JMIR Publications

ABSTRACT

Background:

Intraoperative hypotension (IOH) during artificial liver support system (ALSS) compromises treatment completion and patient outcomes. Existing prediction models rely on high-fidelity arterial waveform analysis and target surgical populations, limiting applicability to ALSS patients with complex pathophysiology and frequent coagulopathy.

Objective:

We aimed to develop HemaFormer, a deep learning model integrating static patient characteristics with intraoperative temporal data to predict IOH during ALSS without requiring invasive waveforms.

Methods:

Methods:

This multicenter retrospective study analyzed 1,405 ALSS sessions from four hospitals (2019-2025). HemaFormer employs a Transformer with a novel feature-wise linear modulation (FiLM) layer. Static features (655-dimensional vector covering demographics, comorbidities, and laboratory parameters) generate modulation parameters that conditionally transform intraoperative temporal features via element-wise affine transformation, enabling personalized interpretation of dynamic signals based on baseline pathological context. The model was trained using composite Focal and Pairwise Ranking Loss, and evaluated on an independent test set with patient-level splitting.

Results:

Results:

IOH occurred in 36.5% (513/1,405) of sessions. Higher BMI (>24 kg/m²), female sex, and lower serum sodium were independently associated with increased IOH risk. Hemodynamic divergence emerged as early as 30 minutes post-initiation, marked by progressive systolic pressure decline and compensatory tachycardia in the IOH cohort. On the independent test set, HemaFormer achieved PR-AUC 0.580 and ROC-AUC 0.582, outperforming XGBoost and logistic regression across all metrics. Subgroup analysis revealed optimal performance in the DPMAS-only modality (PR-AUC 0.614).

Conclusions:

Conclusions:

HemaFormer enables effective IOH prediction during ALSS by fusing static and temporal clinical data. Its core innovation, the FiLM mechanism, modulates dynamic signal interpretation according to static pathological context, offering a novel computational paradigm for understanding hemodynamic instability in complex liver disease without invasive monitoring.