Early prediction of intraoperative hypotension: development and validation of the HypoBridCast hybrid deep learning model

Abstract

Background

Intraoperative hypotension (IOH) is a frequent and clinically important complication associated with adverse postoperative outcomes. Early prediction may facilitate timely intervention, although existing models have limitations in integrating multimodal physiological data and patient-specific characteristics.

Methods

We developed Hypotension Hybrid Forecast (HypoBridCast), a hybrid architecture integrating 1D Convolutional Layer (Conv1D) and Transformer (TF) modules. The model integrates intraoperative waveform data (arterial pressure, electrocardiogram, photoplethysmogram, and capnography) with preoperative clinical variables (including demographic characteristics, medical history, laboratory results, and surgical and anesthetic variables). Data from the VitalDB database (n = 3,369) were used for model development and internal evaluation, and an independent cohort from Zhongda Hospital (n = 437) was used for external evaluation. Model performance for predicting IOH (MAP ≤ 65 mmHg for ≥ 1 min) 5, 10, and 15 min in advance was assessed using AUROC, AUPRC, and calibration metrics.

Results

HypoBridCast achieved strong discriminative performance across prediction horizons in both internal and external evaluations (e.g., internal 5-min AUROC 0.9442 [95% CI 0.9427–0.9456] and AUPRC 0.9387 [0.9376–0.9398]). Compared with Mono-ART models, performance was improved, whereas differences versus multi-channel waveform models were modest. The addition of preoperative variables provided limited and inconsistent gains across datasets. In contrast, performance gains over simple MAP-based baseline models were more pronounced. Calibration was acceptable overall, with some reduction observed in the external cohort, particularly at longer prediction horizons.

Conclusions

The proposed hybrid deep learning framework achieved strong performance for short-term prediction of intraoperative hypotension using routinely collected clinical data. Multimodal integration and preoperative variables provide incremental improvements. Further work is needed to improve generalizability and calibration before clinical deployment.

Trial registration

ChiCTR2500099041

Supplementary Information

The online version contains supplementary material available at 10.1186/s12871-026-03888-8.

Background

Maintaining stable intraoperative blood pressure is fundamental to safe surgical care, ensuring adequate organ perfusion and reducing postoperative complications [1, 2]. Intraoperative hypotension (IOH) presents an immediate and prevalent threat. Substantial evidence shows that both the depth and duration of mean arterial pressure (MAP) below 60–70 mmHg significantly increase the risk of myocardial injury, acute kidney injury, and mortality [3]. Notably, IOH correlates more strongly with 30-day postoperative mortality than hypertension [4].

The dynamic intraoperative environment—shaped by surgical phases, patient variability, and anesthesiologist workload—makes blood pressure fluctuations common and often difficult to manage. Even brief episodes of severe hypotension can lead to organ damage, with risk compounding over time [5]. Timely detection is essential, yet challenging, as anesthesiologists juggle multiple tasks under time pressure and fatigue. This highlights the need for predictive tools that enable proactive IOH management, as both premature and delayed interventions may be suboptimal [5].

Traditional IOH prediction models typically use machine learning (such as logistic regression, [6] naive Bayes, [7] or random forest [8]) on hand-crafted features from arterial waveforms, limiting their capacity to capture complex physiological patterns. While deep learning architectures such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) allow automatic feature learning, [9–11] many existing models analyse only single physiological signals (usually arterial waveforms) and overlook crucial preoperative baseline information (e.g., demographics, comorbidities), thereby restricting personalised risk prediction.

To address these limitations, we propose Hypotension Hybrid Forecast (HypoBridCast), a novel model combining 1D Convolutional Layers (Conv1D) for local pattern extraction [12] with Transformer (TF) architectures to model long-range physiological dependencies [13]. HypoBridCast distinctively integrates analyses of multiple intraoperative waveforms with preoperative baseline characteristics for comprehensive and individualised risk assessment. We hypothesise that HypoBridCast can outperform existing models in predicting hypotension 5, 10, and 15 min in advance. Therefore, this study aims to develop and rigorously validate HypoBridCast as an early IOH warning system, comparing its discrimination, calibration, stability, and generalisability against current mainstream methods through internal and external evaluations.

Materials and methods

This study developed an intraoperative hypotension prediction model for non-cardiac surgery patients using a Conv1D-TF deep learning architecture (Fig. 1), adhering to TRIPOD-AI guidelines [14]. The research received ethical approval from Zhongda Hospital, Southeast University (No. 2024ZDSYLL476-P01) and was registered with the Chinese Clinical Trial Registry (ChiCTR2500099041). As a retrospective study, informed consent was waived, and there was no patient or public involvement in the research.

Fig. 1.

Overall workflow of the Study. Data Collection: The development dataset comes from Vital DB (Seoul National University Hospital, Korea), and the external evaluation dataset is from Zhongda Hospital, Southeast University. The data is divided into training set (60%), validation set (10%), and evaluation set (30%). Data Preprocessing: The waveform data is cleaned (removal of abnormal heart rate, MAP, PPG, EtCO₂) and downsampled to 100 Hz. Static data is processed with Z-score normalization and One-Hot encoding, with missing values imputed. Model Development: Features are passed to the Conv1D-Transformer model for further processing. The model outputs a binary classification for predicting intraoperative hypotension. Model Performance: The model's performance is evaluated using ROC curves, Precision-Recall curves (PR curves), and calibration curves, with SHAP used for interpretability analysis. Visualization: The final interface displays realtime intraoperative waveform data and the model's prediction results for each patient

Study population and data sources

The development dataset utilised the publicly available Vital DB, [15] containing diverse data including intraoperative waveforms from 6,388 non-cardiac surgery patients in South Korea (June 2016–August 2017). An external evaluation dataset from Zhongda Hospital, Southeast University, comprising non-cardiac surgery patients (March 20, 2024–August 9, 2024), was used to assess model generalisability.

Patients and outcome definitions

Adults (≥ 18 years) with ASA physical status I–IV undergoing non-cardiac surgery under general anaesthesia with invasive arterial pressure monitoring were included. Patients without complete or available intraoperative waveform data were excluded.

The primary outcome was intraoperative hypotension, defined as a MAP ≤ 65 mmHg sustained for ≥ 1 min. Non-hypotension was defined as MAP > 65 mmHg for at least 20 min. [11] Time segments not meeting either definition were excluded. No exclusion was applied based on MAP at the start of the prediction window (t = 0).

Included parameters

Data for model development and internal evaluation were obtained from the VitalDB dataset, and external evaluation data from Zhongda Hospital. Parameters included preoperative static features and intraoperative dynamic waveform signals.

Preoperative static features (Table 1) were selected to capture individual variability across four domains: (1) demographics (age, sex, body mass index); (2) medical history (e.g., hypertension, diabetes, arrhythmia); (3) preoperative laboratory tests (e.g., haemoglobin, albumin, creatinine); and (4) surgical and anaesthetic characteristics (ASA classification, emergency status, surgical approach, and position). Missing rates and value ranges are provided in Table S1.

Table 1.

Patients characteristics in the development (Vital DB) and external evaluation (ZDH) set

Characteristics	cases, No. (%)
	Development set (Vital DB)			External Evaluation set (ZDH) (n = 437)
	Training set (n = 2021, 60%)	Validation set (n = 337, 10%)	Internal Evaluation set (n = 1011, 30%)
Preoperative characteristics
Age, Median (IQR), y	60.0 (50.0—69.0)	62.0 (53.0—71.0)	61.0 (50.0—69.0)	65.0 (55.0—73.0)
Sex
Male, n (%)	1,116 (55.2)	200 (59.3)	549 (54.3)	224 (51.3)
Female, n (%)	905 (44.8)	137 (40.7)	462 (45.7)	213 (48.7)
BMI, Median (IQR)	23.0 (20.7—25.2)	22.9 (21.1—25.1)	22.8 (20.8—25.2)	23.9 (21.5—26.3)
Preoperative comorbidities
Hypertension, n (%)	672 (33.3)	112 (33.2)	351 (34.7)	185 (42.3)
Diabetes, n (%)	227 (11.2)	48 (14.2)	126 (12.5)	79 (18.1)
Arrhythmia, n (%)	36 (1.8)	3 (0.9)	11 (1.1)	38 (8.7)
Preoperative laboratory results
Hemoglobin, Median (IQR), g/dL	12.9 (11.6—14.2)	13.0 (11.4—14.1)	12.9 (11.5—14.1)	12.7 (11.6—13.9)
Platelet count, Median (IQR),/nL	226.0 (179.0—278.0)	233.0 (182.0—285.0)	235.0 (193.0—280.0)	189.0 (155.0—233.0)
Albumin, Median (IQR), g/dL	4.1 (3.8—4.4)	4.1 (3.8—4.4)	4.1 (3.8—4.4)	4.1 (3.8—4.3)
Sodium, Median (IQR), mmol/L	140.0 (139.0—142.0)	141.0 (139.0—142.0)	140.0 (139.0—142.0)	141.5 (139.4—143.2)
Potassium, Median (IQR), mmol/L	4.2 (3.9—4.4)	4.2 (4.0—4.4)	4.2 (4.0—4.4)	3.9 (3.7—4.1)
Glucose, Median (IQR), mg/dL	105.0 (94.0—127.0)	105.0 (95.0—128.0)	104.0 (94.0—128.0)	90.7 (80.6—111.2)
Alanine transaminase, Median (IQR), IU/L	18.0 (13.0—27.0)	18.0 (12.0—26.0)	18.0 (13.0—27.0)	17.0 (12.0—27.0)
Aspartate transaminase, Median (IQR), IU/L	21.0 (17.0—26.0)	20.0 (17.0—26.0)	20.0 (17.0—26.0)	20.0 (16.0—27.0)
Activated partial thromboplastin time, Median (IQR), sec	32.1 (30.0—34.5)	32.0 (29.8—34.4)	32.2 (30.0—34.4)	30.0 (28.1—31.9)
Blood urea nitrogen, Median (IQR), mg/dL	14.0 (11.0—18.0)	15.0 (12.0—18.0)	14.0 (11.0—18.0)	16.2 (13.2—19.6)
Creatinine, Median (IQR), mg/dL	0.8 (0.7—1.0)	0.8 (0.7—1.0)	0.8 (0.7—1.0)	0.8 (0.7—1.0)
Surgery information
ASA grade, n (%)
Ⅰ	457 (22.6)	72 (21.3)	217 (21.5)	8 (1.8)
Ⅱ	1,237 (61.2)	221 (65.6)	644 (63.7)	309 (70.7)
Ⅲ	301 (14.9)	40 (11.9)	141(13.9)	107 (24.5)
Ⅳ	26 (1.3)	4 (1.2)	9 (0.9)	13 (3.0)
Emergency operation, n (%)	283 (14.0)	40 (11.9)	118 (11.7)	69 (15.8)
Supine position, n (%)	1,047 (51.8)	171 (50.7)	495 (49.0)	197 (45.1)
Surgery approach, n (%)
Open	852 (42.2)	143 (42.4)	407 (40.3)	314 (71.9)
Videoscopic	1,086 (53.7)	178 (52.8)	570 (56.4)	101 (23.1)
Robotic	83 (4.1)	16 (4.7)	34 (3.4)	22 (5.0)
Postoperative information
Duration of surgery, Median (IQR), min	145.0 (90.0—215.0)	150.0 (95.0—230.0)	145.0 (90.0—218.0)	134.8 (95.8—190.0)
Postop in-hospital stay, Median (IQR), day	6.2 (3.4—8.5)	6.2 (2.6—8.6)	5.6 (3.4—9.3)	5.9 (4.0—9.8)
Postop ICU stay, Median (IQR), day	0.0 (0.0—1.0)	0.0 (0.0—1.0)	0.0 (0.0—1.0)	0.0 (0.0—0.0)
In-hospital mortality, n (%)	22 (1.1)	3 (0.9)	11 (1.1)	0 (0.0)
Dataset composition
Hypotensive event, segments (%)	6357 (41.31)	983 (37.53)	3147 (39.24)	542 (23.92)
per patient, segments (SD)	3.37 (5.24)	3.09 (4.89)	3.31 (4.97)	1.23 (2.67)
Non-hypotensive event, segments (%)	9032 (58.69)	1636 (62.47)	4872 (60.76)	1724 (76.08)
per patient, segments (SD)	4.78 (3.68)	5.14 (3.63)	5.12 (3.76)	3.91 (2.35)

This table presents a summary of demographic characteristics, preoperative comorbidities, laboratory results, surgical information, and postoperative details for patients included in the development cohort (Vital DB) and the external evaluation cohort (Zhongda Hospital). The distribution of intraoperative hypotension (IOH) and non-IOH segments was determined based on the event definitions described in the Methods section. Continuous variables are expressed as mean (SD) or median (interquartile range), as appropriate. Categorical variables are reported as counts and percentages

Abbreviations: ZDH  Zhongda Hospital Southeast University, BMI Body mass index, ASA grade American Society of Anesthesiologists grade

Four commonly used intraoperative waveforms were collected: arterial blood pressure (ART), electrocardiogram (ECG, Lead II), photoplethysmogram (PPG), and capnography (Capno), with sampling frequencies ranging from 62.5 to 500 Hz.

For external evaluation, preoperative features were extracted from electronic medical records and laboratory information systems, and intraoperative waveforms were recorded using Vital Recorder software [16]. Data were independently screened and cross-checked by two researchers to ensure consistency and accuracy.

Data preprocessing

Systematic preprocessing was performed to ensure data uniformity, validity, and reproducibility, including unit standardisation, missing data imputation, normalisation, waveform preprocessing, and resampling. Variable units and formats in the external dataset were standardised to match VitalDB (Table S2).

Static data processing

Preoperative variables (Table S2) underwent quality screening prior to modelling. Continuous variables were screened using a Z-score threshold (|Z|> 3) to flag implausible values, which were treated as missing, consistent with standard statistical practice [17]. Variables with < 40% missingness were imputed using mean imputation. Categorical variables underwent one-hot encoding, with ASA classification retained as an ordinal variable. Continuous variables were standardised using Z-score normalisation.

Vital data processing

Vital data were processed using cardiac cycle segmentation via a peak detection algorithm [11]. Segments likely affected by artifacts were excluded following established criteria, [11] including abnormal heart rates (HR < 20 or > 120 bpm) or out-of-range waveform values (e.g., MAP < 20 or > 160 mmHg; PPG < 0 or > 100; EtCO₂ < 0 or > 70 mmHg) (Figure S1A). All waveform data were downsampled to 100 Hz, and 30-s segments (3,000 data points) were used as model input.

Data partitioning and model development

Only patients with complete recordings for all four waveform types (ART, ECG, PPG, Capno) were included. The dataset was randomly split at the patient level into training (60%), validation (10%, for hyperparameter tuning), and internal test (30%) sets prior to waveform segmentation to prevent data leakage.

For real-time prediction, 30-s waveform segments preceding hypotension onset by 5, 10, and 15 min were extracted as positive samples (Figure S1B). This window length captures several cardiac cycles while maintaining a short temporal context for real-time prediction, and has been used in prior studies [18]. Negative samples were randomly selected from stable non-hypotensive periods, excluding windows within 20 min before hypotension onset.

During model development, candidate algorithms were benchmarked using the internal dataset (Table S3). These comparisons were restricted to the internal data, with the external cohort reserved for independent evaluation.

The proposed architecture integrates Conv1D layers for local feature extraction with TF modules to capture long-range temporal dependencies. Waveform features were combined with preoperative variables to generate a binary prediction (1 = hypotension, 0 = non-hypotension; Figure S1C). Further details are provided in Supplementary Method 1 and Figure S2. Four configurations were evaluated: Mono-ART, Mono-ART with preoperative features (Mono-ART with preop), Four-Channel, and Four-Channel with preoperative features (HypoBridCast). Ablation experiments were conducted by removing either the TF module or Conv1D module (Supplementary Method 2).

To assess performance beyond simple MAP-based representations, two baseline models were implemented: a logistic regression model using MAP-derived features (last MAP, variance, slope) and an ARIMA model for short-term MAP trend prediction. Both were evaluated on the same datasets using identical metrics.

Model performance evaluation

Discrimination

Discrimination was assessed using the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC) [19]. Because hypotension events were relatively infrequent, AUPRC was considered particularly informative as it better reflects the model’s ability to detect true events while limiting false positives. These metrics facilitated comparison against other mainstream models and the evaluation of various physiological signal input combinations (single-channel ART versus multi-channel ART with ECG, PPG, or Capno) to investigate their impact on model performance. Optimal classification thresholds, determined using Youden's index, [20] allowed for the calculation of sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) to comprehensively assess performance, particularly on imbalanced datasets.

Calibration

Calibration was evaluated by the agreement between predicted probabilities and observed event rates. This was quantified using calibration slope (ideally approximating 1) and intercept (ideally near 0) to measure the reliability of the model's probability outputs.

Interpretability

For interpretability, Shapley Additive Explanations (SHAP) values [21] were employed. This method analysed input feature importance and quantified individual contributions to model predictions, offering insights into the model's internal mechanisms and providing interpretable decision support.

Finally, Receiver Operating Characteristic (ROC) curves, Precision-Recall (PR) curves, calibration curves, and SHAP importance plots were generated to visualise the model's multi-dimensional performance.

Sample size

Sample size adequacy was confirmed for both cohorts (details in Supplemental Method 3). For model development, estimates based on Riley et al.'s method [22], accounting for conservative parameters and intra-patient clustering, indicated a minimum requirement of approximately 541 patients (equivalent to 4,328 independent observations or 7,362 waveform segments after adjusting for clustering). Our development set of approximately 25,000 segments from 3,369 patients substantially exceeded this requirement.

For external AUROC evaluation, a binomial approximation indicated a minimum of 196 patients, a requirement met by our 437-patient external cohort.

Statistical analysis

Pairwise comparisons of AUROC between models were performed using the DeLong test for correlated ROC curves. Analyses were conducted separately for each prediction horizon (5, 10, and 15 min). Subgroup analyses were performed in both evaluation datasets across age, ASA physical status, and hypertension status.

Patient characteristics were summarised using descriptive statistics. Continuous variables are presented as mean (standard deviation) or median (interquartile range), and categorical variables as counts (proportions). Key evaluation metrics are reported with 95% confidence intervals, estimated using bootstrap resampling (50 iterations) where appropriate. Analyses were performed using Python 3.7.0 and R version 3.6.3.

Real-time demonstration system

To illustrate potential clinical application, a real-time intraoperative hypotension prediction system was developed, integrating monitoring data to continuously generate risk predictions and early warning signals via a visual interface (see Supplementary Method 4 and Figure S5).

Data sharing statement

The deidentified dataset and statistical code are available from the corresponding author upon reasonable request, subject to submission of a research proposal and institutional review board approval.

Results

A total of 3,369 patients with complete ART, ECG, PPG, and Capno waveform data were included for model development and randomly partitioned into training (n = 2021, 60%), validation (n = 337, 10%), and internal test (n = 1011, 30%) sets. For external evaluation, 887 patients were screened from Zhongda Hospital; after applying eligibility criteria and excluding 72 with incomplete waveform data, 437 patients were included (Fig. 2). Preoperative baseline variables were consistently defined across cohorts (Table 1).

Fig. 2.

Flowchart of patient selection for the development and external evaluation dataset

Prediction performance of hypotension

HypoBridCast demonstrated strong discriminative performance across prediction horizons (5, 10, and 15 min). In the internal evaluation set, the model achieved an AUROC of 0.9442 (95% CI 0.9427–0.9456) and an AUPRC of 0.9358 (95% CI 0.9338–0.9378) for 5-min prediction, with performance gradually decreasing at longer prediction horizons. In the external evaluation set, discrimination was lower but remained acceptable across all prediction horizons (Tables S3–S4).

Multi-waveform inputs yielded higher discrimination than single-channel ART models. However, the incremental gain from adding preoperative variables was limited and inconsistent across datasets and prediction horizons. Pairwise comparisons using the DeLong test showed that HypoBridCast significantly outperformed single-channel ART models across all prediction horizons in the internal dataset (all P < 0.001), indicating superior discrimination. Similar trends were observed in the external dataset at shorter prediction horizons (Table S4). In contrast, differences between HypoBridCast and Four-Channe model were small and not statistically significant in either dataset (P > 0.05), suggesting limited incremental benefit from preoperative variables. In the external dataset, overall performance declined and differences between model configurations were attenuated, particularly at longer prediction horizons (Fig. 3).

Fig. 3.

ROC and PR Curves for Mono-ART and Four-Channel Models at Different Prediction Horizons. These curves illustrate the performance of different models in predicting intraoperative hypotension 5, 10, and 15 min prior to onset. (A) and (B) display the ROC curves, illustrating discrimination performance for the internal and external evaluation sets, respectively. (C) and (D) display the PR curves, which are particularly relevant for evaluating performance on imbalanced datasets, for the internal and external evaluation sets, respectively. Across all plots, the lines represent the following models: HypoBridCast (Four-Channel with Preop; orange), Four-Channel (Waveforms Only; red), Mono-ART + Preop (blue), and Mono-ART (Waveforms Only; green). Area Under the Curve (AUC for ROC in A and B; AUPRC for PR in C and D) values with 95% confidence intervals are provided within each plot's legend

Comparisons of waveform combinations showed that models incorporating multiple physiological signals generally achieved higher point estimates, although no single combination consistently outperformed others across all settings (Table S5). Combinations including PPG and/or ECG tended to show higher point estimates than those adding only Capno.

Ablation experiments demonstrated that removal of either the TF or Conv1D module resulted in reduced AUROC across prediction horizons in both Mono-ART and Four-Channel model, supporting the complementary roles of local feature extraction and temporal modeling (Supplementary Method 2).

At optimal thresholds (Youden’s index), HypoBridCast achieved high positive and negative predictive values for 5-min prediction in both internal (PPV 0.8944, NPV 0.8764) and external datasets (PPV 0.8051, NPV 0.9223) (Table 2). Sensitivity and NPV decreased with longer prediction horizons, whereas PPV remained relatively stable.

Table 2.

Optimal cut-off values were determined using Youden’s index

Model	Value (95% confidence interval)
	Internal Evaluation Set			External Evaluation Set
	5 min prediction	10 min prediction	15 min prediction	5 min prediction	10 min prediction	15 min prediction
Arterial-pressure-only model
Threshold	0.7402	0.6988	0.6834	0.7402	0.6988	0.6834
Sensitivity (%)	0.8523 (0.8503—0.8544)	0.6124 (0.6082—0.6166)	0.5917 (0.5869—0.5964)	0.6136 (0.5919—0.6353)	0.6050 (0.5880—0.6221)	0.5918 (0.5788—0.6047)
Specificity (%)	0.8827 (0.8794—0.8860)	0.9938 (0.9935—0.9941)	0.9940 (0.9932—0.9947)	0.9372 (0.9298—0.9446)	0.9551 (0.9483—0.9620)	0.9243 (0.9151—0.9334)
PPV (%)	0.8238 (0.8194—0.8281)	0.9841 (0.9832—0.9849)	0.9834 (0.9814—0.9855)	0.7107 (0.6785—0.7429)	0.7751 (0.7458—0.8045)	0.6419 (0.6213—0.6625)
NPV (%)	0.9029 (0.9009—0.9048)	0.8047 (0.8021—0.8073)	0.8020 (0.7988—0.8053)	0.9066 (0.9030—0.9102)	0.9049 (0.8980—0.9117)	0.9081 (0.9028—0.9135)
ART + Capno model
Threshold	0.7186	0.6646	0.6693	0.7186	0.6646	0.6693
Sensitivity (%)	0.6806 (0.6755—0.6857)	0.5395 (0.5364—0.5427)	0.4960 (0.4900—0.5020)	0.6878 (0.6587—0.7170)	0.7192 (0.6997—0.7387)	0.6270 (0.6030—0.6510)
Specificity (%)	0.9771 (0.9759—0.9784)	0.9975 (0.9972—0.9978)	0.9973 (0.9968—0.9978)	0.9020 (0.8972—0.9067)	0.8879 (0.8793—0.8965)	0.8617 (0.8516—0.8717)
PPV (%)	0.9505 (0.9474—0.9537)	0.9929 (0.9921—0.9937)	0.9909 (0.9893—0.9925)	0.6346 (0.6159—0.6533)	0.6073 (0.5916—0.6231)	0.5056 (0.4854—0.5258)
NPV (%)	0.8256 (0.8224—0.8288)	0.7700 (0.7679—0.7720)	0.7695 (0.7660—0.7731)	0.9209 (0.9131—0.9288)	0.9296 (0.9248—0.9344)	0.9114 (0.9062—0.9167)
ART + ECG model
Threshold	0.7344	0.6950	0.6729	0.7344	0.6950	0.6729
Sensitivity (%)	0.7457 (0.7421—0.7492)	0.7474 (0.7411—0.7538)	0.6949 (0.6901—0.6997)	0.4526 (0.4244—0.4808)	0.6301 (0.6130—0.6472)	0.5453 (0.5242—0.5664)
Specificity (%)	0.9680 (0.9668—0.9692)	0.9405 (0.9390—0.9421)	0.9588 (0.9574—0.9603)	0.9943 (0.9927—0.9958)	0.9495 (0.9463—0.9528)	0.9726 (0.9691—0.9761)
PPV (%)	0.9376 (0.9353—0.9399)	0.8912 (0.8885—0.8939)	0.9129 (0.9103—0.9154)	0.9522 (0.9400—0.9644)	0.7510 (0.7345—0.7674)	0.8211 (0.7980—0.8441)
NPV (%)	0.8552 (0.8534—0.8569)	0.8510 (0.8464—0.8555)	0.8351 (0.8329—0.8373)	0.8777 (0.8699—0.8856)	0.9141 (0.9108—0.9174)	0.9026 (0.8978—0.9073)
ART + PPG model
Threshold	0.7375	0.6908	0.6794	0.7375	0.6908	0.6794
Sensitivity (%)	0.7399 (0.7358—0.7439)	0.6718 (0.6668—0.6767)	0.6010 (0.5947—0.6073)	0.7309 (0.7081—0.7537)	0.5655 (0.5445—0.5864)	0.5642 (0.5406—0.5878)
Specificity (%)	0.9747 (0.9732—0.9761)	0.9855 (0.9849—0.9860)	0.9915 (0.9910—0.9920)	0.8612 (0.8546—0.8677)	0.9662 (0.9634—0.9690)	0.9338 (0.9300—0.9377)
PPV (%)	0.9495 (0.9467—0.9523)	0.9684 (0.9672—0.9696)	0.9778 (0.9765—0.9791)	0.5752 (0.5544—0.5960)	0.8050 (0.7939—0.8161)	0.6732 (0.6594—0.6869)
NPV (%)	0.8534 (0.8505—0.8564)	0.8192 (0.8168—0.8216)	0.7993 (0.7971—0.8016)	0.9254 (0.9182—0.9326)	0.9005 (0.8960—0.9050)	0.8984 (0.8910—0.9059)
ART + PPG + Capno model
Threshold	0.7337	0.6906	0.6710	0.7337	0.6906	0.6710
Sensitivity (%)	0.6396 (0.6328—0.6465)	0.6892 (0.6838—0.6945)	0.6336 (0.6279—0.6393)	0.7710 (0.7564—0.7855)	0.5848 (0.5629—0.6068)	0.5579 (0.5375—0.5784)
Specificity (%)	0.9924 (0.9916—0.9933)	0.9833 (0.9820—0.9846)	0.9842 (0.9829—0.9855)	0.8513 (0.8405—0.8621)	0.9631 (0.9581—0.9681)	0.9412 (0.9332—0.9492)
PPV (%)	0.9819 (0.9800—0.9838)	0.9643 (0.9616—0.9670)	0.9616 (0.9585—0.9647)	0.5778 (0.5576—0.5980)	0.7869 (0.7642—0.8096)	0.6820 (0.6405—0.7235)
NPV (%)	0.8111 (0.8078—0.8145)	0.8285 (0.8242—0.8328)	0.8113 (0.8083—0.8143)	0.9336 (0.9272—0.9400)	0.9091 (0.9023—0.9160)	0.9044 (0.9001—0.9088)
ART + ECG + Capno model
Threshold	0.7323	0.6786	0.6659	0.7323	0.6786	0.6659
Sensitivity (%)	0.6288 (0.6250—0.6327)	0.6874 (0.6826—0.6921)	0.4613 (0.4566—0.4660)	0.6728 (0.6420—0.7035)	0.7700 (0.7515—0.7884)	0.6132 (0.5924—0.6339)
Specificity (%)	0.9947 (0.9942—0.9953)	0.9726 (0.9712—0.9741)	0.9975 (0.9971—0.9980)	0.9076 (0.9031—0.9122)	0.8071 (0.7966—0.8177)	0.9271 (0.9209—0.9332)
PPV (%)	0.9874 (0.9860—0.9888)	0.9426 (0.9394—0.9457)	0.9916 (0.9901—0.9930)	0.6508 (0.6299—0.6717)	0.4978 (0.4732—0.5223)	0.6660 (0.6442—0.6878)
NPV (%)	0.8038 (0.8014—0.8063)	0.8264 (0.8243—0.8286)	0.7472 (0.7447—0.7498)	0.9154 (0.9077—0.9231)	0.9340 (0.9283—0.9398)	0.9101 (0.9034—0.9167)
ART + ECG + PPG model
Threshold	0.7398	0.6911	0.6755	0.7398	0.6911	0.6755
Sensitivity (%)	0.7494 (0.7458—0.7530)	0.6630 (0.6581—0.6678)	0.6518 (0.6474—0.6562)	0.7020 (0.6800—0.7240)	0.6753 (0.6539—0.6967)	0.5203 (0.4911—0.5495)
Specificity (%)	0.9658 (0.9644—0.9672)	0.9846 (0.9836—0.9855)	0.9764 (0.9749—0.9778)	0.8995 (0.8923—0.9067)	0.9100 (0.9011—0.9189)	0.9772 (0.9736—0.9808)
PPV (%)	0.9341 (0.9319—0.9364)	0.9656 (0.9640—0.9671)	0.9443 (0.9405—0.9481)	0.6461 (0.6289—0.6633)	0.6569 (0.6384—0.6755)	0.8470 (0.8260—0.8680)
NPV (%)	0.8563 (0.8540—0.8587)	0.8178 (0.8155—0.8200)	0.8199 (0.8169—0.8229)	0.9205 (0.9150—0.9261)	0.9170 (0.9105—0.9235)	0.8940 (0.8867—0.9012)
ART + ECG + PPG + Capno (HypoBridCast) model
Threshold	0.7372	0.6965	0.6881	0.7372	0.6965	0.6881
Sensitivity (%)	0.7963 (0.7925—0.8001)	0.6900 (0.6869—0.6932)	0.6630 (0.6587—0.6673)	0.6851 (0.6648—0.7054)	0.6646 (0.6378—0.6915)	0.5361 (0.5045—0.5678)
Specificity (%)	0.9389 (0.9372—0.9407)	0.9732 (0.9718—0.9745)	0.9757 (0.9746—0.9768)	0.9577 (0.9520—0.9634)	0.9457 (0.9419—0.9496)	0.9510 (0.9469—0.9552)
PPV (%)	0.8944 (0.8916—0.8973)	0.9419 (0.9388—0.9451)	0.9427 (0.9398—0.9456)	0.8051 (0.7781—0.8320)	0.7552 (0.7385—0.7719)	0.7061 (0.6718—0.7403)
NPV (%)	0.8764 (0.8742—0.8786)	0.8328 (0.8304—0.8351)	0.8279 (0.8259—0.8299)	0.9223 (0.9169—0.9277)	0.9176 (0.9093—0.9260)	0.9021 (0.8937—0.9105)

All models presented in this table include preoperative data as part of the input features. The classification performance of these prediction models was evaluated across three prediction horizons (5, 10, and 15 min). Sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were calculated using the optimal threshold determined by Youden’s index. Results are reported separately for the internal evaluation dataset (Vital DB) and the external evaluation dataset (Zhongda Hospital). Values are presented as point estimates with 95% confidence intervals

Abbreviations: NPV negative predictive value, PPV positive predictive value

Compared with simple MAP-based baseline models, HypoBridCast demonstrated substantially improved performance (Table S6), with AUROC increases of approximately 0.06–0.08 over the logistic regression model and 0.11–0.12 over the ARIMA model across prediction horizons.

Calibration performance was acceptable overall (Table 3, Table S7, Figure S3). In the internal dataset, calibration slopes were generally close to 1, and HypoBridCast achieved low Expected Calibration Error (ECE) at several prediction horizons (e.g., 0.077 at 5 min). Calibration declined in the external dataset across all models, with HypoBridCast remaining among the better-calibrated configurations.

Table 3.

Expected calibration error in the internal and external evaluation sets

Forecast time	Internal Evaluation Set				External Evaluation Set
	Without Preoperative Feature		With Preoperative Feature		Without Preoperative Feature		With Preoperative Feature
	Mono-ART	Four-Channel	Mono-ART	Four-Channel (HypoBridCast)	Mono-ART	Four-Channel	Mono-ART	Four-Channel (HypoBridCast)
5 min	0.084 (0.050—0.119)	0.097 (0.067—0.125)	0.095 (0.063—0.123)	0.077 (0.039—0.111)	0.125 (0.075—0.174)	0.147 (0.088—0.217)	0.100 (0.059—0.145)	0.092 (0.062—0.123)
10 min	0.101 (0.075—0.125)	0.154 (0.069—0.231)	0.068 (0.043—0.095)	0.091 (0.038—0.147)	0.180 (0.108—0.260)	0.142 (0.077—0.202)	0.116 (0.068—0.162)	0.169 (0.108—0.230)
15 min	0.037 (0.023—0.051)	0.143 (0.079—0.209)	0.167 (0.120—0.201)	0.088 (0.060—0.119)	0.182 (0.109—0.253)	0.108 (0.071—0.147)	0.181 (0.128—0.250)	0.160 (0.102—0.225)

Expected calibration error (ECE) of the four types model (Mono-ART channel with preoperative features, Mono-ART channel without preoperative features, four-channel with preoperative features [HypoBridCast Model], four-channel without preoperative features) in the internal and external evaluation. The forecast time indicates the time before a hypotensive event. Low values represent a low error, thus better calibration. The 95% confidence interval is indicated by the values within the brackets

SHAP analysis indicated that dynamic waveform features, particularly ART, contributed most to model predictions, whereas preoperative variables had smaller contributions (Figure S4).

Analysis of MAP values at prediction time (t = 0) showed that positive samples were distributed across a wide range of MAP values, including values well above the hypotension threshold (> 75 mmHg, Table S8), indicating that predictions were not limited to near-threshold conditions.

Subgroup analyses across age, ASA status, and hypertension demonstrated generally consistent performance in both internal and external datasets, with only modest variation across prediction horizons (Table S9).

Discussion

In this study, we developed HypoBridCast, a hybrid Conv1D–TF architecture integrating intraoperative multi-channel waveforms with preoperative baseline data to predict intraoperative hypotension. Across 5-, 10-, and 15-min prediction horizons, the model achieved the highest discrimination among evaluated models, with generally acceptable calibration.

Analysis of model behavior and feature contribution

The model generally outperformed single-channel ART models across prediction horizons, indicating that multi-channel waveform integration provides additional predictive information beyond arterial pressure alone. However, performance differences between HypoBridCast and Four-Channel model were small, suggesting limited incremental benefit from baseline features.

Different waveform modalities capture complementary physiological processes. PPG reflects peripheral perfusion [23, 24], ECG provides information on cardiac rhythm and early hypoperfusion [25], and capnography reflects respiratory–circulatory coupling, where a declining trend may indicate reduced cardiac output [26]. Their integration enables a more comprehensive representation of haemodynamic status than single-signal models. However, performance did not increase monotonically with additional inputs. For example, adding Capno alone to ART yielded limited or even inferior benefit, whereas its contribution became more apparent when combined with other waveforms, highlighting the importance of appropriate signal integration.

SHAP analysis showed that dynamic waveform features, particularly ART, contributed most to predictions, whereas preoperative variables had smaller effects. This likely reflects the acute physiological nature of intraoperative hypotension, where real-time haemodynamic signals are more informative. Nevertheless, baseline characteristics may still provide complementary context for patient-specific risk assessment [27].

Model generalizability and calibration

External evaluation demonstrated reduced discrimination, particularly at longer prediction horizons. The advantage of HypoBridCast over single-channel ART models was less consistent for 15-min predictions, with largely overlapping discrimination performance between models. Calibration performance varied across prediction horizons. In the internal evaluation set, HypoBridCast did not consistently achieve the best calibration, with slightly higher ECE observed at certain time points (particularly at longer prediction horizons) compared with simpler models. In the external dataset, overall calibration declined, particularly at longer prediction horizons (Figure S3B). These findings may reflect the influence of dataset heterogeneity and increased noise accumulation over longer prediction windows. Further work is needed to improve waveform denoising, feature selection and overall model robustness across diverse datasets. In clinical deployment, post-hoc recalibration methods such as Platt scaling or isotonic regression may be required to ensure reliable risk estimation in new clinical environments.

Analysis of MAP distributions at prediction time showed that positive samples were not restricted to values immediately below the hypotension threshold. This suggests that predictions were not driven solely by proximity to the threshold but also reflect underlying physiological patterns preceding hypotension.

Clinical implications and practical considerations

The model demonstrated relatively high specificity with moderate sensitivity, suggesting a lower false-alarm rate that may help reduce unnecessary interventions, which could otherwise lead to inefficient resource use or potential patient harm [28]. However, the moderate sensitivity indicates that some hypotensive events may be missed; therefore, model predictions should be interpreted alongside real-time monitoring and clinical judgment. The model was developed using invasive arterial waveforms, which provide high-resolution physiological data and are commonly available in higher-risk surgical patients. In contrast, non-invasive blood pressure monitoring is intermittent and has lower temporal resolution, which may limit its suitability for waveform-based real-time prediction. Prior studies have shown that continuous invasive blood pressure monitoring is associated with reduced exposure to IOH [29]. Extension of the model to non-invasive monitoring modalities warrants further investigation.

Comparison with other studies

The Hypotension Prediction Index (HPI) developed by Hatib et al. [6] has been widely studied [28, 30], with variable reported performance and clinical impact [31–34]. Unlike HPI, which excluded MAP values between 65–75 mmHg during training, a design that has been discussed in the literature as a potential source of dataset selection effects, [35, 36] the present model was developed using a consistent definition and evaluated across the full range of blood pressure values. In addition, it relies on standard monitoring signals rather than proprietary hardware, potentially facilitating broader implementation [37]. Consistent with observations reported by Lee et al. [11], models using multiple physiological waveforms may provide improved predictive performance compared with those relying solely on arterial pressure signals. In our study, the four-channel model (ART, ECG, PPG, and Capno) achieved better overall performance. Recent work using transformer-based models with lower-resolution data (sampled every 15 s) has also been reported [38]. In contrast, the present model uses high-resolution continuous waveforms, which may better capture short-term haemodynamic dynamics that reflect physiological responses to clinical interventions [39] without requiring manual input.

Limitations and future works

This study has several limitations. First, the model requires multiple waveform channels and may be sensitive to missing data. Future work will focus on improving robustness under partial signal loss. Second, the development and external evaluation datasets were derived from different countries, and further evaluation across more diverse populations and clinical settings is needed to confirm generalizability. Finally, the impact of alternative input window lengths was not systematically evaluated. Although prior studies suggest that prediction performance is relatively robust to variations in input duration, the absence of a dedicated sensitivity analysis in the present dataset remains a limitation and warrants further investigation.

Conclusions

This study developed HypoBridCast, a deep learning model integrating preoperative and intraoperative multi-channel waveform data for short-term prediction of intraoperative hypotension. The model demonstrated strong discrimination across prediction horizons and improved performance compared with arterial pressure–only models, with generally acceptable calibration. As a scalable approach based on routinely collected data, it supports real-time perioperative application for personalized blood pressure management and potentially improved outcomes.

Abbreviations

IOH

Intraoperative hypotension

MAP

Mean arterial pressure

CNNs

Convolutional Neural Networks

RNNs

Recurrent Neural Networks

HypoBridCast

Hypotension Hybrid Forecast

Conv1D

1D Convolutional Layers

TF

Transformer

ART

Arterial blood pressure

ECG

Electrocardiogram

PPG

Photoplethysmogram

Capno

Capnography

EMR

Electronic medical records

LIS

Laboratory information system

PPV

Positive predictive value

NPV

Negative predictive value

SHAP

Shapley Additive Explanations

ROC

Receiver Operating Characteristic

HPI

Hypotension Prediction Index

Authors’ contributions

YW designed and conducted the design and conduction of the study, acquired and analysed the data, and wrote the text. JS and JX contributed to design of the study, analysis of the data and reviewed the main text. XZ contributed to the design, conduction of the study and analysis of the results. LC and LB contributed to conduction of the study and helped to write the main text. KS contributed to the collection of signal wave. All authors have read and approved the final manuscript.

Funding

This work was supported by the China University-Industry-Research Innovation Fund under Grant No. 2024GY039.

Data availability

The datasets generated and analysed during the current study are available in the Vital DB, https://physionet.org/content/vitaldb/1.0.0/. [15].

Declarations

Ethics approval and consent to participate

The research received ethical approval from Zhongda Hospital, Southeast University (No. 2024ZDSYLL476-P01). As a retrospective study, informed consent was waived, and there was no patient or public involvement in the research.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.