Towards Interpretable Seizure Detection Using Wearables Al-Hussaini, Irfan; Mitchell, Cassie S.
ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)
2023
Conference Proceeding, Journal Article
Odprti dostop
Seizure detection using machine learning is a critical problem for the timely intervention and management of epilepsy. We propose SeizFt, a robust seizure detection framework using EEG from a ...wearable device. It uses features paired with an ensemble of trees, thus enabling further interpretation of the model's results. The efficacy of the underlying augmentation and class-balancing strategy is also demonstrated. This study was performed for the Seizure Detection Challenge 2023, an ICASSP Grand Challenge.
We perform a large-scale meta-analysis of 51 peer-reviewed 3xTg-AD mouse publications to compare Alzheimer's disease (AD) quantitative clinical outcome measures, including amyloid-β (Aβ), total tau, ...and phosphorylated tau (pTau), with cognitive performance in Morris water maze (MWM) and Novel Object Recognition (NOR). "High" levels of Aβ (Aβ40, Aβ42) showed significant but weak trends with cognitive decline (MWM: slope = 0.336, R2 = 0.149, n = 259, p < 0.001; NOR: slope = 0.156, R2 = 0.064, n = 116, p < 0.05); only soluble Aβ or directly measured Aβ meaningfully contribute. Tau expression in 3xTg-AD mice was within 10-20% of wild type and not associated with cognitive decline. In contrast, increased pTau is directly and significantly correlated with cognitive decline in MWM (slope = 0.408, R2 = 0.275, n = 371, p < < 0.01) and NOR (slope = 0.319, R2 = 0.176, n = 113, p < 0.05). While a variety of pTau epitopes (AT8, AT270, AT180, PHF-1) were examined, AT8 correlated most strongly with cognition (slope = 0.586, R2 = 0.521, n = 185, p < < 0.001). Multiple linear regression confirmed pTau is a stronger predictor of MWM performance than Aβ. Despite pTau's lower physical concentration than Aβ, pTau levels more directly and quantitatively correlate with 3xTg-AD cognitive decline. pTau's contribution to neurofibrillary tangles well after Aβ levels plateau makes pTau a viable treatment target even in late-stage clinical AD. Principal component analysis, which included hyperphosphorylation induced by kinases (pGSK3β, GSK3β, CDK5), identified phosphorylated ser9 GSK3β as the primary contributor to MWM variance. In summary, meta-analysis of cognitive decline in preclinical AD finds tauopathy more impactful than Aβ. Nonetheless, complex AD interactions dictate successful therapeutics harness synergy between Aβ and pTau, possibly through the GSK3 pathway.
The ability to translate Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) into different modalities and data types is essential to improve Deep Learning (DL) for predictive ...medicine. This work presents DACMVA, a novel framework to conduct data augmentation in a cross-modal dataset by translating between modalities and oversampling imputations of missing data. DACMVA was inspired by previous work on the alignment of latent spaces in Autoencoders. DACMVA is a DL data augmentation pipeline that improves the performance in a downstream prediction task. The unique DACMVA framework leverages a cross-modal loss to improve the imputation quality and employs training strategies to enable regularized latent spaces. Oversampling of augmented data is integrated into the prediction training. It is empirically demonstrated that the new DACMVA framework is effective in the often-neglected scenario of DL training on tabular data with continuous labels. Specifically, DACMVA is applied towards cancer survival prediction on tabular gene expression data where there is a portion of missing data in a given modality. DACMVA significantly (
<< 0.001, one-sided Wilcoxon signed-rank test) outperformed the non-augmented baseline and competing augmentation methods with varying percentages of missing data (4%, 90%, 95% missing). As such, DACMVA provides significant performance improvements, even in very-low-data regimes, over existing state-of-the-art methods, including TDImpute and oversampling alone.
This work presents SeizFt-a novel seizure detection framework that utilizes machine learning to automatically detect seizures using wearable SensorDot EEG data. Inspired by interpretable sleep ...staging, our novel approach employs a unique combination of data augmentation, meaningful feature extraction, and an ensemble of decision trees to improve resilience to variations in EEG and to increase the capacity to generalize to unseen data. Fourier Transform (FT) Surrogates were utilized to increase sample size and improve the class balance between labeled non-seizure and seizure epochs. To enhance model stability and accuracy, SeizFt utilizes an ensemble of decision trees through the CatBoost classifier to classify each second of EEG recording as seizure or non-seizure. The SeizIt1 dataset was used for training, and the SeizIt2 dataset for validation and testing. Model performance for seizure detection was evaluated using two primary metrics: sensitivity using the any-overlap method (OVLP) and False Alarm (FA) rate using epoch-based scoring (EPOCH). Notably, SeizFt placed first among an array of state-of-the-art seizure detection algorithms as part of the Seizure Detection Grand Challenge at the 2023 International Conference on Acoustics, Speech, and Signal Processing (ICASSP). SeizFt outperformed state-of-the-art black-box models in accurate seizure detection and minimized false alarms, obtaining a total score of 40.15, combining OVLP and EPOCH across two tasks and representing an improvement of ~30% from the next best approach. The interpretability of SeizFt is a key advantage, as it fosters trust and accountability among healthcare professionals. The most predictive seizure detection features extracted from SeizFt were: delta wave, interquartile range, standard deviation, total absolute power, theta wave, the ratio of delta to theta, binned entropy, Hjorth complexity, delta + theta, and Higuchi fractal dimension. In conclusion, the successful application of SeizFt to wearable SensorDot data suggests its potential for real-time, continuous monitoring to improve personalized medicine for epilepsy.
Parkinson's disease (PD) is a movement disorder caused by a dopamine deficit in the brain. Current therapies primarily focus on dopamine modulators or replacements, such as levodopa. Although ...dopamine replacement can help alleviate PD symptoms, therapies targeting the underlying neurodegenerative process are limited. The study objective was to use artificial intelligence to rank the most promising repurposed drug candidates for PD. Natural language processing (NLP) techniques were used to extract text relationships from 33+ million biomedical journal articles from PubMed and map relationships between genes, proteins, drugs, diseases, etc., into a knowledge graph. Cross-domain text mining, hub network analysis, and unsupervised learning rank aggregation were performed in SemNet 2.0 to predict the most relevant drug candidates to levodopa and PD using relevance-based HeteSim scores. The top predicted adjuvant PD therapies included ebastine, an antihistamine for perennial allergic rhinitis; levocetirizine, another antihistamine; vancomycin, a powerful antibiotic; captopril, an angiotensin-converting enzyme (ACE) inhibitor; and neramexane, an N-methyl-D-aspartate (NMDA) receptor agonist. Cross-domain text mining predicted that antihistamines exhibit the capacity to synergistically alleviate Parkinsonian symptoms when used with dopamine modulators like levodopa or levodopa-carbidopa. The relationship patterns among the identified adjuvant candidates suggest that the likely therapeutic mechanism(s) of action of antihistamines for combatting the multi-factorial PD pathology include counteracting oxidative stress, amending the balance of neurotransmitters, and decreasing the proliferation of inflammatory mediators. Finally, cross-domain text mining interestingly predicted a strong relationship between PD and liver disease.
Amyotrophic Lateral Sclerosis (ALS) is a paralyzing, multifactorial neurodegenerative disease with limited therapeutics and no known cure. The study goal was to determine which pathophysiological ...treatment targets appear most beneficial.
A big data approach was used to analyze high copy SOD1 G93A experimental data. The secondary data set comprised 227 published studies and 4,296 data points. Treatments were classified by pathophysiological target: apoptosis, axonal transport, cellular chemistry, energetics, neuron excitability, inflammation, oxidative stress, proteomics, or systemic function. Outcome assessment modalities included onset delay, health status (rotarod performance, body weight, grip strength), and survival duration. Pairwise statistical analysis (two-tailed
-test with Bonferroni correction) of normalized fold change (treatment/control) assessed significant differences in treatment efficacy. Cohen's
quantified pathophysiological treatment category effect size compared to "all" (e.g., all pathophysiological treatment categories combined).
Inflammation treatments were best at delaying onset (
= 0.42,
> 0.05). Oxidative stress treatments were significantly better for prolonging survival duration (
= 0.18,
< 0.05). Excitability treatments were significantly better for prolonging overall health status (
= 0.22,
< 0.05). However, the absolute best pathophysiological treatment category for prolonging health status varied with disease progression: oxidative stress was best for pre-onset health (
= 0.18,
> 0.05); excitability was best for prolonging function near onset (
= 0.34,
< 0.05); inflammation was best for prolonging post-onset function (
= 0.24,
> 0.05); and apoptosis was best for prolonging end-stage function (
= 0.49,
> 0.05). Finally, combination treatments simultaneously targeting multiple pathophysiological categories (e.g., polytherapy) performed significantly (
< 0.05) better than monotherapies at end-stage.
In summary, the most effective pathophysiological treatments change as function of assessment modality and disease progression. Shifting pathophysiological treatment category efficacy with disease progression supports the homeostatic instability theory of ALS disease progression.
A major bottleneck preventing the extension of deep learning systems to new domains is the prohibitive cost of acquiring sufficient training labels. Alternatives such as weak supervision, active ...learning, and fine-tuning of pretrained models reduce this burden but require substantial human input to select a highly informative subset of instances or to curate labeling functions. REGAL (Rule-Enhanced Generative Active Learning) is an improved framework for weakly supervised text classification that performs active learning over labeling functions rather than individual instances. REGAL interactively creates high-quality labeling patterns from raw text, enabling a single annotator to accurately label an entire dataset after initialization with three keywords for each class. Experiments demonstrate that REGAL extracts up to 3 times as many high-accuracy labeling functions from text as current state-of-the-art methods for interactive weak supervision, enabling REGAL to dramatically reduce the annotation burden of writing labeling functions for weak supervision. Statistical analysis reveals REGAL performs equal or significantly better than interactive weak supervision for five of six commonly used natural language processing (NLP) baseline datasets.
Multiple studies have shown that antecedent diseases are less prevalent in amyotrophic lateral sclerosis (ALS) patients than the general age-matched population, which suggests possible ...neuroprotection. Antecedent disease could be protective against ALS or, conversely, the asymptomatic early physiological underpinnings of ALS could be protective against other antecedent disease. Elucidating the impact of antecedent disease on ALS is critical for assessing diagnostic risk factors, prognostic outcomes, and intervention timing. The objective of this study was to examine the relationship between antecedent conditions and ALS onset age and disease duration (i.e. survival). Medical history surveys for 1439 Emory ALS Clinic patients (Atlanta, GA, USA) were assessed for antecedent hypertension, hyperlipidemia, diabetes, obesity, asthma, arthritis, chronic obstructive pulmonary disease (COPD), thyroid, kidney, liver, and other non-ALS neurological diseases. The ALS onset age and disease duration are compared between the antecedent and non-antecedent populations using chi square, Kaplan-Meier, and ordinal logistic regression. When controlled for confounders, antecedent hypertension (high blood pressure), hyperlipidemia (high cholesterol), arthritis, COPD, thyroid disease, and non-ALS neurological disease are found to be statistically associated with a delayed ALS onset age, whereas antecedent obesity body mass index (BMI) > 30 was correlated to earlier ALS onset age. With the potential exceptions of liver disease and diabetes (the latter without other common comorbid conditions), antecedent disease is associated with overall shorter ALS disease duration. The unique potential relationship between antecedent liver disease and longer ALS disease duration warrants further investigation, especially given liver disease was found to be a factor of 4-7 times less prevalent in ALS. Notably, most conditions associated with delayed ALS onset are also associated with shorter disease duration. Pathological homeostatic instability exacerbated by hypervigilant regulation (over-zealous homeostatic regulation due to too high regulatory feedback gains) is a viable hypothesis for explaining the early-life protection against antecedent disease and the overall lower antecedent disease prevalence in ALS patients; the later ALS onset age in patients with antecedent disease; and the inverse relationship between ALS onset age and disease duration.
Background: Recent studies have suggested overlapping pathological features among motor neuron, cognitive and neurodegenerative diseases. Aims/Methods: Secondary analysis of 46 amyotrophic lateral ...sclerosis (ALS) patient autopsies was performed to independently assess pathological feature prevalence (e.g. percent of patients with any positive finding), degree of severity (e.g. mild, moderate, severe), and 2,200+ potential clinical/neuropathological correlations. The possible impact of gender, onset age, onset type (limb vs. bulbar), riluzole treatment, and severe TDP-43 pathology was assessed within patient subgroups. Results: Assessed features (prevalence, severity) include: lateral corticospinal tract degeneration (89%, moderate); Purkinje cell loss (85%, mild); localized neuronal loss (83%, mild to moderate); TDP-43 inclusions (80%, moderate); Betz cell loss (76%, mild); neurofibrillary tangles (78%, severe); anterior corticospinal tract degeneration (72%, moderate); spinal ventral root atrophy (65%, moderate); atherosclerosis (35%, mild); β-amyloid (35%, mild); tauopathy/tau inclusions (17%, mild); ventricular dilation (13%, mild); Lewy body formation (11%, mild); microinfarcts (7%, mild); and α-synuclein (4%, mild). Twenty-two percent of patients met criteria for Alzheimer's disease (AD) and 26% for frontotemporal lobar degeneration. Substantial differences were identified in the AD group and in the different onset age groups. Conclusion: Our findings support the hypothesis that ALS and its variants could comprise a larger neuropathological continuum.
Cargos have been observed exhibiting a “stop-and-go” behavior (i.e. cargo “pause”), and it has generally been assumed that these multi-second pauses can be attributed to equally long pauses of ...cargo-bound motors during motor procession. We contend that a careful examination of the isolated microtubule experimental record does not support motor pauses. Rather, we believe that the data suggests that motor cargo complexes encounter an obstruction that prevents procession, eventually detach and reattach, with this obstructed-detach–reattach sequence being observed in axon as a “pause.” Based on this, along with our quantitative evidence-based contention that slow and fast axonal transport are actually single and multi-motor transport, we have developed a cargo level motor model capable of exhibiting the full range of slow to fast transport solely by changing the number of motors involved. This computational model derived using first-order kinetics is suitable for both kinesin and dynein and includes load-dependence as well as provision for motors encountering obstacles to procession. The model makes the following specific predictions: average distance from binding to obstruction is about 10μm; average motor maximum velocity is at least 6μm/s in axon; a minimum of 10 motors is required for the fastest fast transport while only one motor is required for slow transport; individual in-vivo cargo-attached motors may spend as little as 5% of their time processing along a microtubule with the remainder being spent either obstructed or unbound to a microtubule; and at least in the case of neurofilament transport, kinesin and dynein are largely not being in a “tug-of-war” competition.
•We develop a model of kinesin and dynein motors suitable for use at the cargo level.•We propose that motor “pausing” is obstruction/ensnarement followed by detachment.•We find that motor count can potentially explain fast versus slow transport.•We make several experimentally testable predictions based on this result.