALITY PREDICTION

Continuous arterial blood pressure (ABP) waveforms are routinely recorded in intensive care units (ICUs) but are rarely used beyond displaying mean values. ABP waveforms contain rich physiological information that may reflect previously unknown signs of patient instability and predict adverse outcomes. We investigated whether a transformer-based deep learning model could extract clinically meaningful features from ABP waveforms to predict ICU mortality using the MIMIC-III Waveform Database.

The MIMIC-III Waveform Database contains 67,830 ABP waveform sets linked to 30,000 ICU patients [1,2]. We filtered to physiologically plausible values (30–200 mmHg), removed artefacts, and performed linear interpolation of short gaps (<5 samples; 0.04 s). We adopted the Biosignal Transformer, originally designed for ECG and EMG signals, to ABP waveform prediction tasks [3, 5]. Pre-training was performed on 612,720 hours of ABP waveform data. Model hyperparameters were optimised using Maximal Update Parameterization [6]. A scaling law was developed to estimate the optimal training steps as a function of parameter count and compute budget [4, 5]. Temporal aggregation was used to average the model’s probability predictions over consecutive time windows (e.g., 1- or 12-hour periods) for each patient to reduce short-term variability. Fine-tuning targeted two classification tasks: 1. Mortality within 2 hours, and 2. Mortality within 24 hours. Context window lengths (12–384 s) and embedding sizes were varied to assess the impact of temporal context and model capacity on performance. We visualised per-patient risk trajectories as heatmaps (Figure 1), averaging model probabilities within fixed bins (2-h task: 1-h bins; 24-h task: 12-h bins).

In total, 1510 ABP waveform sets from 919 patients were included in this analysis (post pre-training). Temporal aggregation improved robustness and discrimination across tasks. For 2-hour mortality prediction, validation performance increased from PR-AUC 0.097 to 0.304 (random baseline: 0.002) and AUROC 0.774 to 0.844 with temporal aggregation of model predictions across 1-hour windows (Table 1). For 24-hour mortality prediction, validation performance increased from PR-AUC 0.073 to 0.166 (random baseline: 0.036) and AUROC 0.688 to 0.825 with temporal aggregation of model predictions across 12-hour windows (Table 2). During fine-tuning, we found that models with larger context sizes benefited most from higher embedding dimensions, with AUROC improving from 0.625 to 0.670 for the 48 s model (embedding 128→512), while shorter 12 s models showed minimal change (0.644→0.646). Models with a larger context window showed better overall performance at the window level, achieving 0.644 AUROC for the 12-second model and 0.674 for the 384-second model. Figure 1 shows rising predicted risk approaching death in mortality cases and predominantly low, stable probabilities in non-mortality cases.

This study demonstrated that biosignal transformer models trained on ABP waveforms can predict near-term ICU mortality with clinically meaningful discrimination. Temporal aggregation markedly improves robustness by reducing prediction noise. This study establishes a scalable framework for ABP waveform-based clinical risk modelling and demonstrates the feasibility of transformer architectures for outcome prediction in critical care and other domains.