Kaden McKeen

Using SOTA contrastive pretraining to generate a foundation model for ECG diagnosis using 1.7 million ECGs (800k being from UHN's internal dataset).

Technologies

• PyTorch
• BERT
• Contrastive pretraining

Using multimodal and longitudinal techniques for neurodegenerative disease diagnosis & prognosis; Using 3D brain perfusion scintigraphy & MRI imaging data from Sunnybrook.

Technologies

• PyTorch
• MONAI
• Hugging Face

Developed an unsupervised 3D U-Net model variant for deformable, pairwise image registration of whole-brain MRI data under Dr. Maged Goubran & Ahmadreza Attarpour.

Technologies

• PyTorch
• MONAI
• TensorFlow
• VoxelMorph
• Compute Canada
• Neurite
• ITK-SNAP

Led LearnAI, a free course teaching ML to undergrads, as the club Director of Education.
Acted as curriculum development lead, administrative coordinator, lecturer, TA, and project mentor. Partnered with AI Commons to offer the course to people in emerging countries.

Performed machine learning for health research at U of T's WangLab.

Performed multi-source domain adaptation & generalization for MRI, with test-time input adaptation (conditional image generation) using content-style disentanglement.

Developed a subject-adaptive framework for fMRI-based visual stimuli reconstruction using Versatile Diffusion, a multimodal latent diffusion model, fostering subject-invariant latent spaces using a subject classifier, Gradient Reversal Layer, and ArcFace loss.

Began the SMaRTBRAIN & ECG foundation model projects, carrying into the PhD.

Topics

• Domain adaptation
• Content-style disentanglement
• Loss-conditioned DDPM
• Visual stimuli reconstruction
• Representation learning
• Latent diffusion model

Core developer on CyclOps, an open-source evaluation framework for ML research and deployment in healthcare, as a member of Vector’s AI engineering team.

Researched two risk-prediction use cases utilizing CyclOps: Leading a Vector-WangLab research collaboration predicting cardiovascular risk and all-cause mortality on the GEMINI dataset; Contributing to internal Vector research focusing on all-cause mortality and length of stay in ER on the MIMIC-IV dataset.

Developed generalizable, high-level querying and processor APIs, consisting of general task-based pipelines to query, clean, featurize, normalize, aggregate, impute, and vectorize tabular and temporal clinical data.

Technologies

• PyTorch
• scikit-learn
• MLflow
• NumPy
• Pandas
• SQLAlchemy
• SLURM

Developed a statistical methods & analysis package, consisting of end-to-end workflows to perform climate timeseries forecasting, based on modern statistical research.

Workflows

• EDA, reporting, and cleaning of spatio-temporal timeseries data
• Forecasting methods using climate simulations & observed re-analysis data
• Statistical analysis of temperature & precipitation data

Technologies

• SciPy
• NumPy
• Pandas
• Matplotlib
• xarray
• GCP
• BigQuery
• Jupyter Notebook

Developed computer vision pipelines to help gym users improve their workout form:

• Use deep-learning pose estimation to predict an exercise being performed
• Detect and classify repetitive motion as repetitions of an exercise
• Track multiple users using facial recognition and efficient, real-time object tracking