Mohammad M. Ghassemi

An ICA with reference approach in identification of genetic variation and associated brain networks

To address the statistical challenges associated with genome-wide association studies, we present an independent component analysis (ICA) with reference approach to target a specific genetic variation and associated brain networks. First, a small set of single nucleotide polymorphisms (SNPs) are empirically chosen to reflect a feature of interest and these SNPs are used as a reference when applying ICA to a full genomic SNP array. After extracting the genetic component maximally representing the characteristics of the reference, we test its association with brain networks in functional magnetic resonance imaging (fMRI) data. The method was evaluated on both real and simulated datasets. Simulation demonstrates that ICA with reference can extract a specific genetic factor, even when the variance accounted for by such a factor is so small that a regular ICA fails. Our real data application from 48 schizophrenia patients (SZs) and 40 healthy controls (HCs) include 300K SNPs and fMRI images in an auditory oddball task. Using SNPs with allelic frequency difference in two groups as a reference, we extracted a genetic component that maximally differentiates patients from controls (p < 4 × 10−17), and discovered a brain functional network that was significantly associated with this genetic component (p < 1 × 10−4). The regions in the functional network mainly locate in the thalamus, anterior and posterior cingulate gyri. The contributing SNPs in the genetic factor mainly fall into two clusters centered at chromosome 7q21 and chromosome 5q35. The findings from the schizophrenia application are in concordance with previous knowledge about brain regions and gene function. All together, the results suggest that the ICA with reference can be particularly useful to explore the whole genome to find a specific factor of interest and further study its effect on brain.

Cognitive Tomography Reveals Complex, Task-Independent Mental Representations

Summary Humans develop rich mental representations that guide their behavior in a variety of everyday tasks. However, it is unknown whether these representations, often formalized as priors in Bayesian inference, are specific for each task or subserve multiple tasks. Current approaches cannot distinguish between these two possibilities because they cannot extract comparable representations across different tasks [1–10]. Here, we develop a novel method, termed cognitive tomography, that can extract complex, multidimensional priors across tasks. We apply this method to human judgments in two qualitatively different tasks, “familiarity” and “odd one out,” involving an ecologically relevant set of stimuli, human faces. We show that priors over faces are structurally complex and vary dramatically across subjects, but are invariant across the tasks within each subject. The priors we extract from each task allow us to predict with high precision the behavior of subjects for novel stimuli both in the same task as well as in the other task. Our results provide the first evidence for a single high-dimensional structured representation of a naturalistic stimulus set that guides behavior in multiple tasks. Moreover, the representations estimated by cognitive tomography can provide independent, behavior-based regressors for elucidating the neural correlates of complex naturalistic priors.

Optimal medication dosing from suboptimal clinical examples: A deep reinforcement learning approach

Misdosing medications with sensitive therapeutic windows, such as heparin, can place patients at unnecessary risk, increase length of hospital stay, and lead to wasted hospital resources. In this work, we present a clinician-in-the-loop sequential decision making framework, which provides an individualized dosing policy adapted to each patients evolving clinical phenotype. We employed retrospective data from the publicly available MIMIC II intensive care unit database, and developed a deep reinforcement learning algorithm that learns an optimal heparin dosing policy from sample dosing trails and their associated outcomes in large electronic medical records. Using separate training and testing datasets, our model was observed to be effective in proposing heparin doses that resulted in better expected outcomes than the clinical guidelines. Our results demonstrate that a sequential modeling approach, learned from retrospective data, could potentially be used at the bedside to derive individualized patient dosing policies.

A “datathon” model to support cross-disciplinary collaboration

A “datathon” model combines complementary knowledge and skills to formulate inquiries and drive research that addresses information gaps faced by clinicians. In recent years, there has been a growing focus on the unreliability of published biomedical and clinical research. To introduce effective new scientific contributors to the culture of health care, we propose a “datathon” or “hackathon” model in which participants with disparate, but potentially synergistic and complementary, knowledge and skills effectively combine to address questions faced by clinicians. The continuous peer review intrinsically provided by follow-up datathons, which take up prior uncompleted projects, might produce more reliable research, either by providing a different perspective on the study design and methodology or by replication of prior analyses.

Monitoring and detecting atrial fibrillation using wearable technology

Atrial fibrillation (AFib) is diagnosed by analysis of the morphological and rhythmic properties of the electrocardiogram. It was recently shown that accurate detection of AFib is possible using beat-to-beat interval variations. This raises the question of whether AFib detection can be performed using a pulsatile waveform such as the Photoplethysmogram (PPG). The recent explosion in use of recreational and professional ambulatory wrist-based pulse monitoring devices means that an accurate pulse-based AFib screening algorithm would enable large scale screening for silent or undiagnosed AFib, a significant risk factor for multiple diseases. We propose a noise-resistant machine learning approach to detecting AFib from noisy ambulatory PPG recorded from the wrist using a modern research watch-based wearable device (the Samsung Simband). Ambulatory pulsatile and movement data were recorded from 46 subjects, 15 with AFib and 31 non symptomatic. Single channel electrocardiogram (ECG), multi-wavelength PPG and tri-axial accelerometry were recorded simultaneously at 128 Hz from the non-dominant wrist using the Simband. Recording lengths varied from 3.5 to 8.5 minutes. Pulse (beat) detection was performed on the PPG waveforms, and eleven features were extracted based on beat-to-beat variability and waveform signal quality. Using 10-fold cross validation, an accuracy of 95 % on out-of-sample data was achieved, with a sensitivity of 97%, specificity of 94%, and an area under the receiver operating curve (AUROC) of 0.99. The described approach provides a noise-resistant, accurate screening tool for AFib from PPG sensors located in an ambulatory wrist watch. To our knowledge this is the first study to demonstrate an algorithm with a high enough accuracy to be used in general population studies that does not require an ambulatory Holter electrocardiographic monitor.

Accumulated Deep Sleep Is a Powerful Predictor of LH Pulse Onset in Pubertal Children

CONTEXT During puberty, reactivation of the reproductive axis occurs during sleep, with LH pulses specifically tied to deep sleep. This association suggests that deep sleep may stimulate LH secretion, but there have been no interventional studies to determine the characteristics of deep sleep required for LH pulse initiation. OBJECTIVE The objective of this study was to determine the effect of deep sleep fragmentation on LH secretion in pubertal children. DESIGN AND SETTING Studies were performed in a clinical research center. SUBJECTS Fourteen healthy pubertal children (11.3-14.1 y) participated in the study. INTERVENTIONS Subjects were randomized to two overnight studies with polysomnography and frequent blood sampling, with or without deep sleep disruption via auditory stimuli. RESULTS An average of 68.1 ±10.7 (± SE) auditory stimuli were delivered to interrupt deep sleep during the disruption night, limiting deep sleep to only brief episodes (average length disrupted 1.3 ± 0.2 min vs normal 7.1 ± 0.8 min, P < .001), and increasing the number of transitions between non-rapid eye movement (NREM), REM, and wake (disrupted 274.5 ± 33.4 vs normal 131.2 ± 8.1, P = .001). There were no differences in mean LH (normal: 3.2 ± 0.4 vs disrupted: 3.2 ± 0.5 IU/L), LH pulse frequency (0.6 ± 0.06 vs 0.6 ± 0.07 pulses/h), or LH pulse amplitude (2.8 ± 0.4 vs 2.8 ± 0.4 IU/L) between the two nights. Poisson process modeling demonstrated that the accumulation of deep sleep in the 20 minutes before an LH pulse, whether consolidated or fragmented, was a significant predictor of LH pulse onset (P < .001). CONCLUSION In pubertal children, nocturnal LH augmentation and pulse patterning are resistant to deep sleep fragmentation. These data suggest that, even when fragmented, deep sleep is strongly related to activation of the GnRH pulse generator.

A visualization of evolving clinical sentiment using vector representations of clinical notes

Our objective in this paper was to visualize the evolution of clinical language and sentiment with respect to several common population-level categories including: time in the hospital, age, mortality, gender and race. Our analysis utilized seven years of unstructured free text notes from the Multiparameter Intelligent Monitoring in Intensive Care (MIMIC) database. The text data was partitioned by category and used to generate several high dimensional vector space representations. We generated visualizations of the vector spaces using Distributed Stochastic Neighbor Embedding (tSNE) and Principal Component Analysis (PCA). We also investigated representative words from clusters in the vector space. Lastly, we inferred the general sentiment of the clinical notes toward each parameter by gauging the average distance between positive and negative keywords and all other terms in the space. We found intriguing differences in the sentiment of clinical notes over time, outcome, and demographic features. We noted a decrease in the homogeneity and complexity of clusters over time for patients with poor outcomes. We also found greater positive sentiment for females, unmarried patients, and patients of African ethnicity.

Patient prognosis from vital sign time series: Combining convolutional neural networks with a dynamical systems approach

In this work, we propose a stacked switching vector-autoregressive (SVAR)-CNN architecture to model the changing dynamics in physiological time series for patient prognosis. The SVAR-layer extracts dynamical features (or modes) from the time-series, which are then fed into the CNN-layer to extract higher-level features representative of transition patterns among the dynamical modes. We evaluate our approach using 8-hours of minute-by-minute mean arterial blood pressure (BP) from over 450 patients in the MIMIC-II database. We modeled the time-series using a third-order SVAR process with 20 modes, resulting in first-level dynamical features of size 20×480 per patient. A fully connected CNN is then used to learn hierarchical features from these inputs, and to predict hospital mortality. The combined CNN/SVAR approach using BP time-series achieved a median and interquartile-range AUC of 0.74 [0.69, 0.75], significantly outperforming CNN-alone (0.54 [0.46, 0.59]), and SVAR-alone with logistic regression (0.69 [0.65, 0.72]). Our results indicate that including an SVAR layer improves the ability of CNNs to classify nonlinear and nonstationary time-series.

Newsworthy Rumor Events: A Case Study of Twitter

Rumor events differ in how and where they originate, what topics they address, the emotions they invoke, and how they engage their audience. In this paper, we study various semantic aspects of rumors and analyze the motivational and functional roles they play. Using Twitter as a case study, we develop a framework to characterize rumors. Our characterization covers intrinsic and extrinsic factors, tweet and event-level, as well as usage analysis. We determine the roles various user-types play and analyze rumor propagation from both a re-tweeting and burstiness perspective.

A fast and memory-efficient algorithm for learning and retrieval of phenotypic dynamics in multivariate cohort time series

Robust navigation and mining of physiologic time series databases often requires finding similar temporal patterns of physiological responses. Detection of these complex physiological patterns not only enables demarcation of important clinical events but can also elucidate hidden dynamical structures that may be suggestive of disease processes. Some specific examples where this physiological signal search may be useful include real-time detection of cardiac arrhythmias, sleep staging or detection of seizure onset. In all these cases, being able to identify a cohort of patients who exhibit similar physiological dynamics could be useful in prognosis and informing treatment strategies. However, pattern recognition for physiological time series is complicated by changes between operating regimes and measurement artifacts. Here we briefly describe an approach we have developed for distributed identification of dynamical patterns in physiological time series using a switching linear dynamical system (SLDS). We present a fast and memory-efficient algorithm for learning and retrieval of phenotypic dynamics in large clinical time series databases. Through simulation we show that the proposed algorithm is at least an order of magnitude faster that the state of the art, and provide encouraging preliminary results based on real recordings of vital sign time series from the Multiparameter Intelligent Monitoring in Intensive Care (MIMIC-II) database.