Haoqi Sun

Driver Workload Detection in On-Road Driving Environment Using Machine Learning

Drivers’ high workload caused by distractions has become one of the major concerns for road safety. This paper presents a data-driven method using machine learning algorithms to detect high workload caused by surrogate in-vehicle (IV) secondary tasks performed in an on-road experiment with real traffic. The data were collected using an instrumented vehicle while drivers performed two types of secondary tasks: visual-manual and auditory-vocal tasks. Two types of machine learning methods, support vector machine (SVM) and extreme learning machine (ELM), were applied to detect drivers’ workload via drivers’ visual behaviour (i.e. eye movements) data alone, as well as visual plus driving performance data. The results suggested that both methods can detect drivers’ workload at high accuracy, with ELM outperformed SVM in most cases. We found that for visual intensive workload, using drivers’ visual data alone achieveed an accuracy close to using the combination information from both visual and driving performance data. This study proves that machine learning methods can be used for real driving applications.

Learning polychronous neuronal groups using joint weight-delay spike-timing-dependent plasticity

Polychronous neuronal group (PNG), a type of cell assembly, is one of the putative mechanisms for neural information representation. According to the reader-centric definition, some readout neurons can become selective to the information represented by polychronous neuronal groups under ongoing activity. Here, in computational models, we show that the frequently activated polychronous neuronal groups can be learned by readout neurons with joint weight-delay spike-timing-dependent plasticity. The identity of neurons in the group and their expected spike timing at millisecond scale can be recovered from the incoming weights and delays of the readout neurons. The detection performance can be further improved by two layers of readout neurons. In this way, the detection of polychronous neuronal groups becomes an intrinsic part of the network, and the readout neurons become differentiated members in the group to indicate whether subsets of the group have been activated according to their spike timing. The readout spikes representing this information can be used to analyze how PNGs interact with each other or propagate to downstream networks for higher-level processing.

Large-Scale Automated Sleep Staging

Study Objectives Automated sleep staging has been previously limited by a combination of clinical and physiological heterogeneity. Both factors are in principle addressable with large data sets that enable robust calibration. However, the impact of sample size remains uncertain. The objectives are to investigate the extent to which machine learning methods can approximate the performance of human scorers when supplied with sufficient training cases and to investigate how staging performance depends on the number of training patients, contextual information, model complexity, and imbalance between sleep stage proportions. Methods A total of 102 features were extracted from six electroencephalography (EEG) channels in routine polysomnography. Two thousand nights were partitioned into equal (n = 1000) training and testing sets for validation. We used epoch-by-epoch Cohens kappa statistics to measure the agreement between classifier output and human scorer according to American Academy of Sleep Medicine scoring criteria. Results Epoch-by-epoch Cohens kappa improved with increasing training EEG recordings until saturation occurred (n = ~300). The kappa value was further improved by accounting for contextual (temporal) information, increasing model complexity, and adjusting the model training procedure to account for the imbalance of stage proportions. The final kappa on the testing set was 0.68. Testing on more EEG recordings leads to kappa estimates with lower variance. Conclusion Training with a large data set enables automated sleep staging that compares favorably with human scorers. Because testing was performed on a large and heterogeneous data set, the performance estimate has low variance and is likely to generalize broadly.

Runtime detection of activated polychronous neuronal group towards its spatiotemporal analysis

Due to the precise spike timing in neural coding, spiking neural network (SNN) possesses richer spatiotemporal dynamics compared to neural networks with firing rate coding. One of the distinct features of SNN, polychronous neuronal group (PNG), receives much attention from both computational neuroscience and machine learning communities. However, all existing algorithms detect PNGs from the spike recording collected after simulation in an offline manner. There is currently no algorithm that detects PNGs actually being activated in runtime (online manner), which could be potentially used as inputs to higher level neural processing. We propose a runtime detection algorithm particularly for activated PNGs, using PNG readout neurons, to fill this gap. The proposed algorithm can reveal the spatiotemporal PNG patterns embedded in spike trains, which is higher level neuronal dynamics. We demonstrate through an example that for composed input patterns, new PNGs except the constituent PNGs can be easily found using the proposed algorithm. As an important interpretation, we give further insights on how to use PNG readout neurons to construct layered network structure.

Estimating vigilance from EEG using manifold clustering guided by instantaneous lapse rate

Vigilance decrement happens in prolonged and monotonous tasks such as driving, therefore efficient estimation of vigilance using machine learning becomes a growing research field in road safety. However, the ground truth of vigilance level is often unknown. To address the estimation of brain states with unknown ground truth, we proposed an unsupervised manifold clustering method guided by task performance, namely instantaneous lapse rate, without directly using any artificially labels, using electroencephalogram (EEG) as data source. The proposed algorithm utilizes information from both data structure and task performance, which is especially suitable for applications with unknown ground truth. Future research directions include using advanced manifold clustering algorithms to increase the robustness towards the high nonlinearity in the EEG feature space and the embedded space, as well as allowing the mapping from multiple clusters to one vigilance level.

S176. A Mean field model to study the triphasic waves in acute hepatic encephalopathy

Introduction Triphasic waves (TWs) are a distinctive electroencephalogram (EEG) pattern consisting of consisting ‘triphasic’ epileptiform discharges repeating at relatively regular intervals, most commonly in patients with acute, severe toxic/metabolic encephalopathy (TME). TWs share some similarities with EEG patterns seen in non-convulsive status epilepticus (NCSE), and show similar responses to benzodiazepine administration. Nevertheless the mechanisms underlying TWs are poorly understood, and their relationship to NCSE debated. Here we explore the conditions in a biologically plausible computational model of the EEG that lead to patterns resembling TWs, and attempt to relate these to mechanisms operative in one type of TME: acute hepatic encephalopathy (AHE). Methods Our work builds on an existing neural mean field model (NMFM) developed for how periodic discharges arise in severe cerebral anoxia (an adaptation of the bursting Liley model), through three types of processes: (a) metabolic failure leading to impaired synaptic transmission; and (b) increased neuronal excitability. In addition, the model accounts for effects of (c) GABA-ergic modulation by anesthetics. We identify pathophysiologic mechanisms involved in AHE which can lead to each of these effects, and relate these mechanisms to changes in the parameters of the NMFM. In this way our model relates “microscopic” (biochemical) mechanisms to “macroscopic” observations (EEG patterns). Then, we relate the simulated EEG patterns to clinical EEG data of patients with AHE through by constructing a comparison function, on the basis of different feature extraction methods, similarity measurements and optimization algorithms. Results The model is able to reproduce key qualitative features of EEG data from patients with AHE, including the approximate shape, irregular and slowed background, and semi-periodic pattern of TWs. In addition, we find that the dynamic evolution of EEG activity during AHE can be characterized through the change of four key parameters of proposed NMFM ((recovery time constants), (potentiation factor of EPSP), (decay rate of IPSP)), which reflect gradual changes in underlying physiological mechanisms. Conclusion Known alterations in cerebral physiology in acute hepatic encephalopathy are relatable to parameters of a biologically plausible mean-field model of the EEG. Our model represents a starting point for exploring the dynamical mechanisms underlying the EEG of severe encephalopathy.

S141. Predicting brain age from the electroencephalogram of sleep

Introduction The human electroencephalogram (EEG) of sleep undergoes profound changes with age, such as decreased sleep spindle amplitude and density in non-rapid eye movement stage 2 (NREM2). However, it is unknown how accurately a patient’s age can be predicted from EEG activity during sleep. A quantitative characterization of age-related EEG provides important insights into healthy aging. Moreover, the ability to detect deviations of EEG patterns from those typical for age could provide insights into age-related neurological disorders, and might provide a way to gauge the effects of interventions designed to preserve or improve brain health. Here we develop a model to predict a patient’s age based on large-scale and heterogeneous sleep EEG datasets. The prediction is called “brain age” (BA). Methods Datasets: (1) MGH sleep dataset: 3100 patients aged 18–80 years. (2) sleep-heart health study (SHHS): 3680 paired recordings aged 18–80 years, where each pair is recorded approximately 5 years apart from the same subject. This dataset is used to verify the longitudinal reliability of the brain age. EEG features: 102 features are extracted from 30s-epochs, and then averaged separately for the 5 sleep stages, yielding 510 features to summarize each patient’s overnight sleep. We also analyze medications and clinical variables to identify factors that help account for brain age being older or younger than chronological age. Results For 1000 testing patients from MGH dataset, the Pearson’s correlation between EEG-based brain age and chronological age is 0.86 (95% CI 0.84–0.87). The mean absolute prediction error (MAE) is 6.6 years. For SHHS dataset, training the model on a subset of 2000 records and testing on the other 1680 records achieves correlation at 0.71 (95% CI 0.69–0.73) and MAE 5.9 yrs. The average difference of BA between each pair is 4.8yrs. Training the entire MGH dataset and testing on SHHS achieves correlation at 0.61 (95% CI 0.59–0.63) and MAE 8.6 yrs. The average difference of BA between each pair is 3.7 yrs. In the MGH dataset, older brain age (predicted age greater than chronological age) is associated diabetes (Kruskal-Wallis test p-value 0.01) and weakly associated with wake time after sleep onset (Pearson’s correlation p-value 0.07). Conclusion Our results indicate that, at the population level, chronological age can be accurately predicted from overnight sleep EEG. Moreover, brain age accurately tracks chronological age. Further research is needed to characterize how EEG-based brain age relates to cognitive function and to what degree and by what means brain age is modifiable.

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.