Sunil B. Nagaraj

Automatic Classification of Sedation Levels in Icu Patients Using Heart Rate Variability.

Objective: To explore the potential value of heart rate variability features for automated monitoring of sedation levels in mechanically ventilated ICU patients. Design: Multicenter, pilot study. Setting: Several ICUs at Massachusetts General Hospital, Boston, MA. Patients: Electrocardiogram recordings from 40 mechanically ventilated adult patients receiving sedatives in an ICU setting were used to develop and test the proposed automated system. Measurements and Main Results: Richmond Agitation-Sedation Scale scores were acquired prospectively to assess patient sedation levels and were used as ground truth. Richmond Agitation-Sedation Scale scores were grouped into four levels, denoted “unarousable” (Richmond Agitation- Sedation Scale = –5, –4), “sedated” (–3, –2, –1), “awake” (0), “agitated” (+1, +2, +3, +4). A multiclass support vector machine algorithm was used for classification. Classifier training and performance evaluations were carried out using leave-oneout cross validation. An overall accuracy of 69% was achieved for discriminating between the four levels of sedation. The proposed system was able to reliably discriminate (accuracy = 79%) between sedated (Richmond Agitation-Sedation Scale < 0) and nonsedated states (Richmond Agitation-Sedation Scale > 0). Conclusions: With further refinement, the methodology reported herein could lead to a fully automated system for depth of sedation monitoring. By enabling monitoring to be continuous, such technology may help clinical staff to monitor sedation levels more effectively and to reduce complications related to over- and undersedation.

The Neu-Prem Trial: Neuromonitoring of Brains of Infants Born Preterm During Resuscitation—A Prospective Observational Cohort Study

Objective To determine whether monitoring cerebral oxygen tissue saturation (StO2) with near‐infrared spectroscopy (NIRS) and brain activity with amplitude‐integrated electroencephalography (aEEG) can predict infants at risk for intraventricular hemorrhage (IVH) and death in the first 72 hours of life. Study design A NIRS sensor and electroencephalography leads were placed on 127 newborns <32 weeks of gestational age at birth. Ten minutes of continuous NIRS and aEEG along with heart rate, peripheral arterial oxygen saturation, fraction of inspired oxygen, and mean airway pressure measurements were obtained in the delivery room. Once the infant was transferred to the neonatal intensive care unit, NIRS, aEEG, and vital signs were recorded until 72 hours of life. An ultrasound scan of the head was performed within the first 12 hours of life and again at 72 hours of life. Results Thirteen of the infants developed any IVH or died; of these, 4 developed severe IVH (grade 3‐4) within 72 hours. There were no differences in either cerebral StO2 or aEEG in the infants with low‐grade IVH. Infants who developed severe IVH or death had significantly lower cerebral StO2 from 8 to 10 minutes of life. Conclusions aEEG was not predictive of IVH or death in the delivery room or in the neonatal intensive care unit. It may be possible to use NIRS in the delivery room to predict severe IVH and early death. Trial registration ClinicalTrials.gov: NCT02605733.

ADARRI: a novel method to detect spurious R-peaks in the electrocardiogram for heart rate variability analysis in the intensive care unit

We developed a simple and fully automated method for detecting artifacts in the R-R interval (RRI) time series of the ECG that is tailored to the intensive care unit (ICU) setting. From ECG recordings of 50 adult ICU-subjects we selected 60 epochs with valid R-peak detections and 60 epochs containing artifacts leading to missed or false positive R-peak detections. Next, we calculated the absolute value of the difference between two adjacent RRIs (adRRI), and obtained the empirical probability distributions of adRRI values for valid R-peaks and artifacts. From these, we calculated an optimal threshold for separating adRRI values arising from artifact versus non-artefactual data. We compared the performance of our method with the methods of Berntson and Clifford on the same data. We identified 257,458 R-peak detections, of which 235,644 (91.5%) were true detections and 21,814 (8.5%) arose from artifacts. Our method showed superior performance for detecting artifacts with sensitivity 100%, specificity 99%, precision 99%, positive likelihood ratio of 100 and negative likelihood ratio <0.001 compared to Berntson’s and Clifford’s method with a sensitivity, specificity, precision and positive and negative likelihood ratio of 99%, 78%, 82%, 4.5, 0.013 for Berntson’s method and 55%, 98%, 96%, 27.5, 0.460 for Clifford’s method, respectively. A novel algorithm using a patient-independent threshold derived from the distribution of adRRI values in ICU ECG data identifies artifacts accurately, and outperforms two other methods in common use. Furthermore, the threshold was calculated based on real data from critically ill patients and the algorithm is easy to implement.

The Revised Cerebral Recovery Index Improves Predictions of Neurological Outcome after Cardiac Arrest

OBJECTIVE Analysis of the electroencephalogram (EEG) background pattern helps predicting neurological outcome of comatose patients after cardiac arrest (CA). Visual analysis may not extract all discriminative information. We present predictive values of the revised Cerebral Recovery Index (rCRI), based on continuous extraction and combination of a large set of evolving quantitative EEG (qEEG) features and machine learning techniques. METHODS We included 551 subsequent patients from a prospective cohort study on continuous EEG after CA in two hospitals. Outcome at six months was classified as good (Cerebral Performance Category (CPC) 1-2) or poor (CPC 3-5). Forty-four qEEG features (from time, frequency and entropy domain) were selected by the least absolute shrinkage and selection operator (LASSO) method and used in a Random Forests classification system. We trained and evaluated the system with 10-fold cross validation. For poor outcome prediction, the sensitivity at 100% specificity (Se100) and the area under the receiver operator curve (AUC) were used as performance of the prediction model. For good outcome, we used the sensitivity at 95% specificity (Se95). RESULTS Two hundred fifty-six (47%) patients had a good outcome. The rCRI predicted poor outcome with AUC = 0.94 (95% CI: 0.83-0.91), Se100 = 0.66 (0.65-0.78), and AUC = 0.88 (0.78-0.93), Se100 = 0.60 (0.51-0.75) at 12 and 24 h after CA, respectively. The rCRI predicted good outcome with Se95 = 0.72 (0.61-0.85) and 0.40 (0.30-0.51) at 12 and 24 h after CA, respectively. CONCLUSIONS Results obtained in this study suggest that with machine learning algorithms and large set of qEEG features, it is possible to efficiently monitor patient outcome after CA. We also demonstrate the importance of selection of optimal performance metric to train a classifier model for outcome prediction. SIGNIFICANCE The rCRI is a sensitive, reliable predictor of neurological outcome of comatose patients after CA.

S87. Sedation prediction using EEG spectrogram in ICU patients

Introduction Most mechanically ventilated ICU patients receive sedatives to relieve pain and anxiety, and to provide cardiopulmonary stability. Unfortunately, both excessive and insufficient sedation are common. Brain monitors that track electroencephalogram (EEG) features have been proposed as a real-time, physiologically-based alternative to clinical sedation assessments. However, existing monitors have been tested almost exclusively in the surgical setting, without being optimized for ICU patients. Methods Patients: We analyzed prospectively collected data from 115 mechanically ventilated patients receiving usual ICU care. The Richmond agitation sedation scale (RASS), assessed every 2 h, provides reference sedation levels. In the present work, we consider only assessments with RASS −5 and −4 (deeply sedated) vs −1 and 0 (not sedated). In total, there are 664 RASS assessments. The dataset is split into 69 training, 23 validation and 23 testing patients. Label denoising: RASS assessments are sometimes recorded after a delay or in anticipation of a change in the level of consciousness following adjustment in sedative infusion rate. To reduce such “annotation noise”, before training a classifier we first “denoise” EEG segments whose spectra have a different label than the 10 most similar training segments.Classifier training and testing: We extract EEG power spectra from the 10 min period preceding each RASS assessment, using 10s windows sliding windows spaced 2s apart. The sequence of spectra is used to train a recurrent neural network (LSTM), which performs binary classification (RASS −5 and −4 vs −1 and 0). Performance is measured by area the under the receiver operator curve (AUC). The reported results are the average performance on the testing set from 10 random splits of patients. Strict separation of training and validation data from testing data is maintained throughout all experiments. Results The label denoising procedure alters 15% of RASS scores. The AUC in the testing set is 0.91 (SD 0.02). Visualization of the EEG spectrograms reveals lower total power and higher relative delta power for episodes of RASS −5 and −4; and higher total power and higher relative beta power for RASS −1 and 0. Conclusion Despite heterogeneous medical conditions and varying severity of medical illness in our ICU cohort, our model learns to accurately discriminate deep sedation (RASS −5 and −4) from the awake state (RASS −1and 0). The classifier achieves AUC at 0.91 in a patient-independent manner. The use of recurrent network architecture allows our model to take advantage of long-range temporal information in the EEG, and will allow the extension to take pharmacokinetics and pharmacodynamics information into account, which may further enhance performance and robustness.

F91. Early quantitative EEG reactivity predicts 6-month outcome in hypoxic ischemic brain injury

Introduction Early EEG background reactivity is a strong predictor of neurological recovery after hypoxic-ischemic brain injury despite hypothermia and sedation. Unfortunately, expert interrater-agreement on visual scoring of EEG background reactivity ranges from 24% to 50%. Recent studies indicate that machine-learning approaches using quantitative EEG (QEEG) might yield equivalent or superior performance to current EEG reactivity assessment practices, however its ability to predict outcomes has not been tested. We hypothesized that a QEEG reactivity method can predict long-term functional outcome in hypoxic ischemic brain injury. Methods We retrospectively reviewed clinical and EEG data of cardiac arrest patients managed with hypothermia at two university hospitals. EEG reactivity was tested daily using a structured exam consisting of auditory, tactile, and visual stimulation. Our quantitative EEG method evaluated changes in EEG spectra, entropy, and frequency features during 30 s before and after each stimulation-step (30 QEEG features used). Only the first EEG reactivity assessment for each subject was used in the final analysis. Good outcome was defined as Cerebral Performance Category of 1–2 at six months. A penalized multinomial logistic regression was utilized for feature selection and a random-forest classifier was employed in the training and validation sets. Model performance evaluation metric was the area under the curve. Results Outcome and EEG data was available for a total 77 subjects, and 30 cases were excluded due to presence of burst-suppression, periodic epileptiform discharges, or EEG artifact. Forty-seven subjects were included in the final analysis. Mean age was 57.8 (standard deviation 17.9) years and 29.8% had good outcome. The combination of four features provided best outcome prediction performance with an AUC of 0.87 (Kolmogorov-Smirnov test, skewness, Two-group test, and Renyi entropy). Conclusion Early QEEG reactivity is predictive of good outcome at six months. A quantitative approach to EEG reactivity analysis might facilitate accurate and individualized prognostication in hypoxic-ischemic brain injury.

S85. Automated quantitative EEG reactivity testing using actigraphy monitoring in cardiac arrest coma prognostication

Introduction EEG background reactivity is a strong predictor of coma recovery after cardiac arrest. The current clinical value of EEG background reactivity testing is limited by inadequate inter-rater expert agreement, few number of daily assessments, and unsuitability for quantitative tracking by visual review. We hypothesized that a quantitative EEG reactivity method using actigraphy-triggered events from a wrist-worn wearable can predict long-term functional outcome in comatose cardiac arrest patients treated with targeted temperature management. Methods We prospectively recorded clinical, actigraphy, and EEG data of comatose cardiac arrest patients managed with targeted temperature management at a single U.S. academic hospital. Continuous actigraphy data obtained from a wrist-worn wearable (Affectiva 3) was synchronized to continuous EEG data for up to 96 h after cardiac arrest. Change in actigraphy across time was used as a surrogate of bedside external physical stimulation (i.e. actigraphy-triggered event). Our quantitative EEG method evaluated changes in EEG spectra, entropy, and frequency features during 30 s before and 30 s after each actigraphy-triggered event (46 EEG features used). Actigraphy threshold and EEG window duration (pre and post actigraphy-triggered events) were utilized as hyperparameters in the model. In addition to actigraphy-triggered quantitative EEG background reactivity, EEG reactivity was scored visually as present or absent once daily by a single expert electroencephalographer during clinical care. Good outcome was defined as Cerebral Performance Category of 1–2 at six months. Classification was carried out using extreme learning machine and leave-one-subject-out cross validation. Final outcome class prediction for each individual subject was determined using simple majority voting across all actigraphy-triggered events. Results Ten subjects were monitored. Mean age was 60 (standard deviation (SD) ± 14.6) years, 10% were female, 50% had a shockable rhythm, and 40% had good outcome at six months. A total of 19 (SD ± 8.9) actigraphy-triggered events per patient were detected. Best outcome prediction performance was achieved using 50-s-long EEG window (25-s pre and 25-s post each actigraphy-triggered event). The quantitative EEG reactivity method correctly predicted good outcome in 80% of cases (sensitivity and specificity 80%). Expert rating using once daily visual EEG reactivity assessments had 60% accuracy for good outcome prediction. Conclusion Quantitative EEG reactivity assessments using actigraphy-based triggered events is feasible and may support long-term outcome prediction in cardiac arrest coma. Integration of EEG and wearable sensors using machine learning methods might facilitate the deployment of dynamic multi-modal patient monitoring in the intensive care unit environment.

Heart rate variability as a biomarker for sedation depth estimation in ICU patients

An automated patient-specific system to classify the level of sedation in ICU patients using heart rate variability signal is presented in this paper. ECG from 70 mechanically ventilated adult patients with administered sedatives in an ICU setting were used to develop a support vector machine based system for sedation depth monitoring using several heart rate variability measures. A leave-one-subject-out cross validation was used for classifier training and performance evaluations. The proposed patient-specific system provided a sensitivity, specificity and an AUC of 64%, 84.8% and 0.72, respectively. It is hoped that with the help of additional physiological signals the proposed patient-specific sedation level prediction system could lead to a fully automated multimodal system to assist clinical staff in ICUs to interpret the sedation level of the patient.

Clustering analysis to identify distinct spectral components of encephalogram burst suppression in critically ill patients

Millions of patients are admitted each year to intensive care units (ICUs) in the United States. A significant fraction of ICU survivors develop life-long cognitive impairment, incurring tremendous financial and societal costs. Delirium, a state of impaired awareness, attention and cognition that frequently develops during ICU care, is a major risk factor for post-ICU cognitive impairment. Recent studies suggest that patients experiencing electroencephalogram (EEG) burst suppression have higher rates of mortality and are more likely to develop delirium than patients who do not experience burst suppression. Burst suppression is typically associated with coma and deep levels of anesthesia or hypothermia, and is defined clinically as an alternating pattern of high-amplitude “burst” periods interrupted by sustained low-amplitude “suppression” periods. Here we describe a clustering method to analyze EEG spectra during burst and suppression periods. We used this method to identify a set of distinct spectral patterns in the EEG during burst and suppression periods in critically ill patients. These patterns correlate with level of patient sedation, quantified in terms of sedative infusion rates and clinical sedation scores. This analysis suggests that EEG burst suppression in critically ill patients may not be a single state, but instead may reflect a plurality of states whose specific dynamics relate to a patients underlying brain function.