JMIR Publications

ABSTRACT

Background:

Acute Kidney Injury (AKI) is a common, life-threatening complication among hospitalized patients. Timely prediction is crucial for prevention, while accurate prognosis following onset is essential for guiding care. Although electronic health records (EHRs) provide rich longitudinal data, their high dimensionality, heterogeneity, and temporal complexity pose challenges for traditional machine learning.

Objective:

To develop and evaluate a transformer-based model for accurate and robust AKI prediction and prognosis by modeling long-range temporal dependencies and complex relationships in heterogeneous EHR data with enhanced token-level interpretability.

Methods:

We developed the Biomarker-Aware Transformer for AKI (BAT-AKI), which integrates multi-dimensional embeddings capturing temporal dynamics, textual descriptions, hierarchical ontology structures, and semantic relationships in EHR. To enhance sensitivity to AKI-related biomarker fluctuations, BAT-AKI combines masked language modeling with a novel biomarker abnormality-aware pretraining task. The model was pretrained and fine-tuned for four tasks: AKI-onset, early reversal, recovery at discharge, and 90-day post-AKI mortality. Pretraining and internal validation used data from one Patient-Centered Outcomes Research Network (PCORnet) site, with external validation across three additional sites.

Results:

BAT-AKI outperformed all baseline models across tasks. For AKI-onset prediction, it achieved an AUROC of 0.8546 and an AUPRC of 0.6007, representing improvements of 2.1% and 6.1% respectively over the second-best baseline. Across prognostic tasks, BAT-AKI achieved consistent gains of 2–8%, with AUROCs of 0.5964, 0.6014, and 0.7998 and AUPRCs of 0.6981, 0.6929, and 0.5143 for early reversal, recovery at discharge, and 90-day post-AKI mortality, respectively. Performance of BAT-AKI remained stable across external cohorts, with AKI-onset prediction achieving AUROCs ranging from 0.7755 to 0.7993 and AUPRCs ranging from 0.4759 to 0.5646, corresponding to improvements of 3%–7% in AUROC and 17%–30% in AUPRC over the baseline. In low-resource settings, using only 2% of fine-tuning data, BAT-AKI achieved an AUROC of 0.7877 and an AUPRC of 0.4781 for AKI-onset prediction, corresponding to relative improvements of 5.8% in AUROC and 10.5% in AUPRC over the second-best baseline. Visualization of embeddings and attentions highlighted clinically meaningful relationships and patient-specific trajectories.

Conclusions:

BAT-AKI enables accurate, interpretable, and generalizable AKI prediction and prognosis across institutions, supporting its potential for early warning and AKI care optimization.