A Signal Processing Approach for Detection of Hemodynamic Instability before Decompensation

Prediction of Occult Hemorrhage in the Lower Body Negative Pressure Model: Initial Validation of Machine Learning Approaches

INTRODUCTION

Detection of occult hemorrhage (OH) before progression to clinically apparent changes in vital signs remains an important clinical problem in managing trauma patients. The resource-intensiveness associated with continuous clinical patient monitoring and rescue from frank shock makes accurate early detection and prediction with noninvasive measurement technology a desirable innovation. Despite significant efforts directed toward the development of innovative noninvasive diagnostics, the implementation and performance of the newest bedside technologies remain inadequate. This poor performance may reflect the limitations of univariate systems based on one sensor in one anatomic location. It is possible that when signals are measured with multiple modalities in multiple locations, the resulting multivariate anatomic and temporal patterns of measured signals may provide additional discriminative power over single technology univariate measurements. We evaluated the potential superiority of multivariate methods over univariate methods. Additionally, we utilized machine learning-based models to compare the performance of noninvasive-only to noninvasive-plus-invasive measurements in predicting the onset of OH.

MATERIALS AND METHODS

We applied machine learning methods to preexisting datasets derived using the lower body negative pressure human model of simulated hemorrhage. Employing multivariate measured physiological signals, we investigated the extent to which machine learning methods can effectively predict the onset of OH. In particular, we applied 2 ensemble learning methods, namely, random forest and gradient boosting.

RESULTS

Analysis of precision, recall, and area under the receiver operating characteristic curve showed a superior performance of multivariate approach to that of the univariate ones. In addition, when using both invasive and noninvasive features, random forest classifier had a recall 95% confidence interval (CI) of 0.81 to 0.86 with a precision 95% CI of 0.65 to 0.72. Interestingly, when only noninvasive features were employed, the results worsened only slightly to a recall 95% CI of 0.80 to 0.85 and a precision 95% CI of 0.61 to 0.73.

CONCLUSIONS

Multivariate ensemble machine learning-based approaches for the prediction of hemodynamic instability appear to hold promise for the development of effective solutions. In the lower body negative pressure multivariate hemorrhage model, predictions based only on noninvasive measurements performed comparably to those using both invasive and noninvasive measurements.

Sepsis Trajectory Prediction Using Privileged Information and Continuous Physiological Signals

The aim of this research is to apply the learning using privileged information paradigm to sepsis prognosis. We used signal processing of electrocardiogram and electronic health record data to construct support vector machines with and without privileged information to predict an increase in a given patient's quick-Sequential Organ Failure Assessment score, using a retrospective dataset. We applied this to both a small, critically ill cohort and a broader cohort of patients in the intensive care unit. Within the smaller cohort, privileged information proved helpful in a signal-informed model, and across both cohorts, electrocardiogram data proved to be informative to creating the prediction. Although learning using privileged information did not significantly improve results in this study, it is a paradigm worth studying further in the context of using signal processing for sepsis prognosis.

Prediction of episode of hemodynamic instability using an electrocardiogram based analytic: a retrospective cohort study

BACKGROUND

Predicting the onset of hemodynamic instability before it occurs remains a sought-after goal in acute and critical care medicine. Technologies that allow for this may assist clinicians in preventing episodes of hemodynamic instability (EHI). We tested a novel noninvasive technology, the Analytic for Hemodynamic Instability-Predictive Indicator (AHI-PI), which analyzes a single lead of electrocardiogram (ECG) and extracts heart rate variability and morphologic waveform features to predict an EHI prior to its occurrence.

METHODS

Retrospective cohort study at a quaternary care academic health system using data from hospitalized adult patients between August 2019 and April 2020 undergoing continuous ECG monitoring with intermittent noninvasive blood pressure (NIBP) or with continuous intraarterial pressure (IAP) monitoring.

RESULTS

AHI-PI's low and high-risk indications were compared with the presence of EHI in the future as indicated by vital signs (heart rate > 100 beats/min with a systolic blood pressure < 90 mmHg or a mean arterial blood pressure of < 70 mmHg). 4,633 patients were analyzed (3,961 undergoing NIBP monitoring, 672 with continuous IAP monitoring). 692 patients had an EHI (380 undergoing NIBP, 312 undergoing IAP). For IAP patients, the sensitivity and specificity of AHI-PI to predict EHI was 89.7% and 78.3% with a positive and negative predictive value of 33.7% and 98.4% respectively. For NIBP patients, AHI-PI had a sensitivity and specificity of 86.3% and 80.5% with a positive and negative predictive value of 11.7% and 99.5% respectively. Both groups performed with an AUC of 0.87. AHI-PI predicted EHI in both groups with a median lead time of 1.1 h (average lead time of 3.7 h for IAP group, 2.9 h for NIBP group).

CONCLUSIONS

AHI-PI predicted EHIs with high sensitivity and specificity and within clinically significant time windows that may allow for intervention. Performance was similar in patients undergoing NIBP and IAP monitoring.

Machine learning predicts blood lactate levels in children after cardiac surgery in paediatric ICU

BACKGROUND

Although serum lactate levels are widely accepted markers of haemodynamic instability, an alternative method to evaluate haemodynamic stability/instability continuously and non-invasively may assist in improving the standard of patient care. We hypothesise that blood lactate in paediatric ICU patients can be predicted using machine learning applied to arterial waveforms and perioperative characteristics.

METHODS

Forty-eight post-operative children, median age 4 months (2.9-11.8 interquartile range), mean baseline heart rate of 131 beats per minute (range 33-197), mean lactate level at admission of 22.3 mg/dL (range 6.3-71.1), were included. Morphological arterial waveform characteristics were acquired and analysed. Predicting lactate levels was accomplished using regression-based supervised learning algorithms, evaluated with hold-out cross-validation, including, basing prediction on the currently acquired physiological measurements along with those acquired at admission, as well as adding the most recent lactate measurement and the time since that measurement as prediction parameters. Algorithms were assessed with mean absolute error, the average of the absolute differences between actual and predicted lactate concentrations. Low values represent superior model performance.

RESULTS

The best performing algorithm was the tuned random forest, which yielded a mean absolute error of 3.38 mg/dL when predicting blood lactate with updated ground truth from the most recent blood draw.

CONCLUSIONS

The random forest is capable of predicting serum lactate levels by analysing perioperative variables, including the arterial pressure waveform. Thus, machine learning can predict patient blood lactate levels, a proxy for haemodynamic instability, non-invasively, continuously and with accuracy that may demonstrate clinical utility.

Prediction of Postoperative Deterioration in Cardiac Surgery Patients Using Electronic Health Record and Physiologic Waveform Data

BACKGROUND

Postoperative hemodynamic deterioration among cardiac surgical patients can indicate or lead to adverse outcomes. Whereas prediction models for such events using electronic health records or physiologic waveform data are previously described, their combined value remains incompletely defined. The authors hypothesized that models incorporating electronic health record and processed waveform signal data (electrocardiogram lead II, pulse plethysmography, arterial catheter tracing) would yield improved performance versus either modality alone.

METHODS

Intensive care unit data were reviewed after elective adult cardiac surgical procedures at an academic center between 2013 and 2020. Model features included electronic health record features and physiologic waveforms. Tensor decomposition was used for waveform feature reduction. Machine learning-based prediction models included a 2013 to 2017 training set and a 2017 to 2020 temporal holdout test set. The primary outcome was a postoperative deterioration event, defined as a composite of low cardiac index of less than 2.0 ml min-1 m-2, mean arterial pressure of less than 55 mmHg sustained for 120 min or longer, new or escalated inotrope/vasopressor infusion, epinephrine bolus of 1 mg or more, or intensive care unit mortality. Prediction models analyzed data 8 h before events.

RESULTS

Among 1,555 cases, 185 (12%) experienced 276 deterioration events, most commonly including low cardiac index (7.0% of patients), new inotrope (1.9%), and sustained hypotension (1.4%). The best performing model on the 2013 to 2017 training set yielded a C-statistic of 0.803 (95% CI, 0.799 to 0.807), although performance was substantially lower in the 2017 to 2020 test set (0.709, 0.705 to 0.712). Test set performance of the combined model was greater than corresponding models limited to solely electronic health record features (0.641; 95% CI, 0.637 to 0.646) or waveform features (0.697; 95% CI, 0.693 to 0.701).

CONCLUSIONS

Clinical deterioration prediction models combining electronic health record data and waveform data were superior to either modality alone, and performance of combined models was primarily driven by waveform data. Decreased performance of prediction models during temporal validation may be explained by data set shift, a core challenge of healthcare prediction modeling.

EDITOR’S PERSPECTIVE

Prediction of postoperative cardiac events in multiple surgical cohorts using a multimodal and integrative decision support system

Postoperative patients are at risk of life-threatening complications such as hemodynamic decompensation or arrhythmia. Automated detection of patients with such risks via a real-time clinical decision support system may provide opportunities for early and timely interventions that can significantly improve patient outcomes. We utilize multimodal features derived from digital signal processing techniques and tensor formation, as well as the electronic health record (EHR), to create machine learning models that predict the occurrence of several life-threatening complications up to 4 hours prior to the event. In order to ensure that our models are generalizable across different surgical cohorts, we trained the models on a cardiac surgery cohort and tested them on vascular and non-cardiac acute surgery cohorts. The best performing models achieved an area under the receiver operating characteristic curve (AUROC) of 0.94 on training and 0.94 and 0.82, respectively, on testing for the 0.5-hour interval. The AUROCs only slightly dropped to 0.93, 0.92, and 0.77, respectively, for the 4-hour interval. This study serves as a proof-of-concept that EHR data and physiologic waveform data can be combined to enable the early detection of postoperative deterioration events.

Detection of Hemodynamic Status Using an Analytic Based on an Electrocardiogram Lead Waveform

OBJECTIVES

Delayed identification of hemodynamic deterioration remains a persistent issue for in-hospital patient care. Clinicians continue to rely on vital signs associated with tachycardia and hypotension to identify hemodynamically unstable patients. A novel, noninvasive technology, the Analytic for Hemodynamic Instability (AHI), uses only the continuous electrocardiogram (ECG) signal from a typical hospital multiparameter telemetry monitor to monitor hemodynamics. The intent of this study was to determine if AHI is able to predict hemodynamic instability without the need for continuous direct measurement of blood pressure.

DESIGN

Retrospective cohort study.

SETTING

Single quaternary care academic health system in Michigan.

PATIENTS

Hospitalized adult patients between November 2019 and February 2020 undergoing continuous ECG and intra-arterial blood pressure monitoring in an intensive care setting.

INTERVENTIONS

None.

MEASUREMENTS AND MAIN RESULTS

One million two hundred fifty-two thousand seven hundred forty-two 5-minute windows of the analytic output were analyzed from 597 consecutive adult patients. AHI outputs were compared with vital sign indications of hemodynamic instability (heart rate > 100 beats/min, systolic blood pressure < 90 mm Hg, and shock index of > 1) in the same window. The observed sensitivity and specificity of AHI were 96.9% and 79.0%, respectively, with an area under the curve (AUC) of 0.90 for heart rate and systolic blood pressure. For the shock index analysis, AHI's sensitivity was 72.0% and specificity was 80.3% with an AUC of 0.81.

CONCLUSIONS

The AHI-derived hemodynamic status appropriately detected the various gold standard indications of hemodynamic instability (hypotension, tachycardia and hypotension, and shock index > 1). AHI may provide continuous dynamic hemodynamic monitoring capabilities in patients who traditionally have intermittent static vital sign measurements.

Discrete Heart Rate Values or Continuous Streams? Representation, Variability, and Meaningful Use of Vital Sign Data

Documentation and review of patient heart rate are a fundamental process across a myriad of clinical settings. While historically recorded manually, bedside monitors now provide for the automated collection of such data. Despite the availability of continuous streaming data, patients' charts continue to reflect only a subset of this information as snapshots recorded throughout a hospitalization. Over the past decade, prominent works have explored the implications of such practices and established fundamental differences in the alignment of discrete charted vitals and steaming data captured by monitoring systems. Limited work has examined the temporal properties of these differences, how they manifest, and their relation to clinical applications. The work presented in this article addresses this disparity, providing evidence that differences between charting techniques extend to measures of variability. Our results demonstrate how variability manifests with respect to temporal elements of charting timing and how it can facilitate personalized care by contextualizing deviations in magnitude. This work also highlights the utility of variability metrics with relation to clinical measures including associations to severity scores and a case study utilizing complex variability metrics derived from the complete set of monitor data.

Early Detection of Hypotension Using a Multivariate Machine Learning Approach

INTRODUCTION

The ability to accurately detect hypotension in trauma patients at the earliest possible time is important in improving trauma outcomes. The earlier an accurate detection can be made, the more time is available to take corrective action. Currently, there is limited research on combining multiple physiological signals for an early detection of hemorrhagic shock. We studied the viability of early detection of hypotension based on multiple physiologic signals and machine learning methods. We explored proof of concept with a small (5 minutes) prediction window for application of machine learning tools and multiple physiologic signals to detecting hypotension.

MATERIALS AND METHODS

Multivariate physiological signals from a preexisting dataset generated by an experimental hemorrhage model were employed. These experiments were conducted previously by another research group and the data made available publicly through a web portal. This dataset is among the few publicly available which incorporate measurement of multiple physiological signals from large animals during experimental hemorrhage. The data included two hemorrhage studies involving eight sheep. Supervised machine learning experiments were conducted in order to develop deep learning (viz., long short-term memory or LSTM), ensemble learning (viz., random forest), and classical learning (viz., support vector machine or SVM) models for the identification of physiological signals that can detect whether or not overall blood loss exceeds a predefined threshold 5 minutes ahead of time. To evaluate the performance of the machine learning technologies, 3-fold cross-validation was conducted and precision (also called positive predictive value) and recall (also called sensitivity) values were compared. As a first step in this development process, 5 minutes prediction windows were utilized.

RESULTS

The results showed that SVM and random forest outperform LSTM neural networks, likely because LSTM tends to overfit the data on small sized datasets. Random forest has the highest recall (84%) with 56% precision while SVM has 62% recall with 82% precision. Upon analyzing the feature importance, it was observed that electrocardiogram has the highest significance while arterial blood pressure has the least importance among all other signals.

CONCLUSION

In this research, we explored the viability of early detection of hypotension based on multiple signals in a preexisting animal hemorrhage dataset. The results show that a multivariate approach might be more effective than univariate approaches for this detection task.

Multimodal tensor-based method for integrative and continuous patient monitoring during postoperative cardiac care

Patients recovering from cardiovascular surgeries may develop life-threatening complications such as hemodynamic decompensation, making the monitoring of patients for such complications an essential component of postoperative care. However, this need has given rise to an inexorable increase in the number and modalities of data points collected, making it challenging to effectively analyze in real time. While many algorithms exist to assist in monitoring these patients, they often lack accuracy and specificity, leading to alarm fatigue among healthcare practitioners. In this study we propose a multimodal approach that incorporates salient physiological signals and EHR data to predict the onset of hemodynamic decompensation. A retrospective dataset of patients recovering from cardiac surgery was created and used to train predictive models. Advanced signal processing techniques were employed to extract complex features from physiological waveforms, while a novel tensor-based dimensionality reduction method was used to reduce the size of the feature space. These methods were evaluated for predicting the onset of decompensation at varying time intervals, ranging from a half-hour to 12 h prior to a decompensation event. The best performing models achieved AUCs of 0.87 and 0.80 for the half-hour and 12-h intervals respectively. These analyses evince that a multimodal approach can be used to develop clinical decision support systems that predict adverse events several hours in advance.

Analysis of Risk Factors for Adverse Cardiovascular Events in Elderly Patients with Acute Myocardial Infarction and Non-Alcoholic Fatty Liver Disease (NAFLD)

Background

The present research aimed to explore the risk factors for adverse cardiovascular events in elderly patients with acute myocardial infarction (AMI) combined with NAFLD.

Material/Methods

We included 325 AMI patients hospitalized in the Department of Cardiology. AMI patients underwent emergency thrombolysis or percutaneous coronary intervention (PCI). AMI patients were classified into NAFLD group and non-NAFLD group. General clinical data, creatinine and myocardial enzyme, GRACE scores of AMI patients were evaluated and compared between two groups. Incidence of adverse cardiovascular events, including ECG instability, hemodynamic instability and death were evaluated.

Results

Compared to patients in the non-NAFLD group, patients in the NAFLD group had remarkably lower proportions of diabetic patients (p=0.001), coronary heart disease (CHD) patients (p=0.027), and CABG/PCI patients (p<0.001), and had significantly higher EF values (p=0.042). Meanwhile, the proportion of adverse cardiovascular events (ECG instability (p<0.001), hemodynamic instability (p=0.033), and deaths (p=0.016)) in patients in the NAFLD group was significantly higher compared to patients in the non-NAFLD group. Multivariate logistic regression analysis showed that GRACE score >140 (OR: 3.005, 95% CI: 1.504–6.032), EF <35% (OR: 2.649, 95% CI: 1.364–4.346), diabetes (OR: 1.308, 95% CI: 1.072–1.589), and NAFLD (OR: 1.112, 95% CI: 1.043–1.324) were independent predictors for elderly AMI patients’ adverse cardiovascular events.

Conclusions

The risk for adverse cardiovascular events in elderly acute myocardial infarction patients who also had NAFLD was significantly higher. Therefore, strengthening monitoring and active treatment for elderly AMI patients who also have NAFLD could reduce the incidence of adverse cardiovascular events and improve survival rate prognosis.

Preprocessing Sequence Coverage Data for More Precise Detection of Copy Number Variations

Copy number variation (CNV) is a type of genomic/genetic variation that plays an important role in phenotypic diversity, evolution, and disease susceptibility. Next generation sequencing (NGS) technologies have created an opportunity for more accurate detection of CNVs with higher resolution. However, efficient and precise detection of CNVs remains challenging due to high levels of noise and biases, data heterogeneity, and the "big data" nature of NGS data. Sequence coverage (readcount) data are mostly used for detecting CNVs, specially for whole exome sequencing data. Readcount data are contaminated with several types of biases and noise that hinder accurate detection of CNVs. In this work, we introduce a novel preprocessing pipeline for reducing noise and biases to improve the detection accuracy of CNVs in heterogeneous NGS data, such as cancer whole exome sequencing data. We have employed several normalization methods to reduce readcount's biases that are due to GC content of reads, read alignment problems, and sample impurity. We have also developed a novel efficient and effective smoothing approach based on Taut String to reduce noise and increase CNV detection power. Using simulated and real data we showed that employing the proposed preprocessing pipeline significantly improves the accuracy of CNV detection.

Evaluation of Pulse Arrival Times during Lower Body Negative Pressure Test for the Non-Invasive Detection of Hypovolemia

The early detection of occult bleeding is a difficult problem for clinicians because physiological variables such as heart rate and blood pressure that are measured with standard patient monitoring equipment are insensitive to blood loss. In this study, the pulse arrival time (PAT) was investigated as an easily recorded, non-invasive indicator of hypovolemia. A lower body negative pressure (LBNP) study with a stepwise increase of negative pressure was conducted to induce central hypovolemia in a study population of 30 subjects. PAT values were extracted from simultaneous recordings of the electrocardiogram (ECG) and photoplethysmographic (PPG) recordings both from the index finger and from within the outer ear canal. Stroke volume (SV) was recorded as a reference measure by transthoracic echocardiography. An inter- and intra-individual correlation analysis between changes in SV and the PAT measurements was performed. Furthermore, it was assessed if PAT measurements can indicate a diminished SV in this scenario. It could be demonstrated that the measured PAT values are significantly increased at the lowest LBNP pressure level. A very strong intra-individual correlation (ρ ≥ 0.8) and a moderate inter-individual correlation (ρ ≥ 0.5) between PAT and SV measurements were found. Thus, PAT measurements could be a viable tool to monitor patient specific volemic trends. Further research is needed to investigate if PAT information can be utilized for a more robust inter-subject quantification of the degree of hypovolemia.

Noise cancellation using total variation for copy number variation detection

BACKGROUND

Due to recent advances in sequencing technologies, sequence-based analysis has been widely applied to detecting copy number variations (CNVs). There are several techniques for identifying CNVs using next generation sequencing (NGS) data, however methods employing depth of coverage or read depth (RD) have recently become a main technique to identify CNVs. The main assumption of the RD-based CNV detection methods is that the readcount value at a specific genomic location is correlated with the copy number at that location. However, readcount data's noise and biases distort the association between the readcounts and copy numbers. For more accurate CNV identification, these biases and noise need to be mitigated. In this work, to detect CNVs more precisely and efficiently we propose a novel denoising method based on the total variation approach and the Taut String algorithm.

RESULTS

To investigate the performance of the proposed denoising method, we computed sensitivities, false discovery rates and specificities of CNV detection when employing denoising, using both simulated and real data. We also compared the performance of the proposed denoising method, Taut String, with that of the commonly used approaches such as moving average (MA) and discrete wavelet transforms (DWT) in terms of sensitivity of detecting true CNVs and time complexity. The results show that Taut String works better than DWT and MA and has a better power to identify very narrow CNVs. The ability of Taut String denoising in preserving CNV segments' breakpoints and narrow CNVs increases the detection accuracy of segmentation algorithms, resulting in higher sensitivities and lower false discovery rates.

CONCLUSIONS

In this study, we proposed a new denoising method for sequence-based CNV detection based on a signal processing technique. Existing CNV detection algorithms identify many false CNV segments and fail in detecting short CNV segments due to noise and biases. Employing an effective and efficient denoising method can significantly enhance the detection accuracy of the CNV segmentation algorithms. Advanced denoising methods from the signal processing field can be employed to implement such algorithms. We showed that non-linear denoising methods that consider sparsity and piecewise constant characteristics of CNV data result in better performance in CNV detection.

Surgical data science: The new knowledge domain

Healthcare in general, and surgery/interventional care in particular, is evolving through rapid advances in technology and increasing complexity of care, with the goal of maximizing the quality and value of care. Whereas innovations in diagnostic and therapeutic technologies have driven past improvements in the quality of surgical care, future transformation in care will be enabled by data. Conventional methodologies, such as registry studies, are limited in their scope for discovery and research, extent and complexity of data, breadth of analytical techniques, and translation or integration of research findings into patient care. We foresee the emergence of surgical/interventional data science (SDS) as a key element to addressing these limitations and creating a sustainable path toward evidence-based improvement of interventional healthcare pathways. SDS will create tools to measure, model, and quantify the pathways or processes within the context of patient health states or outcomes and use information gained to inform healthcare decisions, guidelines, best practices, policy, and training, thereby improving the safety and quality of healthcare and its value. Data are pervasive throughout the surgical care pathway; thus, SDS can impact various aspects of care, including prevention, diagnosis, intervention, or postoperative recovery. The existing literature already provides preliminary results, suggesting how a data science approach to surgical decision-making could more accurately predict severe complications using complex data from preoperative, intraoperative, and postoperative contexts, how it could support intraoperative decision-making using both existing knowledge and continuous data streams throughout the surgical care pathway, and how it could enable effective collaboration between human care providers and intelligent technologies. In addition, SDS is poised to play a central role in surgical education, for example, through objective assessments, automated virtual coaching, and robot-assisted active learning of surgical skill. However, the potential for transforming surgical care and training through SDS may only be realized through a cultural shift that not only institutionalizes technology to seamlessly capture data but also assimilates individuals with expertise in data science into clinical research teams. Furthermore, collaboration with industry partners from the inception of the discovery process promotes optimal design of data products as well as their efficient translation and commercialization. As surgery continues to evolve through advances in technology that enhance delivery of care, SDS represents a new knowledge domain to engineer surgical care of the future.