Predicting Intraoperative Hypotension Using Machine Learning: A Comprehensive Analysis of the MOVER Dataset

2. Literature Review

2.1. Intraoperative Hypotension: Definition and Clinical Impact

Intraoperative hypotension lacks a universal definition, with studies using various thresholds including systolic blood pres- sure (SBP) ¡ 80-90 mmHg, mean arterial pressure (MAP) ¡ 60-70 mmHg, or relative decreases from baseline [1]. Despite definitional variability, evidence consistently demonstrates that even brief periods of hypotension increase the risk of adverse outcomes. Walsh et al. [2] found that MAP ¡ 55 mmHg for even 1-5 minutes was associated with increased risk of acute kidney injury and myocardial injury. Salmasi et al. [3] demonstrated that the risk of myocardial and kidney injury increases with both the depth and duration of hypotension.

2.2. Machine Learning in Hypotension Prediction

Recent advances in machine learning have enabled new approaches to hypotension prediction. Hatib et al. [6] devel- oped the Hypotension Prediction Index (HPI) using arterial waveform analysis, achieving an AUC of 0.95 for predicting hypotension 5-15 minutes before onset. However, this ap- proach requires invasive arterial line placement and specialized monitoring.

Several studies have explored non-invasive approaches using routinely monitored vital signs. Lee et al. [7] used gradient boosting machines to predict hypotension during cesarean sections, achieving 83% accuracy. Kendale et al. [8] employed neural networks for intraoperative hypotension prediction with moderate success. However, these studies were limited by relatively small sample sizes and single-center data.

2.3. Deep Learning for Time Series Prediction

Recurrent neural networks, particularly Long Short-Term Memory (LSTM) networks, have shown promise for physi- ological time series prediction. Chen et al. [9] used LSTMs to predict hypotension in intensive care unit patients, while Sadeghi et al. [10] employed attention-based models for intra- operative blood pressure forecasting. These approaches capture temporal dependencies but require substantial training data and computational resources.

2.4. The MOVER Dataset

The MOVER dataset represents a significant advancement in publicly available intraoperative data. Unlike previous datasets that often lacked temporal alignment between vitals and events or had limited sample sizes, MOVER provides synchronized, high-frequency measurements from multiple institutions [5]. This enables robust model development and external valida- tion, addressing key limitations of prior research.

2.5. Research Gap

Despite progress in hypotension prediction, several gaps remain: (1) limited external validation of models across diverse patient populations and institutions; (2) insufficient compari- son of different algorithmic approaches on the same dataset; (3) lack of interpretability analysis to understand prediction drivers; and (4) limited evaluation of real-time implementation feasibility. This study addresses these gaps through compre- hensive analysis of the MOVER dataset.

3.1. Dataset Description

The MOVER dataset [5] contains intraoperative data col- lected from surgical procedures at multiple medical centers. The dataset includes:

• Patient Demographics: Age, sex, height, weight, ASA physical status classification
• Vital Signs: Heart rate (HR), systolic blood pressure (SBP), diastolic blood pressure (DBP), mean arterial pressure (MAP), oxygen saturation (SpO2), respiratory rate (RR), end-tidal CO2 (EtCO2), temperature
• Ventilation Parameters: Tidal volume, peak inspiratory pressure, positive end-expiratory pressure (PEEP), minute ventilation
• Medication Administration: Timestamps and doses of anesthetics, vasopressors, and other medications
• Clinical Events: Hypotension, hypertension, tachycardia, bradycardia, desaturation events with timestamps
• Surgical Information: Procedure type, duration, urgency status

Vital signs are recorded at 1-minute intervals, with some parameters (e.g., arterial waveform) available at higher fre- quencies. Table 1 summarizes the dataset characteristics.

3.2. Data Preprocessing

1)

Data Cleaning: The raw MOVER dataset underwent extensive cleaning procedures:

1)

Missing Value Handling: Vital sign measurements with missing values were handled using forward-fill for gaps

≤ 5 minutes and linear interpolation for longer gaps.

Records with ¿30% missing data were excluded.

2)

Outlier Detection and Removal: Physiological implau- sible values were removed using clinically validated thresholds: HR (20-200 bpm), SBP (40-250 mmHg), DBP (20-150 mmHg), MAP (30-180 mmHg), SpO2 (50-100%), RR (4-40 breaths/min).

3)

Artifact Removal: Motion artifacts and measurement errors were identified using median filtering and local outlier factor algorithms.

4)

Temporal Alignment: All vital sign measurements were aligned to uniform 1-minute intervals with timestamps synchronized to procedure start time.

2)

Event Definition and Labeling: Hypotension was defined as MAP ¡ 65 mmHg for at least 1 minute, consistent with recent clinical guidelines [4]. For each hypotensive event, prediction windows were created at 5, 10, and 15 minutes before onset. Non-hypotensive periods were sampled from time points at least 30 minutes away from any hypotensive event to ensure clean negative samples.

3.3. Feature Engineering

I engineered a comprehensive set of features across multiple domains:

1)

Statistical Features: For each vital sign, rolling statistics were calculated over windows of 5, 10, 15, and 30 minutes:

• Mean, median, standard deviation
• Minimum, maximum, range

3.4. Machine Learning Models

I implemented and compared five categories of machine learning algorithms.

1)

Tree-Based Ensemble Methods:XGBoost (Extreme Gradient Boosting): XGBoost implements gradient boosted decision trees with regularization to prevent overfitting. The objective function is:

• Skewness, kurtosis
• 25th, 50th, 75th percentiles
• Rate of change (first derivative)

L(ϕ) =

Σ l(yˆ_i, y_i) +

i

Σ Ω(f_k)

(1)

k

• Acceleration (second derivative)

2)

Temporal Features:

• Time since procedure start
• Time since last medication administration
• Time since last clinical event
• Cumulative duration of hypotension in last hour
• Trends using linear regression slopes over multiple win- dows

3)

Interaction Features:

• Heart rate - blood pressure product (rate-pressure prod- uct)
• Shock index (HR/SBP)
• Pulse pressure (SBP - DBP)
• Mean arterial pressure variation
• Oxygen delivery index (MAP × SpO2)

4)

Contextual Features:

• Patient demographics (age, sex, ASA class)
• Surgical procedure type and duration
• Anesthetic agents used
• Vasopressor administration history
• Fluid administration rates

Table 2 presents the top 20 most important features identi- fied through mutual information analysis.

where l is the loss function, yˆ_iis prediction, y_i is true label, and Ω(f_k) = γT + ¹ λ||w||² regularizes tree complexity.

Hyperparameters: learning rate=0.05, max depth=6, n estimators=200, subsample=0.8, colsample bytree=0.8.

Random Forest: An ensemble of decision trees trained on bootstrapped samples with random feature selection. I used 300 trees with minimum samples split=10 and max features=’sqrt’.

2)

Gradient Boosting Machines:Histogram-based Gra- dient Boosting (HGB): Optimized implementation that bins continuous features to improve computational efficiency. Pa- rameters: max iter=200, learning rate=0.1, max depth=5.

3)

Support Vector Machines:SVM with RBF Kernel: Maps input features to high-dimensional space for nonlinear classification. I used C=1.0, gamma=’scale’, and class weight balancing.

4)

Neural Networks:Long Short-Term Memory (LSTM) Networks: Designed to capture temporal dependencies in vital sign sequences. Architecture:

• Input layer: sequence length = 60 minutes
• LSTM layer 1: 128 units, return sequences=True
• Dropout: 0.3
• LSTM layer 2: 64 units
• Dropout: 0.3
• Dense layer: 32 units, ReLU activation
• Output layer: 1 unit, sigmoid activation

Multilayer Perceptron (MLP): Three hidden layers with 256, 128, and 64 neurons, ReLU activation, dropout=0.2, batch normalization.

5)

K-Nearest Neighbors: KNN with k=15 after empirical optimization, using Euclidean distance and uniform weights.

3.5. Experimental Setup

1)

Train-Test Split: The dataset was partitioned as follows:

• Training set (70%): 3,739 procedures
• Validation set (15%): 801 procedures
• Test set (15%): 802 procedures

Stratification ensured similar distributions of hypotension events across splits. No patients appeared in multiple splits.

2)

Cross-Validation Strategy: I employed 5-fold cross- validation on the training set for hyperparameter tuning. Time- series cross-validation with expanding windows was used to prevent look-ahead bias.

3)

Class Imbalance Handling: The dataset exhibited im- balance between hypotensive and non-hypotensive periods. I addressed this using:

• SMOTE (Synthetic Minority Over-sampling Technique)
• Class weight adjustment (balanced mode)
• Undersampling of majority class in training

4)

Evaluation Metrics: Models were evaluated using:

• TP +TN TP +TN +FP +FN
• Precision: ^{T P}

4)

Patient factors: ASA status and age remained impor- tant, suggesting baseline vulnerability influences hy- potension risk.

5)

Temporal patterns: Time since last vasopressor ad- ministration captured medication effects and washout periods.

3.6. Error Analysis

I analyzed misclassified cases to understand model limitations:

TP +FP

TP

• Recall (Sensitivity): TP +FN
• Specificity: ^{T N}
• Precision×Recall Precision+Recall
• AUC-ROC: Area under Receiver Operating Characteristic

curve

• AUC-PR: Area under Precision-Recall curve
• Time-to-event prediction accuracy

3.7. Implementation Details

All models were implemented in Python 3.9 using scikit- learn 1.2, XGBoost 1.7, TensorFlow 2.11, and PyTorch 2.0. Experiments were conducted on an NVIDIA A100 GPU with 40GB memory. Hyperparameter optimization used Optuna with 100 trials per model.

4. Results

4.1. Model Performance Comparison

Table 3 presents comprehensive performance metrics for all models at 5-minute prediction windows.

4.2. Prediction Window Analysis

As expected, performance decreased with longer prediction windows:

• 5-minute: AUC-ROC = 0.973 ± 0.005
• 10-minute: AUC-ROC = 0.942 ± 0.008
• 15-minute: AUC-ROC = 0.908 ± 0.011

4.3. Confusion Matrix Analysis

XGBoost demonstrated the best balance between false pos- itives and false negatives:

• True Positives: 1,847
• True Negatives: 1,891
• False Positives: 119
• False Negatives: 107

4.4. Feature Importance Analysis

Key findings from feature importance analysis:

1)

MAP trends: Recent MAP trends (last 5-10 minutes) were the strongest predictors, highlighting the impor- tance of trajectory rather than absolute values.

2)

Heart rate variability: Increased variability often pre- ceded hypotensive events, possibly reflecting autonomic instability.

3)

Shock index: The combination of heart rate and blood pressure proved more predictive than either parameter alone.

1)

Borderline MAP values (65-70 mmHg): 34% of false positives occurred when MAP was 65-70 mmHg but not meeting hypotension threshold.

2)

Rapid hemodynamic changes: 28% of false negatives involved sudden hypotension development (¡2 minutes) that prediction windows missed.

3)

Surgical manipulation: 22% of errors coincided with major surgical events (aortic cross-clamping, rapid blood loss) not captured in features.

4)

Medication effects: 16% of errors occurred shortly after vasoactive drug administration with unpredictable responses.

4.5. Computational Performance

For real-time deployment, computational efficiency is cru- cial:

• XGBoost inference time: 4.2 ± 0.8 ms per prediction
• LSTM inference time: 28.5 ± 3.2 ms per prediction
• Feature engineering: 12.3 ± 1.5 ms per window
• Total pipeline: ¡50 ms, suitable for 1-minute prediction cycles

5.1. Principal Findings

This study demonstrates that machine learning models trained on the MOVER dataset can predict intraoperative hypotension with high accuracy up to 15 minutes before onset. XGBoost achieved the best performance (AUC-ROC = 0.973 for 5-minute prediction), outperforming more complex deep learning approaches. This finding aligns with prior research showing that gradient-boosted trees often excel on structured tabular data with rich feature engineering [11].

The degradation in performance with longer prediction windows (15-minute AUC-ROC = 0.908) reflects the inherent uncertainty in forecasting hemodynamic events further in ad- vance. However, even 15-minute predictions provide clinically valuable lead time for preventive interventions.

5.2. Comparison with Prior Work

My results compare favorably with previous studies:

• Hatib et al. [6]: AUC-ROC = 0.95 (arterial waveform- based HPI)
• Lee et al. [7]: Accuracy = 83% (cesarean section patients)
• Kendale et al. [8]: AUC-ROC = 0.82 (neural networks)

The superior performance in my study likely reflects: (1) larger, more diverse dataset, (2) comprehensive feature en- gineering, (3) rigorous cross-validation, and (4) inclusion of contextual patient and procedural factors.

5.3. Clinical Implications

The developed models have several potential clinical appli- cations:

1)

Early Warning Systems: Real-time alerts could notify anesthesiologists of impending hypotension, enabling proactive interventions such as fluid boluses, vasopressor administration, or adjustment of anesthetic depth.

2)

Risk Stratification: Preoperative risk assessment could identify high-risk patients for enhanced monitoring and preventive strategies.

3)

Treatment Guidance: Models could help optimize va- sopressor timing and dosing by predicting response to interventions.

4)

Quality Improvement: Aggregate analysis of predicted vs. actual events could identify systematic issues in hemodynamic management.

5.4. Interpretability and Clinical Trust

Feature importance analysis revealed clinically interpretable patterns, enhancing trust in model predictions. The prominence of MAP trends aligns with clinical intuition that trajectory matters. The predictive value of heart rate variability and shock index reflects underlying physiological mechanisms of hemodynamic decompensation.

5.5. Limitations

Several limitations warrant consideration:

1)

Single hypotension definition: Using MAP ¡ 65 mmHg may not capture all clinically relevant hypotensive events. Alternative definitions (e.g., relative decreases from baseline) might yield different results.

2)

Data quality variability: Despite preprocessing, some measurement artifacts may persist, potentially affecting model performance.

3)

Generalizability: While MOVER includes multiple cen- ters, all are academic medical centers; performance in community hospitals requires validation.

4)

Intervention confounding: The dataset includes real- world clinical interventions (vasopressors, fluids) that alter the natural history of hypotension, potentially in- troducing confounding.

5)

Binary classification: My approach treats hypotension prediction as binary classification, while continuous risk scoring might provide more nuanced information.

5.6. Future Directions

Based on my findings, I recommend several directions for future research:

1)

Multi-center prospective validation: Deploy models in prospective studies across diverse clinical settings to assess real-world performance.

2)

Integration with electronic health records: Develop pipelines for real-time data extraction and model deploy- ment within existing clinical workflows.

3)

Personalized prediction: Explore patient-specific model adaptation to account for individual hemodynamic responses.

4)

Causal modeling: Investigate causal relationships be- tween features and hypotension to guide interventions.

5)

Multi-task learning: Simultaneously predict multiple outcomes (hypotension, hypertension, desaturation) for comprehensive risk assessment.

6)

Explainable AI: Develop enhanced interpretability methods to provide actionable insights for clinicians.