Viktor Tóth

We recorded stereo EEG signals (SEEG) from epilepsy patients in our lab at Feinstein while we stimulated their vagus nerve electrically. I analyzed the recorded brain signals to discover between-patient consistent brain responses to continuous and rapid on-off vagus nerve stimulation. I then employed a wavelet analysis and developed a novel local consistency analysis method that compares responses in temporal and spectral domains by matching the responses of electrodes near patients using a customized cross-correlation method.

I further performed non-negative matrix factorization on the evoked spectrograms to reveal shared features of brain activity between electrodes and subjects.

I designed an algorithm and data pipeline that is the first to segment cross-sectional images of the pig vagus nerve on a single-fiber level. The vagus nerve consists of around 50 fascicles, each enclosing thousands of fibers.

I combined a custom computer vision algorithm with a deep convolutional instance segmentation network to achieve an over 95% accuracy in fiber detection.

Tools Used: Gaussian mixture models, Voronoi tessellation, Detectron2 Mask-RCNN, geometric manipulations

Overnight vital sign monitoring can be critical to a patient's well-being. Still, it's not a medical procedure without costs as it disrupts sleep, and sleep disruption is associated with delirium, cognitive impairment, weakened immunity, hypertension, increased stress, and mortality.

I developed a machine learning model to predict whether a patient's health is likely to deteriorate overnight using a deep recurrent neural network (LSTM) trained on sequences of 26 million vital sign measurements, e.g., heart rate, temperature, blood pressure, and so on.

Patients who are highly unlikely to deteriorate then don't receive unnecessary overnight vital monitoring—enabling more than 25% of patient nights to be left undisturbed.

The model was published in Nature Digital Medicine and is being deployed in multiple hospitals.

I implemented a deep recurrent variational autoencoder to encode images into sound sequences. The deep learning network takes images, represents them as amplitude, frequency, and spatial modulations of sound, then reproduces the original image in its decoder part by drawing on a canvas.

The network training is engineered to produce sound modulations that are audible for humans, so a small visual difference in the image will result in a slight but discernable sound shift. Psychoacoustical and brain imaging studies supported all the assumptions embedded in the sound modulation. I blindfolded myself for five days without an interruption to test the algorithm and successfully learned to translate visual information into sound in two separate tasks.

I designed an algorithm to cluster more than 100,000 web pages to detect changes over time that may introduce cybersecurity vulnerabilities.

I created a custom DBSCAN algorithm in MapReduce that clusters web pages according to their locality-sensitive hash (SimHash) by building a search tree out of the simhashes. I then compared the DOM of web pages of sufficiently similar simhashes using a custom optimized tree edit distance algorithm. Next, I reduced the runtime of DBSCAN by 4x and scaled up to hundreds of thousands of web pages on computing clusters which sped up tree edit distance computation by 6x.

After the web pages are clustered and compared, their differences (e.g., an added input text field) are surfaced to employed penetration testers to assess whether a vulnerability is introduced.

Thoughtflow is designed to extract sentiment from sentences and group them according to similarity using deep learning NLP methods. You think of a search bar, type in your thought as a sentence, and get grouped up with others who recently typed in analogous sentences. For instance, get matched with others who just had issues with their pets, who just celebrated their birthdays, or just recently tried to access a website that went down.

I hosted the deep learning model on GCP and provided a web interface to search and chat handling requests in Flask.