Liangyou Chen

Decision tool for the early diagnosis of trauma patient hypovolemia

We present a classifier for use as a decision assist tool to identify a hypovolemic state in trauma patients during helicopter transport to a hospital, when reliable acquisition of vital-sign data may be difficult. The decision tool uses basic vital-sign variables as input into linear classifiers, which are then combined into an ensemble classifier. The classifier identifies hypovolemic patients with an area under a receiver operating characteristic curve (AUC) of 0.76 (standard deviation 0.05, for 100 randomly-reselected patient subsets). The ensemble classifier is robust; classification performance degrades only slowly as variables are dropped, and the ensemble structure does not require identification of a set of variables for use as best-feature inputs into the classifier. The ensemble classifier consistently outperforms best-features-based linear classifiers (the classification AUC is greater, and the standard deviation is smaller, p<0.05). The simple computational requirements of ensemble classifiers will permit them to function in small fieldable devices for continuous monitoring of trauma patients.

Automated beat onset and peak detection algorithm for field-collected photoplethysmograms

Recent reports suggest that photoplethysmo-graphy (PPG), which is a component of routine pulse oximetry, may be useful for detecting hypovolemia. An essential step in extracting and analyzing common PPG features is the robust identification of onset and peak locations of the vascular beats, despite varying beat morphologies and major oscillations in the baseline. Some prior reports used manual analysis of the PPG waveform; however, for systematic widespread use, an automated method is required. In this paper, we report an algorithm that automatically detects beat onsets and peaks from noisy field-collected PPG waveforms. We validated the algorithm by clinician evaluation of 100 randomly selected PPG waveform samples. For 99% of the beats, the algorithm was able to credibly identify the onsets and peaks of vascular beats, although the precise locations were ambiguous, given the very noisy data from actual clinical operations. The algorithm appears promising, and future consideration of its diagnostic capabilities and limitations is warranted.

The association between vital signs and major hemorrhagic injury is significantly improved after controlling for sources of measurement variability

PURPOSE Measurement error and transient variability affect vital signs. These issues are inconsistently considered in published reports and clinical practice. We investigated the association between major hemorrhagic injury and vital signs, successively applying analytic techniques that excluded unreliable measurements, reduced transient variation, and then controlled for ambiguity in individual vital signs through multivariate analysis. METHODS Vital sign data from 671 adult prehospital trauma patients were analyzed retrospectively. Computer algorithms were used to identify and exclude unreliable data and to apply time averaging. An ensemble classifier was developed and tested by cross-validation. Primary outcome was hemorrhagic injury plus red cell transfusion. Areas under receiver operating characteristic curves (ROC AUCs) were compared by the test of DeLong et al. RESULTS Of initial vital signs, systolic blood pressure (BP) had the highest ROC AUC of 0.71 (95% confidence interval, 0.64-0.78). The ROC AUCs improved after excluding unreliable data, significantly for heart rate and respiratory rate but not significantly for BP. Time averaging to reduce temporal variability further increased AUCs, significantly for BP and not significantly for heart rate and respiratory rate. The ensemble classifier yielded a final ROC AUC of 0.84 (95% confidence interval, 0.80-0.89) in cross-validation. CONCLUSIONS Techniques to reduce variability in vital sign data can lead to significantly improved diagnostic performance. Failure to consider such variability could significantly reduce clinical effectiveness or confound research investigations.

Are Standard Diagnostic Test Characteristics Sufficient for the Assessment of Continual Patient Monitoring

Background. For diagnostic processes involving continual measurements from a single patient, conventional test characteristics, such as sensitivity and specificity, do not consider decision consistency, which might be a distinct, clinically relevant test characteristic. Objective. The authors investigated the performance of a decision-support classifier for the diagnosis of traumatic injury with blood loss, implemented with three different data-processing methods. For each method, they computed standard diagnostic test characteristics and novel metrics related to decision consistency and latency. Setting. Prehospital air ambulance transport. Patients. A total of 557 trauma patients. Design. Continually monitored vital-sign data from 279 patients (50%) were randomly selected for classifier development, and the remaining were used for testing. Three data-processing methods were evaluated over 16 min of patient monitoring: a 2-min moving window, time averaging, and postprocessing with the sequential probability ratio test (SPRT). Measurements. Sensitivity and specificity were computed. Consistency was quantified through cumulative counts of decision changes over time and the fraction of patients affected by false alarms. Latency was evaluated by the fraction of patients without a decision. Results. All 3 methods showed very similar final sensitivities and specificities. Yet, there were significant differences in terms of the fraction of patients affected by false alarms, decision changes through time, and latency. For instance, use of the SPRT led to a 75% reduction in the number of decision changes and a 36% reduction in the number of patients affected by false alarms, at the expense of 3% unresolved final decisions. Conclusion. The proposed metrics of decision consistency and decision latency provided additional information beyond what could be obtained from test sensitivity and specificity and are likely to be clinically relevant in some applications involving temporal decision making.

Diagnosis of Hemorrhage in a Prehospital Trauma Population Using Linear and Nonlinear Multiparameter Analysis of Vital Signs

In this study, we analyzed a dataset of time-series vital-signs data collected by standard Propaq travel monitor during helicopter transport of 898 civilian trauma casualties from the scene of injury to a receiving trauma center. The goals of the analysis are two fold. First, to determine which combination of the automatically-collected and -qualified vital signs provides the best discrimination between casualties with and without major hemorrhage. Second, to determine whether nonlinear classifiers provide improved discrimination over simpler, linear classifiers. Major hemorrhage is defined by the presence of injuries consistent with hemorrhage in casualties who received one or more units of blood. We randomly selected a subset of the casualties to train and test the classifiers with multiple combinations of the vital-signs variables, and used the area under the receiver operating characteristic curve (ROC AUC) as a decision metric. Based on the results of 100 simulations, we observe that: (i) the best two features obtained are systolic blood pressure and heart rate (mean AUC = 0.75 from a linear classifier), and (ii) the use of nonlinear classifiers does not improve discrimination. These results support earlier findings that the interaction of systolic blood pressure and heart rate is useful for the identification of trauma hemorrhage and that linear classifiers are adequate for many real-world applications.

Development and Validation of a Portable Platform for Deploying Decision-Support Algorithms in Prehospital Settings

BACKGROUND Advanced decision-support capabilities for prehospital trauma care may prove effective at improving patient care. Such functionality would be possible if an analysis platform were connected to a transport vital-signs monitor. In practice, there are technical challenges to implementing such a system. Not only must each individual component be reliable, but, in addition, the connectivity between components must be reliable. OBJECTIVE We describe the development, validation, and deployment of the Automated Processing of Physiologic Registry for Assessment of Injury Severity (APPRAISE) platform, intended to serve as a test bed to help evaluate the performance of decision-support algorithms in a prehospital environment. METHODS We describe the hardware selected and the software implemented, and the procedures used for laboratory and field testing. RESULTS The APPRAISE platform met performance goals in both laboratory testing (using a vital-sign data simulator) and initial field testing. After its field testing, the platform has been in use on Boston MedFlight air ambulances since February of 2010. CONCLUSION These experiences may prove informative to other technology developers and to healthcare stakeholders seeking to invest in connected electronic systems for prehospital as well as in-hospital use. Our experiences illustrate two sets of important questions: are the individual components reliable (e.g., physical integrity, power, core functionality, and end-user interaction) and is the connectivity between components reliable (e.g., communication protocols and the metadata necessary for data interpretation)? While all potential operational issues cannot be fully anticipated and eliminated during development, thoughtful design and phased testing steps can reduce, if not eliminate, technical surprises.

Exploiting the existence of temporal heart-rate patterns for the detection of trauma-induced hemorrhage

Unattended hemorrhage is a major source of mortality in trauma casualties. In this study, we explore a set of prehospital heart rate (HR) time-series data collected from 358 civilian casualties to examine whether temporal HR patterns can be used for automated hemorrhage identification. Continuous and reliable HR time series are fragmented into overlapping segments of 128 s, with a 118-s overlap between each two neighboring segments, which are projected into a wavelet coefficient space using the Haar wavelet function. A supervised nearest-neighbor clustering algorithm is developed to explore the existence of temporal HR patterns represented by the wavelet coefficients to discriminate casualties with and without (control) major hemorrhage. The clustering algorithm identifies 162 HR patterns. The most frequent pattern is observed in 11 (23%) hemorrhage and 16 (5%) control patients, which is a significant association (p<0.05, chi-square test). When the top 10 patterns are combined for hemorrhage detection, their sensitivity and specificity are 0.68 and 0.79, respectively, and when the top 20 patterns are used sensitivity increases to 0.77 and specificity decreases to 0.71.

LAM: A landscape matching algorithm for respiratory data alignment

Respiratory waveforms and their derived respiratory rate time-series data can become misaligned from each other when they are collected by vital signs monitors under sub-optimal field conditions. The monitor-provided waveforms and rates can be re-aligned by independently calculating respiratory rates from the waveforms and then aligning them with the monitor-provided rates. However, substantially different rates may be generated from the same waveform due to the presence of ambiguous breaths at noisy positions in the waveform. This paper reports a landscape matching (LAM) algorithm to align respiratory rate time-series data with the waveform that they are derived from by using rates that are calculated by different means. The algorithm exploits the intermittent matches between two respiratory rate time series to generate a matching score for an alignment. The best alignment exhibits the highest matching score. The alignment performance of the LAM algorithm is compared to that of a correlation matching (CM) algorithm using field-collected respiratory data. Alignment performance is evaluated by: (1) comparing the ability of the two algorithms to return a shifted waveform to its original, known position; and (2) comparing the percent of points that match between the monitor-provided and calculated respiratory rate time-series data after re-alignment. The LAM alignment algorithm outperforms the CM algorithm in both comparisons at a statistically significant level (p<0.05). Out of 67 samples with shifted time-series data, on average, the LAM aligns respiratory rates within 44 seconds of the original position, which is significantly better the CM-calculated alignment (136 seconds). Out of 465 samples, the LAM performs better, worse, and equal to the CM algorithm in percentage of points matching in 73%, 11%, and 16% of the cases, respectively. This robust alignment algorithm supports the use of reliable post-hoc monitor-provided respiratory rates for data mining purposes.

Using confidence intervals to assess the reliability of instantaneous heart rate and respiratory rate

Physiological waveform signals collected from unstructured environments are noisy, requiring automated algorithms to assess the reliability of the derived vital signs, such as heart rate (HR) and respiratory rate (RR), before they can be used for automated decision support. We recently proposed a weighted regularized least squares method to estimate instantaneous HR (HR<inf>R</inf>), which readily provides analytically based confidence intervals (CIs). Accordingly, this method can be extended to the estimation of instantaneous RR (RR<inf>R</inf>). In this study, we aim to investigate whether we can use CIs to select reliable HR<inf>R</inf> and RR<inf>R</inf>. We calculated HR<inf>R</inf> and RR<inf>R</inf> for 532 and 370 trauma patients, respectively, grouped the rates according to their CIs, and investigated their reliability by determining their ability to diagnose major hemorrhage. The areas under a receiver operating characteristic curve of HR<inf>R</inf> and RR<inf>R</inf> with CI ≤ 5 bpm (beats per minute for HR and breaths per minute for RR) were 0.70 and 0.66, respectively. RR<inf>R</inf> was superior to the average output of the clinical monitor (p < 0.05 by DeLongs test), while HR<inf>R</inf> was equivalent. HR<inf>R</inf> and RR<inf>R</inf> provide a new approach to systematically and automatically assess the reliability of noisy, field-collected vital signs.

The matching of sinus arrhythmia to respiration: are trauma patients without serious injury comparable to healthy laboratory subjects?

We sought to better understand the physiology underlying the metrics of heart rate variability (HRV) in trauma patients without serious injury, compared to healthy laboratory controls. In trauma patients without serious injury (110 subjects, 470 2-min data segments), we studied the correlation between sinus arrhythmia (SA) rate, heart rate (HR), and respiratory rate (RR). Most segments with 2.4 <; HR/RR <; 4.8 exhibited SA-RR matching, whereas rate matching was absent in 81% of the segments with HR/RR <; 2.4 and in 86% of the segments with HR/RR > 4.8. The findings were comparable, in some cases remarkably so, to previous reports from healthy laboratory subjects. The presence (or absence) of SA-RR matching, when SA is largely controlled by respiration, can be anticipated in this trauma population. This work provides a valuable step towards the definition of patterns of HRV found in trauma patients with and without life-threatening injury.