To systematically examine IDH recurrence patterns and evaluate their impact on model performance, while identifying methodological improvements.
Approach:
Data Analysis: Retrospective analysis of 12,767 hemodialysis sessions from 66 patients, focusing on recurrent IDH events defined as occurring ≥30 min after the initial IDH.
Model Comparison: Comparison of deep learning models (ConvMixer, temporal convolutional network, long short-term memory with attention) against a naïve baseline that predicted IDH solely from prior occurrences.
Model Enhancement: Incorporation of recurrence-aware features and loss weighting to improve predictive performance, evaluated across systolic blood pressure subgroups.
Key Findings:
The probability of IDH increased from 0.7–10.4% before initial events to 11.7–65.7% thereafter.
Conventional evaluation methods overestimated model performance, particularly in distinguishing between initial and recurrent IDH predictions.
Incorporating recurrence-aware features improved AUROC by up to 8.0 percentage points across different architectures.
Adversarial training reduced subgroup disparities while maintaining overall model performance.
Interpretation:
Incorporating recurrence patterns into IDH prediction models enhances accuracy and robustness, suggesting a need for standardized evaluation protocols that account for recurrence.
Limitations:
The study was retrospective and did not perform formal sample size calculations, which may limit the reliability of the findings.
Only sessions lasting at least 2 hours were analyzed, which may limit generalizability.
Conclusion:
Incorporating recurrence patterns into IDH prediction models demonstrated improvements in accuracy and robustness, highlighting the need for standardized evaluation protocols.
