Week one:

Spectro-temporal Feature Generation Method for Atrial Fibrillation Detection in ECGs, presented by Ruhi Sharmin (Pavlos Vlachos, advisor)

Abstract: This work presents a framework for extracting a novel set of features from Lead II ECGs and using these features to detect atrial fibrillation. Our method pre-processes the signal using a denoising step, identification of the QRS complexes, and beat segmentation. Subsequently, we use a series of signal processing methods including proper orthogonal decomposition (POD), the continuous wavelet transform (CWT), discrete cosine transform (DCT) and the standard cross correlation to extract a total of 48 signal features. The extracted features are classified into four groups: correlation (C), morphology (M), Discrete cosine transform (D), and entropy (E). In general, these features represent novel and robust ways to characterize and assess key aspects of arrhythmias, such as beat-to-beat variability and the presence of fibrillatory waves. Subsequently, we use the feature set to train a machine learning based ECG classifier using XGBoost (eXtreme Gradient Boosting). For designing, training, and testing of our method, we used the 2017 PhysioNet Challenge dataset which annotates ECGs into categories of ‘Normal’, ‘Atrial Fibrillation’, ‘Other’, and ‘Noisy’. For this purpose, first, we took an exhaustive grid search approach for hyper-parameter optimization of the classifier algorithm. Second, a k-fold cross-validation was used for creating k-unique subsets from the original undivided dataset for enhanced generalization of our model. Third, we partitioned each subset into different 80-20 train-test splits for building the model. The drop in the weighted average of the F1 scores of our proposed model from the training to the testing phase was only 3% whereas it was ~10% for prior works. This proves that our model is free from overfit and more generalized towards ECG dataset. On the other hand, mean F1-scores on the test dataset were 84% and 94% for the Normal and AFib categories, respectively. Moreover, the F1-score of AFib at 94% showed significant improvement (~10%) over established methods (82%~85%), while also using significantly fewer total features comparably. This demonstrates that our novel feature XGBoost arrythmia classifier can differentiate AFib and normal sinus rhythm from other arrhythmias in Lead II ECGs effectively. Moreover, by using a smaller feature set, our method is adaptable for real-time arrhythmia detection and is effective even for short (<10 beat) ECGs.

Evaluation link for Ruhi Sharmin: https://purdue.ca1.qualtrics.com/jfe/form/SV_eIZzuVuHE8dCC2i

Profiling Neuron Cytokine Expression Stimulated from Metabolites Associated with Alzheimer’s Disease & Type 2 Diabetes by Brendan Ball (Doug Brubaker, advisor)

Abstract: Alzheimer’s disease (AD) is the most common form of dementia that leads to memory loss, behavioral changes, and decline of cognitive function. In the United States, more than six million people are living with AD. Of this population, 81% of people with AD also reported to have impaired glucose levels or type 2 diabetes (T2D), a metabolic disease that affects blood sugar regulation. While there is a connection between T2D and AD, the biological pathway and mechanism in which T2D exacerbates AD progression is not well-understood. A literature review revealed nine candidate metabolites that are upregulated in the presence of people with T2D, AD, or both. We compared metabolites associated with neuroprotective or neurodegenerative properties, which include lauric acid, asparagine, fructose, arachidonic acid, aminoadipic acid, sorbitol, retinol, tryptophan, and niacinamide. The nine metabolites associated with the potential development of AD or T2D were used to individually stimulate cultures of primary neurons collected from embryonic CD1 mice. After neuronal stimulation, the cell media was sampled and analyzed using a Luminex assay. Our findings indicate that cytokines involved in the JAK/STAT and chemokine receptor pathways may be involved in T2D or AD development. Specifically, we have found that cytokines including MCP-1, MIP-1alpha, and RANTES are consistently active across all metabolite stimulations, despite associations to different disease groups. These cytokines may serve an important role in AD and T2D progression. Improving our understanding of the relationship between AD and T2D will enable further studies that utilize biosystems for the development of potential treatments.

Evaluation link for Brendan Ball: https://purdue.ca1.qualtrics.com/jfe/form/SV_doH4Ktq6v86aBZs

*Join Zoom Meeting https://purdue-edu.zoom.us/j/98811885943?pwd=Nll1MlE3TTYyOHZJL3hIYlpGSXhUUT09

Meeting ID: 988 1188 5943; Passcode: biomedical