Forecasting cumulative intraoperative hypotension burden: patient-disjoint evaluation of quantile boosting and alert policies

Yunfan ZhaoarXiv 2026 · arXiv

TL;DR

This paper shifts the focus of intraoperative hypotension (IOH) forecasting from simple binary "yes/no" events to predicting the cumulative burden—the depth and duration of low blood pressure. By applying quantile boosting to high-resolution surgical data, the study establishes a rigorous benchmark that separates statistical forecasting accuracy from the actual clinical utility of resulting alerts.

What Problem Does This Paper Address?

Intraoperative hypotension (IOH) is a common surgical complication linked to serious postoperative issues like acute kidney injury and even mortality. However, most existing AI models treat IOH as a binary classification problem: "Will the patient's blood pressure drop below a threshold in the next few minutes?"

The authors argue this is insufficient. A brief, shallow dip in blood pressure carries a very different clinical risk than a prolonged, deep excursion. Clinicians need to know the burden—the cumulative area under the threshold curve. Furthermore, a model that is statistically "accurate" (minimizing mathematical error) might still produce clinically "useless" alerts that are poorly timed or too frequent, leading to alarm fatigue. This paper aims to bridge the gap between statistical performance and clinical decision support.

Methods

The researchers utilized Quantile Boosting (specifically using XGBoost) to predict the 90th and 95th percentiles of the "MAP deficit" (Mean Arterial Pressure below 65 mmHg) over the next 15 minutes.

Key methodological choices included:

  • Target Functional: Instead of predicting a single mean value, they predicted upper quantiles to capture "worst-case" scenarios.
  • Direct vs. Residual Models: They compared "direct" models (predicting burden from raw history) against "residual" models (predicting the error of a simple baseline that assumes current pressure will persist).
  • Patient-Disjoint Evaluation: To ensure the model generalizes to new people, they split the data so that no patient's data appeared in both training and testing sets.
  • Alert Replay: They simulated how these forecasts would trigger actual bedside alerts using different policies (e.g., simple thresholds vs. hysteresis rules that require a significant recovery before turning off).

Datasets

The study used VitalDB, a large, publicly available dataset of de-identified intraoperative recordings.

  • Scope: 3,423 surgical cases and 3,304 unique subjects.
  • Data Points: Over 100,000 labeled 15-minute windows derived from 10-second median blood pressure readings.
  • Benchmark: The authors established a reproducible benchmark for future researchers to compare against, emphasizing explicit cohort derivation and "test freezing" (locking the analysis before final evaluation) to ensure integrity.

Key Insights and Ideas

  • Direct is Better than Residual: Interestingly, models that predicted the burden directly from history outperformed "residual" models that tried to correct a baseline persistence forecast. This suggests that the complex patterns in blood pressure history contain more information than just the "current state."
  • The Statistical-Clinical Disconnect: A major "aha moment" was confirming that lower forecast loss (better math) does not automatically lead to better alerts. A model might be very good at predicting the amount of burden but still struggle with the timing required to give a useful early warning.
  • Weighting Matters: The study highlights how technical implementation details, like how you normalize patient weights in XGBoost, can significantly shift model rankings. This underscores the need for transparent reporting in medical AI.

Key Results

The primary metric for forecasting was patient-weighted pinball loss (where lower is better).

Table 3: Patient-weighted pinball loss results. Direct models consistently outperformed baselines and residual models.
  • Forecasting Performance: At the 90th percentile, the direct model achieved a loss of 2.196 mmHg·min, significantly better than the "calibrated persistence" baseline of 2.900.
  • Alert Performance: While the models were statistically superior, their clinical utility remained modest. At a fixed budget of 0.25 false alerts per hour, the best model only captured about 16.3% of eligible hypotension episodes.
Table 5: Test alert results at the validation budget of 0.25 unmatched alerts per decision-exposure hour. False rates and total activation rates are distinct.

Limitations and Future Work

The authors are careful to note that this is a retrospective internal evaluation.

  • Treatment Bias: The data reflects "treated" patients; future blood pressure is influenced by the actions clinicians took in response to what they saw on the monitors.
  • Single-Center: The results might not "transport" to other hospitals with different patient populations or monitoring equipment.
  • Standalone Utility: The low sensitivity and precision suggest these models are not yet ready to be standalone "early warning" systems.

Future Work: The authors suggest exploring multi-channel data (adding heart rate or medications) and investigating whether Large Language Models (LLMs) could help summarize these numerical forecasts into more actionable bedside insights, though they emphasize that an LLM is not a substitute for the underlying numerical forecasting engine.

Summary updated

Abstract

Background: Binary hypotension forecasts do not describe the depth and duration of future low pressure. Burden forecasts may support graded early warning, but better forecast scores need not yield more useful alerts. Objective: To evaluate conditional quantiles of the next 15-minute mean arterial pressure (MAP) deficit below 65 mmHg and distinguish forecasting performance from episode-warning performance. Methods: Public VitalDB numeric arterial MAP was reduced to 10-second medians without forward filling. Subject-disjoint training, validation and test splits included 70,402, 14,270 and 15,575 labeled windows. We compared climatology, calibrated persistence, last-MAP-only boosting, and direct and residual quantile boosting, with and without additional MAP-history features. Patient-weighted pinball loss and subject-bootstrap intervals evaluated forecasting. Validation selected alert thresholds before a hash-recorded test evaluation. Results: In the test split, the prespecified residual-versus-persistence comparison at the 90th percentile yielded losses of 2.256 and 2.900 mmHg-min, respectively (difference -0.644; 95% interval -0.814 to -0.487). Direct prediction achieved 2.196, and direct prediction with additional MAP features achieved 2.193. Episode-warning results were sensitive to the alert rule and operating threshold. Conclusions: This retrospective benchmark separates burden-forecast skill from warning-policy performance. Findings support further study of direct burden estimation; they do not establish patient benefit, prospective reliability or transportability beyond the source hospital.
BibTeX
@misc{zhao2026forecasting,
  author = {Yunfan Zhao},
  title = {{Forecasting cumulative intraoperative hypotension burden: patient-disjoint evaluation of quantile boosting and alert policies}},
  year = {2026},
  eprint = {2611.09259},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.09259},
}
Figure