Open-Source Neural Network vs. Simple Hemodynamic Rules for Intraoperative Hypotension Prediction

Introduction

Intraoperative hypotension (IOH) is a common complication of general anesthesia and is associated with postoperative acute kidney injury, myocardial injury, stroke, and 30-day mortality.¹⁻³ Machine learning algorithms that predict impending hypotension have been developed to support hemodynamic management. The Hypotension Prediction Index (HPI; Edwards Lifesciences) is a US Food and Drug Administration-cleared algorithm that derives features from the arterial pressure waveform to estimate the probability of a hypotensive event within 5 to 15 minutes.⁴ In the HYPE randomized trial, HPI-guided care reduced the depth and duration of IOH.⁵ However, recent evidence questions whether HPI outperforms simpler alternatives. Mulder and colleagues found that HPI did not outperform a simple mean arterial pressure (MAP) threshold for predicting IOH during noncardiac surgery, using a single outcome definition.⁶

An underappreciated source of variability is the absence of a universal definition of IOH. Most studies use absolute MAP thresholds of <65 or <55 mm Hg or a relative decrease from patient baseline for event durations of 1 to 5 minutes.²˒⁷⁻⁹ Because these definitions capture different physiologic events and occur at different frequencies, a model that performs well under one definition need not perform well under another. A recent evaluation of the Epic sepsis model showed that performance varied substantially across three sepsis definitions,¹⁰ illustrating how sensitive a fixed algorithm can be to the label it is judged against. An analogous multi-definition evaluation has not been performed for hypotension prediction. Moreover, evaluations of these algorithms have emphasized discrimination (area under the receiver operating characteristic curve [AUROC]) over other deployment-relevant metrics, i.e. calibration, decision-curve net benefit, alert burden, and lead time, that determine whether an alert is clinically usable.

We performed an independent comparative evaluation of an open-source convolutional neural network (CNN) for hypotension prediction against two simple rule-based comparators across four commonly used IOH definitions. The two comparators were a MAP threshold rule and a three-feature trend-based logistic regression model (hereafter "Trend LR"), each described in detail in the Methods. Our primary objective was to characterize whether a complex machine learning approach meaningfully outperforms these simple alternatives across multiple definitions, and whether any observed advantage persists when evaluated against deployment-relevant metrics. We considered a difference "meaningful" only if it was both statistically supported (non-overlapping 95% confidence intervals for AUROC) and large enough to change deployment behavior (for example, a clinically relevant change in net benefit, alert burden, or lead time); nominal differences in AUROC that did not meet both conditions were treated as equivalent.

Methods

Study Design and Data Source
This diagnostic study compared a CNN against two comparators across four outcome definitions in an independent surgical cohort, following the TRIPOD+AI reporting guideline.¹¹ TRIPOD+AI (Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis, extended for artificial intelligence) is a consensus checklist that specifies how prediction-model studies should report their data sources, model development, and evaluation so that results can be appraised and reproduced. Briefly, we report the cohort derivation, the exact outcome definitions, the model specifications, and the full set of evaluation metrics in accordance with that guideline. We used VitalDB, a publicly available database that contains high-resolution intraoperative biosignals from adults who received general anesthesia at a tertiary academic center between 2016 and 2017.¹² Arterial waveforms recorded at up to 500 Hz were downsampled to 100 Hz, a rate that preserves the beat-to-beat MAP information used by all three models while reducing computational and storage burden. The overall study design is summarized in Figure 1.

Ethical Considerations
VitalDB comprises fully de-identified records that are freely available for research. The University of Virginia Human Research Participant Program reviewed this project and determined that it does not meet the definition of research with human subjects or a clinical investigation and is therefore not subject to institutional review board oversight, in accordance with 45 CFR 46.102 (Non-Human Subject Determination, Project No. 864; July 1, 2026). No additional informed consent was required beyond that obtained by the original data collectors.

Cohort and Signal Processing
We included adults (≥18 years) undergoing noncardiac surgery with continuous invasive arterial monitoring, because invasive monitoring provides the beat-to-beat pressure signal on which waveform- and MAP-based prediction depends. Cases were excluded if fewer than 70% of waveform samples fell within 30 to 200 mm Hg, if recording duration was under 20 minutes, or if median MAP fell outside 40 to 130 mm Hg. These criteria were chosen to remove recordings dominated by artifact, transducer disconnection, or non-physiologic values that would otherwise corrupt event labeling: pressures outside 30 to 200 mm Hg are physiologically implausible in this setting, very short recordings capture too little of the operative course to yield meaningful prediction windows, and an implausible median MAP signals a miscalibrated or malfunctioning transducer. From eligible cases, 500 were randomly sampled using a fixed seed. Values outside 30 to 200 mm Hg were treated as artifact; beat-to-beat MAP was approximated with a 10-second rolling mean.

Outcome Definitions
We evaluated four definitions of IOH drawn from the perioperative literature. Each captures a clinically recognized pattern of low blood pressure, and they differ in the pressure threshold, the required duration, and whether the threshold is absolute or relative to the patient’s own baseline:
1. MAP <65 mm Hg for ≥1 minute (the HYPE/HPI definition).¹˒⁵
2. MAP <65 mm Hg for ≥5 minutes.¹˒⁸
3. MAP <55 mm Hg for ≥1 minute.²˒⁸
4. ≥20% relative decrease from baseline MAP for ≥1 minute.²˒⁷
Definitions 1 and 2 share the same 65 mm Hg threshold but differ in the required duration (1 vs 5 minutes); we retained both because brief and sustained hypotension carry different physiologic and prognostic implications and are used interchangeably across studies. Baseline MAP was taken to be the median of valid samples between 5 and 10 minutes after recording onset.

Windowing
We extracted nonoverlapping 30-second windows and labeled whether a qualifying event began 5 to 15 minutes after each window. Windows during ongoing events were censored to ensure prospective rather than concurrent prediction. Windows with more than 50% missing data were excluded.

Prediction Models
We compared three models, each trained or specified separately for every definition. In brief, the models span a range of complexity: a one-line threshold rule, a three-feature regression, and a neural network.

Open-source CNN. The CNN was a 1-dimensional network with four convolutional blocks (16, 32, 64, and 128 channels, with batch normalization and rectified linear unit activation), global average pooling, 30% dropout, and a fully connected output layer (61,953 parameters). It was trained on normalized 30-second MAP windows with weighted binary cross-entropy loss, the Adam optimizer, and early stopping on validation AUROC. A separate model was trained for each definition on a single consumer-grade graphics processing unit.

MAP threshold. The first comparator ("MAP threshold") was a one-line rule mapping the current MAP to a risk score via a sigmoid transformation, representing the simplest possible predictor.

Trend LR. The second comparator ("Trend LR") was a logistic regression using three interpretable features: mean MAP within the window, the slope of MAP over the preceding 3 minutes, and the deviation of current MAP from the patient’s baseline. This model was designed to capture short-term hemodynamic trajectory and, uniquely among the three, the patient’s individual baseline.

Data Splits and Statistical Analysis
Cases were split at the patient level into training (60%; n=300), validation (20%; n=100), and test (20%; n=100) sets, preserving splits across all window-level analyses to prevent data leakage. Discrimination was assessed by AUROC with 95% confidence intervals from 200 patient-level bootstrap resamples. Calibration was assessed by expected calibration error (ECE; 10 equal-width bins) and Brier score before and after post-hoc Platt scaling fit on validation-set predictions.¹⁴ Net benefit was computed across decision thresholds of 0.01 to 0.50 following Vickers and Elkin.¹³ At recalibrated thresholds of 0.05, 0.10, and 0.20, we computed alert burden (alerts per hour), positive predictive value (PPV), sensitivity, number needed to alert (NNA = 1/PPV), and lead time (interval from window end to event onset). Analyses used Python 3.14 with pandas, NumPy, scikit-learn, PyTorch, and matplotlib.

Results

Cohort Characteristics
Of 500 cases, 91% met quality criteria; median case duration was 226 minutes (interquartile range [IQR], 173–301). Among 209,213 windows, event prevalence was 13.9% for definition 1 (MAP<65/1min), 5.0% for definition 2 (MAP<65/5min), 3.4% for definition 3 (MAP<55/1min), and 15.1% for definition 4 (relative decrease from baseline). In short, the two clinically common definitions (1 and 4) were roughly three to four times as prevalent as the stricter definitions (2 and 3). 

Discrimination and Calibration
Across the three absolute definitions, all three models produced overlapping AUROC values (0.711–0.747), with the CNN and MAP threshold nearly indistinguishable (differences <0.01) (Table 1; Figure 2). For the relative-decrease definition, performance diverged: the MAP threshold performed worse than chance (0.485; 95% CI, 0.435–0.533), the CNN was near chance (0.543; 95% CI, 0.511–0.576), whereas Trend LR achieved 0.782 (95% CI, 0.740–0.817). All models were substantially miscalibrated before adjustment (ECE 0.212–0.403), meaning the probabilities they emitted did not match observed event rates; mean predicted probabilities exceeded observed rates 3- to 14-fold, consistent with class-weighted training. Platt scaling, a standard post-hoc recalibration step that rescales predicted probabilities using a held-out set, reduced ECE to 0.002–0.017 without changing AUROC, so that the reported probabilities could be interpreted at face value for the decision-curve and alert analyses (Table 1).

Net Benefit
On recalibrated predictions, decision-curve analysis showed near-identical maximum net benefit across models for each of the IOH definitions (differences <0.001). For definition 1 (MAP <65 for ≥1 min), maximum net benefit was approximately 0.097 for all three models (Figure 3). For definition 4 (relative decrease from baseline), only Trend LR provided meaningful net benefit across a wide threshold range. In other words, added model complexity did not translate into greater clinical value for the absolute definitions.

Operational Characteristics
At a recalibrated threshold of 0.10, the three models were operationally indistinguishable for definition 1 (MAP<65/1min), with 44 to 48 alerts per hour, PPV of 18.5% to 19.3%, sensitivity of 66% to 69%, NNA of 5.2 to 5.4, and median lead time of 7.4 to 7.5 minutes (Table 2). The CNN’s 61,953 parameters conferred no operational advantage over the one-line MAP threshold model. Regardless, such frequent alert rates would likely be clinically untenable for all models. For definition 4 (relative decrease from baseline), the MAP threshold and CNN produced approximately 118 to 120 alerts per hour at approximately 10% PPV, alerting on nearly every window, whereas the Trend LR model produced 38.6 alerts per hour at 21.9% PPV and 70.9% sensitivity. Lead time distributions were similar across models for the common definitions.

Effect of Definition
For each model, performance varied markedly across definitions. The MAP threshold model ranged from 0.485 to 0.738 in AUROC, the CNN ranged from 0.543 to 0.747, and Trend LR was most stable, ranging from 0.718 to 0.782. Maximum net benefit varied with event prevalence for each IOH definition, from approximately 0.097 for higher-prevalence definitions (i.e., 1 and 4) to 0.014–0.022 for rarer definitions (i.e., 2 and 3). Choice of IOH definition altered performance by an order of magnitude more than the choice of model did for any of the absolute definitions (i.e., 1, 2, or 3).

Discussion

In this independent evaluation using open-access surgical data, an open-source CNN trained on MAP time series did not meaningfully outperform a simple MAP threshold model across three absolute IOH definitions on any of five deployment-relevant metrics (discrimination, calibration, net benefit, alert burden, and lead time). These findings extend the observation of Mulder et al⁶ that the HPI does not outperform a MAP threshold, applying it to a deep learning model evaluated on independent data across multiple definitions and metrics.

Because our CNN was trained on the MAP time series rather than the raw arterial waveform, our results speak directly to MAP-derived prediction and cannot exclude the possibility that raw waveform features carry additional signal.¹⁵ The practical implication is nonetheless notable: when MAP is the input, added model capacity yielded no benefit over a simple threshold. For definition 4 (relative decrease from baseline), both the MAP threshold and the CNN performed at or near chance, while the three-feature Trend LR model, which incorporates patient baseline, achieved an AUROC of 0.78 and the only deployable alert pattern. This gap was driven by a single clinically motivated feature rather than by architecture, illustrating that feature-outcome alignment can dominate model complexity. The relative success of the Trend LR model in predicting IOH under definition 4 is partly structural, as its algorithm explicitly accounts for current deviation from the baseline MAP.

Several findings indicate that AUROC overstates clinical utility. All models were severely miscalibrated before adjustment, with predicted probabilities 3- to 14-fold above observed rates. After recalibration, net benefit was indistinguishable across models for absolute definitions despite nominally different AUROC values, and alert burden and lead time were likewise similar. Even at optimal thresholds, alert rates near one per minute would challenge clinical deployment. These observations argue for routine reporting of calibration, net benefit, alert burden, and lead time alongside AUROC.

Performance varied substantially across definitions for every model, mirroring findings for the Epic sepsis model.¹⁰ Published metrics reported under a single definition may not generalize. Institutions should align a model’s training definition with the clinical concern motivating deployment, or apply definition-specific recalibration. Prior deep learning models trained on richer inputs have reported higher discrimination,¹⁵ reinforcing that input representation drives performance.

We did not adjust for the duration of anesthesia. Longer cases contribute more prediction windows and greater cumulative exposure to hypotension, so case duration could in principle influence both event prevalence and model performance. Median case duration in our cohort was 226 minutes, and windows from all cases were pooled with patient-level splitting to avoid leakage. Because the comparison of interest was between models evaluated on the same windows, differential case duration is unlikely to bias the head-to-head comparison, but future work stratifying performance by case length would clarify whether prediction is harder early or late in long operations.

Limitations
This study has several limitations, several of which point to concrete next steps. First, the proprietary HPI cannot be evaluated directly; we studied an open-source CNN targeting the same task, and our architecture and training population differ from HPI. A prospective head-to-head comparison against the deployed HPI device, where licensing permits, would close this gap. Second, our CNN was trained on smoothed MAP rather than the raw waveform, which likely limits its ceiling and is the principal caveat to generalizing our conclusions to waveform-based algorithms. Retraining on the full arterial waveform would test whether richer inputs recover an advantage for deep learning. Third, VitalDB is a single-institution dataset, and we used a 500-case subsample, yielding wide confidence intervals for rare definitions, where lead time estimates were based on as few as 13 true positives; external multicenter validation with larger samples would tighten these estimates. Fourth, the relative-decrease definition depends on accurate baseline MAP estimation, which may be noisy in patients with limited recorded preoperative monitoring. Incorporating preoperative ward or clinic blood pressures could yield a more robust baseline. Fifth, alert burden and lead time analyses were performed at a single threshold (0.10) on recalibrated probabilities. Sensitivity to threshold selection should be considered when interpreting these results, and prospective work should pre-specify thresholds tied to an acceptable alert rate. Finally, decision curve analysis assumes that the harm of a false positive is constant across thresholds and that all true positives are equally valuable. In practice, the actionability of an alert and the cost of intervention may vary across patients and clinical contexts, which future utility-weighted analyses could capture.

Conclusions

For predicting intraoperative hypotension from mean arterial pressure, a neural network was no better than a one-line pressure threshold across three absolute definitions and every deployment metric we examined. For a relative, percentage-drop definition, a simple three-feature model that included the patient’s own baseline was the only approach that worked. The practical message for clinicians and institutions is that more complex is not automatically better. Rather, better outcomes can be achieved by choosing a model whose inputs matched the specific kind of hypotension being targeted. These models should ultimately be judged with metrics that predict realistic benefit at the bedside, e.g. how often the model alerts, how far in advance, and whether its risk estimates are trustworthy, rather than on discrimination alone. Before adopting any hypotension-prediction tool, teams should confirm that its definition of hypotension matches their clinical concern and that its alert rate and lead time are compatible with real intraoperative workflow.

Data and Code Availability

The VitalDB dataset is publicly available at https://github.com/kmshannon1016/ioh-prediction.