VitalDB, a high-fidelity multi-parameter vital signs database in surgical patients

Atmospheric particulate matter and hypoxaemia in Korean children receiving general anaesthesia: A retrospective analysis

BACKGROUND

The association between the concentration of atmospheric particulate matter on the day of surgery and the occurrence of intra-operative hypoxaemia in children receiving general anaesthesia is unclear.

OBJECTIVE

To investigate the association between the exposure to particulate matter on the day of surgery and the occurrence of intra-operative hypoxaemia, defined as a pulse oximetry oxygen saturation of less than 90% for more than 1 min, in children.

DESIGN

Retrospective study.

SETTING

Single-centre.

PARTICIPANTS

Children aged 18 years or younger who received general anaesthesia between January 2019 and October 2020.

INTERVENTION

Information on daily levels of particulate matter with a diameter 10 μm or less and 2.5 μm or less measured within a neighbourhood corresponding to the area defined by the hospital's zip code was obtained from publicly available air-quality data.

MAIN OUTCOME MEASURES

The primary outcome was intra-operative hypoxaemia, defined as a pulse oximetry oxygen saturation of less than 90% lasting for more than 1 min, manually verified by anaesthesiologists using vital sign registry data extracted at 2 s intervals.

RESULTS

Of the patients finally analysed, 3.85% (489/13 175) experienced intra-operative hypoxaemia. Higher levels of particulate matter 10 μm or less in diameter (≥81 μg m -3 , 17/275, 6.2%) were associated with an increased occurrence of intra-operative hypoxaemia compared with lower particulate matter concentrations [<81 μg m -3 , 472/12 900, 3.7%; adjusted odds ratio, 1.71; 95% confidence interval (CI), 1.04 to 2.83; P  = 0.035].

CONCLUSION

The level of particulate matter on the day of surgery pose a risk of intra-operative hypoxaemia in children receiving general anaesthesia. If the concentrations of particulate matter 10 μm or less in diameter on the day of surgery are high, children receiving general anaesthesia should be managed with increased caution.

Decision Curve Analysis of In-Hospital Mortality Prediction Models: The Relative Value of Pre- and Intraoperative Data For Decision-Making

BACKGROUND

Clinical prediction modeling plays a pivotal part in modern clinical care, particularly in predicting the risk of in-hospital mortality. Recent modeling efforts have focused on leveraging intraoperative data sources to improve model performance. However, the individual and collective benefit of pre- and intraoperative data for clinical decision-making remains unknown. We hypothesized that pre- and intraoperative predictors contribute equally to the net benefit in a decision curve analysis (DCA) of in-hospital mortality prediction models that include pre- and intraoperative predictors.

METHODS

Data from the VitalDB database featuring a subcohort of 6043 patients were used. A total of 141 predictors for in-hospital mortality were grouped into preoperative (demographics, intervention characteristics, and laboratory measurements) and intraoperative (laboratory and monitor data, drugs, and fluids) data. Prediction models using either preoperative, intraoperative, or all data were developed with multiple methods (logistic regression, neural network, random forest, gradient boosting machine, and a stacked learner). Predictive performance was evaluated by the area under the receiver-operating characteristic curve (AUROC) and under the precision-recall curve (AUPRC). Clinical utility was examined with a DCA in the predefined risk preference range (denoted by so-called treatment threshold probabilities) between 0% and 20%.

RESULTS

AUROC performance of the prediction models ranged from 0.53 to 0.78. AUPRC values ranged from 0.02 to 0.25 (compared to the incidence of 0.09 in our dataset) and high AUPRC values resulted from prediction models based on preoperative laboratory values. A DCA of pre- and intraoperative prediction models highlighted that preoperative data provide the largest overall benefit for decision-making, whereas intraoperative values provide only limited benefit for decision-making compared to preoperative data. While preoperative demographics, comorbidities, and surgery-related data provide the largest benefit for low treatment thresholds up to 5% to 10%, preoperative laboratory measurements become the dominant source for decision support for higher thresholds.

CONCLUSIONS

When it comes to predicting in-hospital mortality and subsequent decision-making, preoperative demographics, comorbidities, and surgery-related data provide the largest benefit for clinicians with risk-averse preferences, whereas preoperative laboratory values provide the largest benefit for decision-makers with more moderate risk preferences. Our decision-analytic investigation of different predictor categories moves beyond the question of whether certain predictors provide a benefit in traditional performance metrics (eg, AUROC). It offers a nuanced perspective on for whom these predictors might be beneficial in clinical decision-making. Follow-up studies requiring larger datasets and dedicated deep-learning models to handle continuous intraoperative data are essential to examine the robustness of our results.

Prediction for Perioperative Stroke Using Intraoperative Parameters

BACKGROUND

Perioperative stroke is a severe complication following surgery. To identify patients at risk for perioperative stroke, several prediction models based on the preoperative factors were suggested. Prediction models often focus on preoperative patient characteristics to assess stroke risk. However, most existing models primarily base their predictions on the patient's baseline characteristics before surgery. We aimed to develop a machine-learning model incorporating both pre- and intraoperative variables to predict perioperative stroke.

METHODS AND RESULTS

This study included patients who underwent noncardiac surgery at 2 hospitals with the data of 15 752 patients from Seoul National University Hospital used for development and temporal internal validation, and the data of 449 patients from Boramae Medical Center used for external validation. Perioperative stroke was defined as a newly developed ischemic lesion on diffusion-weighted imaging within 30 days of surgery. We developed a prediction model composed of pre- and intraoperative factors (integrated model) and compared it with a model consisting of preoperative features alone (preoperative model). Perioperative stroke developed in 109 (0.69%) patients in the Seoul National University Hospital group and 11 patients (2.45%) in the Boramae Medical Center group. The integrated model demonstrated superior predictive performance with area under the curve values of 0.824 (95% CI, 0.762-0.880) versus 0.584 (95% CI, 0.499-0.667; P<0.001) in the internal validation; and 0.716 (95% CI, 0.560-0.859) versus 0.505 (95% CI, 0.343-0.654; P=0.018) in the external validation, compared to the preoperative model.

CONCLUSIONS

We suggest that incorporating intraoperative factors into perioperative stroke prediction models can improve their accuracy.

Prediction of intraoperative hypotension using deep learning models based on non-invasive monitoring devices

PURPOSE

Intraoperative hypotension is associated with adverse outcomes. Predicting and proactively managing hypotension can reduce its incidence. Previously, hypotension prediction algorithms using artificial intelligence were developed for invasive arterial blood pressure monitors. This study tested whether routine non-invasive monitors could also predict intraoperative hypotension using deep learning algorithms.

METHODS

An open-source database of non-cardiac surgery patients ( https://vitadb.net/dataset ) was used to develop the deep learning algorithm. The algorithm was validated using external data obtained from a tertiary Korean hospital. Intraoperative hypotension was defined as a systolic blood pressure less than 90 mmHg. The input data included five monitors: non-invasive blood pressure, electrocardiography, photoplethysmography, capnography, and bispectral index. The primary outcome was the performance of the deep learning model as assessed by the area under the receiver operating characteristic curve (AUROC).

RESULTS

Data from 4754 and 421 patients were used for algorithm development and external validation, respectively. The fully connected model of Multi-head Attention architecture and the Globally Attentive Locally Recurrent model with Focal Loss function were able to predict intraoperative hypotension 5 min before its occurrence. The AUROC of the algorithm was 0.917 (95% confidence interval [CI], 0.915-0.918) for the original data and 0.833 (95% CI, 0.830-0.836) for the external validation data. Attention map, which quantified the contributions of each monitor, showed that our algorithm utilized data from each monitor with weights ranging from 8 to 22% for determining hypotension.

CONCLUSIONS

A deep learning model utilizing multi-channel non-invasive monitors could predict intraoperative hypotension with high accuracy. Future prospective studies are needed to determine whether this model can assist clinicians in preventing hypotension in patients undergoing surgery with non-invasive monitoring.

Non-invasive prediction of massive transfusion during surgery using intraoperative hemodynamic monitoring data

OBJECTIVE

Failure to receive prompt blood transfusion leads to severe complications if massive bleeding occurs during surgery. For the timely preparation of blood products, predicting the possibility of massive transfusion (MT) is essential to decrease morbidity and mortality. This study aimed to develop a model for predicting MT 10 min in advance using non-invasive bio-signal waveforms that change in real-time.

METHODS

In this retrospective study, we developed a deep learning-based algorithm (DLA) to predict intraoperative MT within 10 min. MT was defined as the transfusion of 3 or more units of red blood cells within an hour. The datasets consisted of 18,135 patients who underwent surgery at Seoul National University Hospital (SNUH) for model development and internal validation and 621 patients who underwent surgery at the Boramae Medical Center (BMC) for external validation. We constructed the DLA by using features extracted from plethysmography (collected at 500 Hz) and hematocrit measured during surgery.

RESULTS

Among 18,135 patients in SNUH and 621 patients in BMC, 265 patients (1.46%) and 14 patients (2.25%) received MT during surgery, respectively. The area under the receiver operating characteristic curve (AUROC) of DLA predicting intraoperative MT before 10 min was 0.962 (95% confidence interval [CI], 0.948-0.974) in internal validation and 0.922 (95% CI, 0.882-0.959) in external validation, respectively.

CONCLUSION

The DLA can successfully predict intraoperative MT using non-invasive bio-signal waveforms.

SQI-DOANet: electroencephalogram-based deep neural network for estimating signal quality index and depth of anaesthesia

Objective. Monitoring the depth of anaesthesia (DOA) during surgery is of critical importance. However, during surgery electroencephalography (EEG) is usually subject to various disturbances that affect the accuracy of DOA. Therefore, accurately estimating noise in EEG and reliably assessing DOA remains an important challenge. In this paper, we proposed a signal quality index (SQI) network (SQINet) for assessing the EEG signal quality and a DOA network (DOANet) for analyzing EEG signals to precisely estimate DOA. The two networks are termed SQI-DOANet.Approach. The SQINet contained a shallow convolutional neural network to quickly determine the quality of the EEG signal. The DOANet comprised a feature extraction module for extracting features, a dual attention module for fusing multi-channel and multi-scale information, and a gated multilayer perceptron module for extracting temporal information. The performance of the SQI-DOANet model was validated by training and testing the model on the large VitalDB database, with the bispectral index (BIS) as the reference standard.Main results. The proposed DOANet yielded a Pearson correlation coefficient with the BIS score of 0.88 in the five-fold cross-validation, with a mean absolute error (MAE) of 4.81. The mean Pearson correlation coefficient of SQI-DOANet with the BIS score in the five-fold cross-validation was 0.82, with an MAE of 5.66.Significance. The SQI-DOANet model outperformed three compared methods. The proposed SQI-DOANet may be used as a new deep learning method for DOA estimation. The code of the SQI-DOANet will be made available publicly athttps://github.com/YuRui8879/SQI-DOANet.

Enhancing Hypotension Prediction in Real-time Patient Monitoring Through Deep Learning: A Novel Application of XResNet with Contrastive Learning and Value Attention Mechanisms

The precise prediction of hypotension is vital for advancing preemptive patient care strategies. Traditional machine learning approaches, while instrumental in this field, are hampered by their dependence on structured historical data and manual feature extraction techniques. These methods often fall short of recognizing the intricate patterns present in physiological signals. Addressing this limitation, our study introduces an innovative application of deep learning technologies, utilizing a sophisticated end-to-end architecture grounded in XResNet. This architecture is further enhanced by the integration of contrastive learning and a value attention mechanism, specifically tailored to analyze arterial blood pressure (ABP) waveform signals. Our approach improves the performance of hypotension prediction over the existing state-of-theart ABP model [7]. This research represents a step towards optimizing patient care, embodying the next generation of AI-driven healthcare solutions. Through our findings, we demonstrate the promise of deep learning in overcoming the limitations of conventional prediction models, thereby offering an avenue for enhancing patient outcomes in clinical settings.

INSPIRE, a publicly available research dataset for perioperative medicine

We present the INSPIRE dataset, a publicly available research dataset in perioperative medicine, which includes approximately 130,000 surgical operations at an academic institution in South Korea over a ten-year period between 2011 and 2020. This comprehensive dataset includes patient characteristics such as age, sex, American Society of Anesthesiologists physical status classification, diagnosis, surgical procedure code, department, and type of anaesthesia. The dataset also includes vital signs in the operating theatre, general wards, and intensive care units (ICUs), laboratory results from six months before admission to six months after discharge, and medication during hospitalisation. Complications include total hospital and ICU length of stay and in-hospital death. We hope this dataset will inspire collaborative research and development in perioperative medicine and serve as a reproducible external validation dataset to improve surgical outcomes.

Robust arterial compliance estimation with Katz's fractal dimension of photoplethysmography

Arterial compliance (AC) plays a crucial role in vascular aging and cardiovascular disease. The ability to continuously estimate aortic AC or its surrogate, pulse pressure (PP), through wearable devices is highly desirable, given its strong association with daily activities. While the single-site photoplethysmography (PPG)-derived arterial stiffness indices show reasonable correlations with AC, they are susceptible to noise interference, limiting their practical use. To overcome this challenge, our study introduces a noise-resistant indicator of AC: Katz's fractal dimension (KFD) of PPG signals. We showed that KFD integrated the signal complexity arising from compliance changes across a cardiac cycle and vascular structural complexity, thereby decreasing its dependence on individual characteristic points. To assess its capability in measuring AC, we conducted a comprehensive evaluation using both in silico studies with 4374 virtual human data and real-world measurements. In the virtual human studies, KFD demonstrated a strong correlation with AC (r = 0.75), which only experienced a slight decrease to 0.66 at a signal-to-noise ratio of 15dB, surpassing the best PPG-morphology-derived AC measure (r = 0.41) under the same noise condition. In addition, we observed that KFD's sensitivity to AC varied based on the individual's hemodynamic status, which may further enhance the accuracy of AC estimations. These in silico findings were supported by real-world measurements encompassing diverse health conditions. In conclusion, our study suggests that PPG-derived KFD has the potential to continuously and reliably monitor arterial compliance, enabling unobtrusive and wearable assessment of cardiovascular health.

LDSG-Net: an efficient lightweight convolutional neural network for acute hypotensive episode prediction during ICU hospitalization

Objective. Acute hypotension episode (AHE) is one of the most critical complications in intensive care unit (ICU). A timely and precise AHE prediction system can provide clinicians with sufficient time to respond with proper therapeutic measures, playing a crucial role in saving patients' lives. Recent studies have focused on utilizing more complex models to improve predictive performance. However, these models are not suitable for clinical application due to limited computing resources for bedside monitors.Approach. To address this challenge, we propose an efficient lightweight dilated shuffle group network. It effectively incorporates shuffling operations into grouped convolutions on the channel and dilated convolutions on the temporal dimension, enhancing global and local feature extraction while reducing computational load.Main results. Our benchmarking experiments on the MIMIC-III and VitalDB datasets, comprising 6036 samples from 1304 patients and 2958 samples from 1047 patients, respectively, demonstrate that our model outperforms other state-of-the-art lightweight CNNs in terms of balancing parameters and computational complexity. Additionally, we discovered that the utilization of multiple physiological signals significantly improves the performance of AHE prediction. External validation on the MIMIC-IV dataset confirmed our findings, with prediction accuracy for AHE 5 min prior reaching 93.04% and 92.04% on the MIMIC-III and VitalDB datasets, respectively, and 89.47% in external verification.Significance. Our study demonstrates the potential of lightweight CNN architectures in clinical applications, providing a promising solution for real-time AHE prediction under resource constraints in ICU settings, thereby marking a significant step forward in improving patient care.

Internal and External Validation of Machine Learning Models for Predicting Acute Kidney Injury Following Non-Cardiac Surgery Using Open Datasets

This study developed and validated a machine learning model to accurately predict acute kidney injury (AKI) after non-cardiac surgery, aiming to improve patient outcomes by assessing its clinical feasibility and generalizability. We conducted a retrospective cohort study using data from 76,032 adults who underwent non-cardiac surgery at a single tertiary medical center between March 2019 and February 2021, and used data from 5512 patients from the VitalDB open dataset for external model validation. The predictive variables for model training consisted of demographic, preoperative laboratory, and intraoperative data, including calculated statistical values such as the minimum, maximum, and mean intraoperative blood pressure. When predicting postoperative AKI, our gradient boosting machine model incorporating all the variables achieved the best results, with AUROC values of 0.868 and 0.757 for the internal and external validations using the VitalDB dataset, respectively. The model using intraoperative data performed best in internal validation, while the model with preoperative data excelled in external validation. In this study, we developed a predictive model for postoperative AKI in adult patients undergoing non-cardiac surgery using preoperative and intraoperative data, and external validation demonstrated the efficacy of open datasets for generalization in medical artificial modeling research.

CiGNN: A Causality-Informed and Graph Neural Network Based Framework for Cuffless Continuous Blood Pressure Estimation

Causalityholds profound potentials to dissipate confusion and improve accuracy in cuffless continuous blood pressure (BP) estimation, an area often neglected in current research. In this study, we propose a two-stage framework, CiGNN, that seamlessly integrates causality and graph neural network (GNN) for cuffless continuous BP estimation. The first stage concentrates on the generation of a causal graph between BP and wearable features from the the perspective of causal inference, so as to identify features that are causally related to BP variations. This stage is pivotal for the identification of novel causal features from the causal graph beyond pulse transit time (PTT). We found these causal features empower better tracking in BP changes compared to PTT. For the second stage, a spatio-temporal GNN (STGNN) is utilized to learn from the causal graph obtained from the first stage. The STGNN can exploit both the spatial information within the causal graph and temporal information from beat-by-beat cardiac signals for refined cuffless continuous BP estimation. We evaluated the proposed method with three datasets that include 305 subjects (102 hypertensive patients) with age ranging from 20-90 and BP at different levels, with the continuous Finapres BP as references. The mean absolute difference (MAD) for estimated systolic blood pressure (SBP) and diastolic blood pressure (DBP) were 3.77 mmHg and 2.52 mmHg, respectively, which outperformed comparison methods. In all cases including subjects with different age groups, while doing various maneuvers that induces BP changes at different levels and with or without hypertension, the proposed CiGNN method demonstrates superior performance for cuffless continuous BP estimation. These findings suggest that the proposed CiGNN is a promising approach in elucidating the causal mechanisms of cuffless BP estimation and can substantially enhance the precision of BP measurement.

Development of Prediction Model for Intensive Care Unit Admission Based on Heart Rate Variability: A Case-Control Matched Analysis

This study aimed to develop a predictive model for intensive care unit (ICU) admission by using heart rate variability (HRV) data. This retrospective case-control study used two datasets (emergency department [ED] patients admitted to the ICU, and patients in the operating room without ICU admission) from a single academic tertiary hospital. HRV metrics were measured every 5 min using R-peak-to-R-peak (R-R) intervals. We developed a generalized linear mixed model to predict ICU admission and assessed the area under the receiver operating characteristic curve (AUC). Odds ratios (ORs) with 95% confidence intervals (CIs) were calculated from the coefficients. We analyzed 610 (ICU: 122; non-ICU: 488) patients, and the factors influencing the odds of ICU admission included a history of diabetes mellitus (OR [95% CI]: 3.33 [1.71-6.48]); a higher heart rate (OR [95% CI]: 3.40 [2.97-3.90] per 10-unit increase); a higher root mean square of successive R-R interval differences (RMSSD; OR [95% CI]: 1.36 [1.22-1.51] per 10-unit increase); and a lower standard deviation of R-R intervals (SDRR; OR [95% CI], 0.68 [0.60-0.78] per 10-unit increase). The final model achieved an AUC of 0.947 (95% CI: 0.906-0.987). The developed model effectively predicted ICU admission among a mixed population from the ED and operating room.

Measurement error of pulse pressure variation

Dynamic preload parameters are used to guide perioperative fluid management. However, reported cut-off values vary and the presence of a gray zone complicates clinical decision making. Measurement error, intrinsic to the calculation of pulse pressure variation (PPV) has not been studied but could contribute to this level of uncertainty. The purpose of this study was to quantify and compare measurement errors associated with PPV calculations. Hemodynamic data of patients undergoing liver transplantation were extracted from the open-access VitalDatabase. Three algorithms were applied to calculate PPV based on 1 min observation periods. For each method, different durations of sampling periods were assessed. Best Linear Unbiased Prediction was determined as the reference PPV-value for each observation period. A Bayesian model was used to determine bias and precision of each method and to simulate the uncertainty of measured PPV-values. All methods were associated with measurement error. The range of differential and proportional bias were [- 0.04%, 1.64%] and [0.92%, 1.17%] respectively. Heteroscedasticity influenced by sampling period was detected in all methods. This resulted in a predicted range of reference PPV-values for a measured PPV of 12% of [10.2%, 13.9%] and [10.3%, 15.1%] for two selected methods. The predicted range in reference PPV-value changes for a measured absolute change of 1% was [- 1.3%, 3.3%] and [- 1.9%, 4%] for these two methods. We showed that all methods that calculate PPV come with varying degrees of uncertainty. Accounting for bias and precision may have important implications for the interpretation of measured PPV-values or PPV-changes.

A multimodal stacked ensemble model for cardiac output prediction utilizing cardiorespiratory interactions during general anesthesia

This study examined the possibility of estimating cardiac output (CO) using a multimodal stacking model that utilizes cardiopulmonary interactions during general anesthesia and outlined a retrospective application of machine learning regression model to a pre-collected dataset. The data of 469 adult patients (obtained from VitalDB) with normal pulmonary function tests who underwent general anesthesia were analyzed. The hemodynamic data in this study included non-invasive blood pressure, plethysmographic heart rate, and SpO2. CO was recorded using Vigileo and EV1000 (pulse contour technique devices). Respiratory data included mechanical ventilation parameters and end-tidal CO2 levels. A generalized linear regression model was used as the metalearner for the multimodal stacking ensemble method. Random forest, generalized linear regression, gradient boosting machine, and XGBoost were used as base learners. A Bland-Altman plot revealed that the multimodal stacked ensemble model for CO prediction from 327 patients had a bias of - 0.001 L/min and - 0.271% when calculating the percentage of difference using the EV1000 device. Agreement of model CO prediction and measured Vigileo CO in 142 patients reported a bias of - 0.01 and - 0.333%. Overall, this model predicts CO compared to data obtained by the pulse contour technique CO monitors with good agreement.

Quantifying the Suitability of Biosignals Acquired During Surgery for Multimodal Analysis

Goal: Recently, large datasets of biosignals acquired during surgery became available. As they offer multiple physiological signals measured in parallel, multimodal analysis - which involves their joint analysis - can be conducted and could provide deeper insights than unimodal analysis based on a single signal. However, it is unclear what percentage of intraoperatively acquired data is suitable for multimodal analysis. Due to the large amount of data, manual inspection and labelling into suitable and unsuitable segments are not feasible. Nevertheless, multimodal analysis is performed successfully in sleep studies since many years as their signals have proven suitable. Hence, this study evaluates the suitability to perform multimodal analysis on a surgery dataset (VitalDB) using a multi-center sleep dataset (SIESTA) as reference. Methods: We applied widely known algorithms entitled "signal quality indicators" to the common biosignals in both datasets, namely electrocardiography, electroencephalography, and respiratory signals split in segments of 10 s duration. As there are no multimodal methods available, we used only unimodal signal quality indicators. In case, all three signals were determined as being adequate by the indicators, we assumed that the whole signal segment was suitable for multimodal analysis. Results: 82% of SIESTA and 72% of VitalDB are suitable for multimodal analysis. Unsuitable signal segments exhibit constant or physiologically unreasonable values. Histogram examination indicated similar signal quality distributions between the datasets, albeit with potential statistical biases due to different measurement setups. Conclusions: The majority of data within VitalDB is suitable for multimodal analysis.

Automatic segmentation of atrial fibrillation and flutter in single-lead electrocardiograms by self-supervised learning and Transformer architecture

OBJECTIVES

Automatic detection of atrial fibrillation and flutter (AF/AFL) is a significant concern in preventing stroke and mitigating hemodynamic instability. Herein, we developed a Transformer-based deep learning model for AF/AFL segmentation in single-lead electrocardiograms (ECGs) by self-supervised learning with masked signal modeling (MSM).

MATERIALS AND METHODS

We retrieved data from 11 open-source databases on PhysioNet; 7 of these databases included labeled ECGs, while the other 4 were without labels. Each database contained ECG recordings with durations of ≥30 s. A total of 24 intradialytic ECGs with paroxysmal AF/AFL during 4 h of hemodialysis sessions at Seoul National University Hospital were used for external validation. The model was pretrained by predicting masked areas of ECG signals and fine-tuned by predicting AF/AFL areas. Cross-database validation was used for evaluation, and the intersection over union (IOU) was used as a main performance metric in external database validation.

RESULTS

In the 7 labeled databases, the areas marked as AF/AFL constituted 41.1% of the total ECG signals, ranging from 0.19% to 51.31%. In the evaluation per ECG segment, the model achieved IOU values of 0.9254 and 0.9477 for AF/AFL segmentation and other segmentation tasks, respectively. When applied to intradialytic ECGs with paroxysmal AF/AFL, the IOUs for the segmentation of AF/AFL and non-AF/AFL were 0.9896 and 0.9650, respectively. Model performance by different training procedure indicated that pretraining with MSM and the application of an appropriate masking ratio both contributed to the model performance. It also showed higher IOUs of AF/AFL labels than in previous studies when training and test databases were matched.

CONCLUSION

The present model with self-supervised learning by MSM performs robustly in segmenting AF/AFL.

Association between cumulative duration of deep anesthesia and postoperative acute kidney injury after noncardiac surgeries: a retrospective observational study

BACKGROUND

Bispectral index (BIS) is a processed electroencephalography monitoring tool and is widely used in anesthetic depth monitoring. Deep anesthesia exposure may be associated with multiple adverse outcomes. However, the relationship between anesthetic depth and postoperative acute kidney injury (AKI) remains unclear. We sought to determine the effect of BIS-based deep anesthesia duration on postoperative AKI following noncardiac surgery.

METHODS

This retrospective study used data from the Vital Signs DataBase, including patients undergoing noncardiac surgeries with BIS monitoring. The BIS values were collected every second during anesthesia. Restricted cubic splines and logistic regression were used to assess the association between the cumulative duration of deep anesthesia and postoperative AKI.

RESULTS

4774 patients were eligible, and 129 (2.7%) experienced postoperative AKI. Restricted cubic splines showed that a cumulative duration of BIS < 45 was nonlinearly associated with postoperative AKI (P-overall = 0.033 and P-non-linear = 0.023). Using the group with the duration of BIS < 45 less than 15 min as the reference, ORs of postoperative AKI were 2.59 (95% confidence interval [CI]:0.60 to 11.09, p = 0.200) in the 15-100 min group, and 4.04 (95%CI:0.92 to 17.76, p = 0.064) in the ≥ 100 min group after adjusting for preoperative and intraoperative covariates in multivariable logistic regression.

CONCLUSIONS

The cumulative duration of BIS < 45 was independently and nonlinearly associated with the risk of postoperative AKI in patients undergoing noncardiac surgery.

Analysis of the Chaotic Component of Photoplethysmography and Its Association with Hemodynamic Parameters

Wearable technologies face challenges due to signal instability, hindering their usage. Thus, it is crucial to comprehend the connection between dynamic patterns in photoplethysmography (PPG) signals and cardiovascular health. In our study, we collected 401 multimodal recordings from two public databases, evaluating hemodynamic conditions like blood pressure (BP), cardiac output (CO), vascular compliance (C), and peripheral resistance (R). Using irregular-resampling auto-spectral analysis (IRASA), we quantified chaotic components in PPG signals and employed different methods to measure the fractal dimension (FD) and entropy. Our findings revealed that in surgery patients, the power of chaotic components increased with vascular stiffness. As the intensity of CO fluctuations increased, there was a notable strengthening in the correlation between most complexity measures of PPG and these parameters. Interestingly, some conventional morphological features displayed a significant decrease in correlation, indicating a shift from a static to dynamic scenario. Healthy subjects exhibited a higher percentage of chaotic components, and the correlation between complexity measures and hemodynamics in this group tended to be more pronounced. Causal analysis showed that hemodynamic fluctuations are main influencers for FD changes, with observed feedback in most cases. In conclusion, understanding chaotic patterns in PPG signals is vital for assessing cardiovascular health, especially in individuals with unstable hemodynamics or during ambulatory testing. These insights can help overcome the challenges faced by wearable technologies and enhance their usage in real-world scenarios.

Improved U-Net Model to Estimate Cardiac Output Based on Photoplethysmography and Arterial Pressure Waveform

We aimed to estimate cardiac output (CO) from photoplethysmography (PPG) and the arterial pressure waveform (ART) using a deep learning approach, which is minimally invasive, does not require patient demographic information, and is operator-independent, eliminating the need to artificially extract a feature of the waveform by implementing a traditional formula. We aimed to present an alternative to measuring cardiac output with greater accuracy for a wider range of patients. Using a publicly available dataset, we selected 543 eligible patients and divided them into test and training sets after preprocessing. The data consisted of PPG and ART waveforms containing 2048 points with the corresponding CO. We achieved an improvement based on the U-Net modeling framework and built a two-channel deep learning model to automatically extract the waveform features to estimate the CO in the dataset as the reference, acquired using the EV1000, a commercially available instrument. The model demonstrated strong consistency with the reference values on the test dataset. The mean CO was 5.01 ± 1.60 L/min and 4.98 ± 1.59 L/min for the reference value and the predicted value, respectively. The average bias was -0.04 L/min with a -1.025 and 0.944 L/min 95% limit of agreement (LOA). The bias was 0.79% with a 95% LOA between -20.4% and 18.8% when calculating the percentage of the difference from the reference. The normalized root-mean-squared error (RMSNE) was 10.0%. The Pearson correlation coefficient (r) was 0.951. The percentage error (PE) was 19.5%, being below 30%. These results surpassed the performance of traditional formula-based calculation methods, meeting clinical acceptability standards. We propose a dual-channel, improved U-Net deep learning model for estimating cardiac output, demonstrating excellent and consistent results. This method offers a superior reference method for assessing cardiac output in cases where it is unnecessary to employ specialized cardiac output measurement devices or when patients are not suitable for pulmonary-artery-catheter-based measurements, providing a viable alternative solution.

Development and validation of a prediction model for postoperative intensive care unit admission in patients with non-cardiac surgery

BACKGROUND

Accurately forecasting patients admitted to the intensive care units (ICUs) after surgery may improve clinical outcomes and guide the allocation of expensive and limited ICU resources. However, studies on predicting postoperative ICU admission in non-cardiac surgery have been limited.

OBJECTIVE

To develop and validate a prediction model combining pre- and intraoperative variables to predict ICU admission after non-cardiac surgery.

METHODS

This study is based on data from the Vital Signs DataBase (VitalDB) database. Predictors were selected using the least absolute shrinkage and selection operator regression method and logistic regression to develop a nomogram and an online web calculator. The model was internally verified by 1000-Bootstrap resampling. Performance of model was evaluated using area under the receiver operating characteristic curve (AUC), calibration curve and Brier score. The Youden's index was used to find the optimal nomogram's probability threshold. Clinical utility was assessed by decision curve analysis.

RESULTS

This study included 5216 non-cardiac surgery patients; of these, 812 (15.6%) required postoperative ICU admission. Potential predictors included age, ASA classification, surgical department, emergency surgery, preoperative albumin level, preoperative urea nitrogen level, intraoperative crystalloid, intraoperative transfusion, intraoperative catheterization, and surgical time. A nomogram was constructed with an AUC of 0.917 (95% CI: 0.907-0.926) and a Brier score of 0.077. The Bootstrap-adjusted AUC was 0.914; the adjusted Brier score was 0.078. The calibration curve showed good agreement between predicted and actual probabilities; and the decision curve indicated clinical usefulness. Finally, we established an online web calculator for clinical application (https://xuzhikun.shinyapps.io/postopICUadmission1/).

CONCLUSION

We developed and internally validated an easy-to-use nomogram for predicting ICU admission after non-cardiac surgery.

Availability of information needed to evaluate algorithmic fairness - A systematic review of publicly accessible critical care databases

BACKGROUND

Machine learning (ML) may improve clinical decision-making in critical care settings, but intrinsic biases in datasets can introduce bias into predictive models. This study aims to determine if publicly available critical care datasets provide relevant information to identify historically marginalized populations.

METHOD

We conducted a review to identify the manuscripts that report the training/validation of ML algorithms using publicly accessible critical care electronic medical record (EMR) datasets. The datasets were reviewed to determine if the following 12 variables were available: age, sex, gender identity, race and/or ethnicity, self-identification as an indigenous person, payor, primary language, religion, place of residence, education, occupation, and income.

RESULTS

7 publicly available databases were identified. Medical Information Mart for Intensive Care (MIMIC) reports information on 7 of the 12 variables of interest, Sistema de Informação de Vigilância Epidemiológica da Gripe (SIVEP-Gripe) on 7, COVID-19 Mexican Open Repository on 4, and eICU on 4. Other datasets report information on 2 or fewer variables. All 7 databases included information about sex and age. Four databases (57%) included information about whether a patient identified as native or indigenous. Only 3 (43%) included data about race and/or ethnicity. Two databases (29%) included information about residence, and one (14%) included information about payor, language, and religion. One database (14%) included information about education and patient occupation. No databases included information on gender identity and income.

CONCLUSION

This review demonstrates that critical care publicly available data used to train AI algorithms do not include enough information to properly look for intrinsic bias and fairness issues towards historically marginalized populations.

Exploring the clinical relevance of vital signs statistical calculations from a new-generation clinical information system

New information on the intensive care applications of new generation 'high-density data clinical information systems' (HDDCIS) is increasingly being published in the academic literature. HDDCIS avoid data loss from bedside equipment and some provide vital signs statistical calculations to promote quick and easy evaluation of patient information. Our objective was to study whether manual records of continuously monitored vital signs in the Paediatric Intensive Care Unit could be replaced by these statistical calculations. Here we conducted a prospective observational clinical study in paediatric patients with severe diabetic ketoacidosis, using a Medlinecare® HDDCIS, which collects information from bedside equipment (1 data point per parameter, every 3-5 s) and automatically provides hourly statistical calculations of the central trend and sample dispersion. These calculations were compared with manual hourly nursing records for patient heart and respiratory rates and oxygen saturation. The central tendency calculations showed identical or remarkably similar values and strong correlations with manual nursing records. The sample dispersion calculations differed from the manual references and showed weaker correlations. We concluded that vital signs calculations of central tendency can replace manual records, thereby reducing the bureaucratic burden of staff. The significant sample dispersion calculations variability revealed that automatic random measurements must be supervised by healthcare personnel, making them inefficient.

Heart Rate Variability Code: Does It Exist and Can We Hack It?

A code is generally defined as a system of signals or symbols for communication. Experimental evidence is synthesized for the presence and utility of such communication in heart rate variability (HRV) with particular attention to fetal HRV: HRV contains signatures of information flow between the organs and of response to physiological or pathophysiological stimuli as signatures of states (or syndromes). HRV exhibits features of time structure, phase space structure, specificity with respect to (organ) target and pathophysiological syndromes, and universality with respect to species independence. Together, these features form a spatiotemporal structure, a phase space, that can be conceived of as a manifold of a yet-to-be-fully understood dynamic complexity. The objective of this article is to synthesize physiological evidence supporting the existence of HRV code: hereby, the process-specific subsets of HRV measures indirectly map the phase space traversal reflecting the specific information contained in the code required for the body to regulate the physiological responses to those processes. The following physiological examples of HRV code are reviewed, which are reflected in specific changes to HRV properties across the signal-analytical domains and across physiological states and conditions: the fetal systemic inflammatory response, organ-specific inflammatory responses (brain and gut), chronic hypoxia and intrinsic (heart) HRV (iHRV), allostatic load (physiological stress due to surgery), and vagotomy (bilateral cervical denervation). Future studies are proposed to test these observations in more depth, and the author refers the interested reader to the referenced publications for a detailed study of the HRV measures involved. While being exemplified mostly in the studies of fetal HRV, the presented framework promises more specific fetal, postnatal, and adult HRV biomarkers of health and disease, which can be obtained non-invasively and continuously.

Open datasets in perioperative medicine: a narrative review

With the growing interest of researchers in machine learning and artificial intelligence (AI) based on large data, their roles in medical research have become increasingly prominent. Despite the proliferation of predictive models in perioperative medicine, external validation is lacking. Open datasets, defined as publicly available datasets for research, play a crucial role by providing high-quality data, facilitating collaboration, and allowing an objective evaluation of the developed models. Among the available datasets for surgical patients, VitalDB has been the most widely used, with the Medical Informatics Operating Room Vitals and Events Repository recently launched and the Informative Surgical Patient dataset for Innovative Research Environment expected to be released soon. For critically ill patients, the available resources include the Medical Information Mart for Intensive Care, the eICU Collaborative Research Database, the Amsterdam University Medical Centers Database, and the High time Resolution ICU Dataset, with the anticipated release of the Intensive Care Network with Million Patients' information for the AI Clinical decision support system Technology dataset. This review presents a detailed comparison of each to enrich our understanding of these open datasets for data science and AI research in perioperative medicine.

Demographic Information Fusion Using Attentive Pooling In CNN-GRU Model For Systolic Blood Pressure Estimation

Fusing demographic information into deep learning models has become of interest in recent end-to-end cuff-less blood pressure (BP) estimation studies in order to achieve improved performance. Conventionally, the demographic feature vector is concatenated with the pooled embedding vector. Here, using an attention-based convolutional neural network-gated recurrent unit (CNN-GRU), we present a new approach and fuse the demographic information into the attentive pooling module. Our results demonstrate that, under calibration-based testing protocol, the proposed approach provides improved systolic blood pressure (SBP) estimation accuracy (with R2=0.86 and mean absolute error (MAE)=4.90 mmHg) compared to both the baseline model with no demographic information fused, and the conventional approach of fusing demographic information. Our work showcases the feasibility of using attention-based methods to combine demographic features with deep learning models, and suggests new ways for fusing demographic information in deep learning models to achieve improved BP estimation accuracy.

Intensive care photoplethysmogram datasets and machine-learning for blood pressure estimation: Generalization not guarantied

The large MIMIC waveform dataset, sourced from intensive care units, has been used extensively for the development of Photoplethysmography (PPG) based blood pressure (BP) estimation algorithms. Yet, because the data comes from patients in severe conditions-often under the effect of drugs-it is regularly noted that the relationship between BP and PPG signal characteristics may be anomalous, a claim that we investigate here. A sample of 12,000 records from the MIMIC waveform dataset was stacked up against the 219 records of the PPG-BP dataset, an alternative public dataset obtained under controlled experimental conditions. The distribution of systolic and diastolic BP data and 31 PPG pulse morphological features was first compared between datasets. Then, the correlation between features and BP, as well as between the features themselves, was analysed. Finally, regression models were trained for each dataset and validated against the other. Statistical analysis showed significant p < 0.001 differences between the datasets in diastolic BP and in 20 out of 31 features when adjusting for heart rate differences. The eight features showing the highest rank correlation ρ   >   0.40 to systolic BP in PPG-BP all displayed muted correlation levels ρ   <   0.10 in MIMIC. Regression tests showed twice higher baseline predictive power with PPG-BP than with MIMIC. Cross-dataset regression displayed a practically complete loss of predictive power for all models. The differences between the MIMIC and PPG-BP dataset exposed in this study suggest that BP estimation models based on the MIMIC dataset have reduced predictive power on the general population.

Development and implementation of databases to track patient and safety outcomes

PURPOSE OF REVIEW

Recent advancements in big data analytical tools and large patient databases have expanded tremendously the opportunities to track patient and safety outcomes.We discuss the strengths and limitations of large databases and implementation in practice with a focus on the current opportunities to use technological advancements to improve patient safety.

RECENT FINDINGS

The most used sources of data for large patient safety observational studies are administrative databases, clinical registries, and electronic health records. These data sources have enabled research on patient safety topics ranging from rare adverse outcomes to large cohort studies of the modalities for pain control and safety of medications. Implementing the insights from big perioperative data research is augmented by automating data collection and tracking the safety outcomes on a provider, institutional, national, and global level. In the near future, big data from wearable devices, physiological waveforms, and genomics may lead to the development of personalized outcome measures.

SUMMARY

Patient safety research using large databases can provide actionable insights to improve outcomes in the perioperative setting. As datasets and methods to gain insights from those continue to grow, adopting novel technologies to implement personalized quality assurance initiatives can significantly improve patient care.