סמינר המחלקה להנדסה ביו-רפואית
סמינר מחלקתי למסטרנטים
Artificial intelligence identifies a modest influence of character and emotions on blood glucose in type 1 diabetes
Gon Shoham, MD*1,2, Roy Eldor, MD, PhD2,3, Irit Hochberg, MD4,5, Ran Gilad-Bachrach, PhD 1
People with T1DM and their healthcare providers perceive that mood and sleep quality influence blood glucose (BG). Artificial pancreas algorithms use only BG estimates to predict future BG and determine insulin requirements. We aimed to improve the understanding of the effect of emotional states and personality on temporal BG changes using machine learning.
Patients with T1DM using continuous glucose monitoring (CGM) completed questionnaires on demographic, socioeconomic and medical information and personality characteristics. For several days, using a mobile app, patients recorded their present valence, arousal state and sleep quality. The data was synced with patient CGM readings and evaluated using gradient boosting decision tree (GBDT) over 5-fold cross-validation. Area under the receiver operating characteristic curve (AUROC) and marginal contribution were used to determine model feature importance. BG was divided into three classes: normal (70-180), high (>180), low (<70). For normal BG values, models were trained to predict BG class in the next 15, 30, and 45 minutes.
A total of 64 participants (age 4815 years SEM, years of T1D 1813, 59% women) from two medical centers participated for 17.315.2 SEM days, each reporting valence and arousal 5653 times. Models using personality and emotional features to predict future BG outperformed models using only BG, albeit by a small margin. Following BG and time-of-day features, both neuroticism and arousal were found to be predictive. Neuroticism levels and sleep were significantly different between participants with HbA1c values under 7.6% and over 9.5%. (p=0.001, p=0.02, respectively).
1 Department of Biomedical Engineering Tel Aviv University, Ramat Aviv, Tel Aviv, 6997801, Israel
2 Sackler Faculty of Medicine, Tel Aviv University, Ramat Aviv, Tel Aviv, 6997801, Israel
3 Diabetes Unit, Institute of Endocrinology, Metabolism and Hypertension, Tel Aviv Sourasky Medical Center, Tel Aviv, Israel
4 Institute of Endocrinology, Diabetes and Metabolism, Rambam Health Care Campus, Haifa, Israel
5 Bruce Rappaport Faculty of Medicine, Technion – Israel Institute of Technology, Haifa, Israel
Quantifying Chorea, Daily-Living Physical Activity and Gait from a Wrist-Worn Accelerometer in Patients with Huntington’s Disease
Huntington's Disease (HD) is a progressive neurodegenerative, inherited disorder that leads to a combination of motor, cognitive, and behavioral symptoms. Over time, HD progresses to a state of severe hypokinesia and dementia. Chorea, one of the most common HD motor symptoms, is a hyperkinetic movement disorder that is characterized by excessive spontaneous movements that are irregularly timed, randomly distributed, and abrupt. Chorea severity is commonly assessed using a clinical test, i.e., subitems of the Unified-Huntington's Disease Rating Scale (UHDRS). However, this test is limited by inter-and interrater variability, subjective error, and its categorical design. Therefore, the first aim of this project was to objectively identify chorea-like movements using a wearable sensor. The second aim of this work is to quantify daily-living walking. Healthy gait requires a balance between various neuro-physiological systems and is considered an important indicator of a subject’s physical and cognitive health status. Previous work has generally relied on self-report to quantify walking in HD. Accurate measurement of daily-living walking of HD patients, which takes into account involuntary movements (e.g. chorea), for evaluating the quantity and quality of gait are needed. Our work addresses these needs by leveraging a wrist-worn accelerometer worn for 7 days. In this seminar, I will describe the signal processing methods that I developed to automatically identify irregular movements, daily-living walking, and gait quality from a wrist-worn device. Initial results indicate that irregular movements are much more common in patients with HD than in healthy controls, that daily-living walking quantity and quality are reduced in patients, and that the extracted measures are associated with clinical tests, supporting their validity.
RV Strain Classification from Volumetric CTPA scans using an Enhanced 3D DenseNet with Residual Attention Blocks
By Noa Cahan
Pulmonary embolism (PE) is a life-threatening condition, often without warning signs or symptoms. Early diagnosis and accurate risk stratification are critical for decreasing mortality rates. High-risk PE relies on the presence of right ventricular (RV) dysfunction resulting from acute pressure overload.PE severity classificationand specifically, high-risk PE diagnosis are crucial for the planning of appropriate therapy. Computed tomographypulmonary angiography (CTPA) is the golden standard in the diagnostic workup of suspected PE. Therefore, it can link between diagnosis and risk stratification strategies.
In this work, we address the problem of RV strain classification from 3D volumetric CTPA scans.
We retrieved data of consecutive patients who underwent CTPA and were diagnosed with PE and extracted a single binary label of “RV strain biomarker” from the CTPA scan report. This label was used as a weak label for classification. Our solution applies a 3D DenseNet network architecture, further improved by integrating residual attention blocks into the network’s layers. This model achieved an area under the receiver operating characteristic curve(AUC) of 0.88 for classifying RV strain, a sensitivity of 87% and specificity of 73.47%. Our solution outperformsstate-of-the-art 3D CNN networks. The proposed design allows for a fully automated network that can be trainedeasily in an end-to-end manner without requiring computationally intensive and time-consuming preprocessingor strenuous labeling of the data.
This current solution demonstrates that a small dataset of readily available unmarked CTPAs can be used for effective RV strain classification and later on for PE risk stratification.
To our knowledge, this is the firstwork that attempts to solve the problem of RV strain classification from CTPA scans and this is the first workwhere medical images are used in such an architecture. Our generalized self-attention blocks can be incorporated into various existing classification architectures making this a general methodology that can be applied to 3D medical datasets.
Under the instruction of Prof. Hayit Greenspan
Efficient Information Coding Over Living Organisms
Alon Akiva, B.Sc.
Department of Electrical Engineering, Faculty of Engineering,
Tel Aviv University, Israel
Encoding information over the genome of living organisms is a fundamental problem with various applications in synthetic biology such as biosensors, biological treatments and very long term storage. In addition, this topic is important for understanding fundamental aspects in molecular evolution and genetics.
This topic is specifically challenging due to the fact that living organisms constantly undergo evolution that can cause deletion and corruption to the encoded information. Unlike artificial gene synthesis, in which virtually any DNA sequence can be synthesized, DNA synthesis in living cells imposes different mechanisms of data corruption, and requires new sophisticated analysis methods. The aim of this research is to study the problem of information coding over living organisms from a digital communication system perspective.
We start by modelling the living organism as a discrete communication channel, in a way that captures the major biological phenomena, and derive the channel capacity for the model as a function of those major biological parameters. Then, we use computational molecular evolution approaches in order to estimate those important biological parameters. Next, we present a coding scheme which allows reliable communication over this channel, and analyze its performance. The design goals of the coding scheme is efficiency (in terms of storage density), flexibility (in terms of scalability and minimization of environment dependency), and decoder simplicity (in terms of minimal side information needed). To evaluate our approach we performed numerical simulations over real genomic data based on some micro organism models. The simulations include random channel behaviour based on parameters estimations from evolutional models and genomic data. In addition, we design an experimental framework for evaluating our models.