T-CAIREM Trainee Rounds: Austin A. Barr & Valentina Tamayo Velasquez

 Join us for these presentations

Trainee Rounds Presentations (Session 5)

Austin A. Barr

Austin A. Barr is a medical student at the University of Calgary and co-founder of the school’s AI society. His research focuses on generating high-fidelity synthetic clinical data. Prior to medical school, Austin served as a Program Officer at the Canadian Institute for Advanced Research (CIFAR), where he supported the Pan-Canadian AI Strategy. He has a background in clinical research, having worked at Sunnybrook Hospital’s Research Institute, and holds a BSc from McMaster University. During his time at McMaster, he contributed to the design, implementation, and teaching of an introductory AI course. Austin is also active online, sharing developments on the integration of novel technologies in healthcare: @alvie_barr

Abstract title

Diffusion-Generated Synthetic Neuroimaging Enables Transfer Learning for Anatomical Landmark Localization in Low-Data Settings

Abstract

Machine learning in neuroimaging is often limited by the difficulty of assembling and sharing large, high-quality datasets. To address this, we evaluated whether a denoising diffusion probabilistic model (DDPM) could generate realistic synthetic lateral cervical spine radiographs and whether these data could support transfer learning for anatomical landmark localization in low-data settings. A DDPM trained on 4,963 radiographs from the Cervical Spine X-ray Atlas generated synthetic images that were reviewed in blinded fashion by six neuroradiologists and two neurosurgeons. Experts could not reliably identify the real images, and realism ratings did not significantly differ between real and synthetic images. Nearest-neighbour analysis found no evidence of explicit memorization of training data. After quality control, we released a synthetic dataset including 20,063 radiographs. We used this synthetic corpus to train a landmark-localization backbone and evaluated transfer learning on the external CLX-34 dataset using a 22-keypoint task. Fine-tuning from the synthetic backbone substantially outperformed de novo training on CLX-34 data. These findings show that synthetic cervical spine radiographs can be both realistic and useful for downstream model training when labeled data are limited. We are now expanding this approach to MR imaging for brain metastases.

Valentina Tamayo Velasquez

Valentina Tamayo is a PhD candidate at the Institute of Medical Science, University of Toronto, under the supervision of Dr. Bernard Le Foll and Dr. Andrea Waddell. Her research focuses on development, implementation, and adoption strategies for integrating artificial intelligence (AI) and machine learning (ML) technologies in mental healthcare settings, with a particular emphasis on patient safety and outcomes. Her thesis project examines the development and implementation of an ML model, the Predictive Risk Identification for Mental Health Events (PRIME) tool, to predict mental health-specific adverse events, alongside co-design and stakeholder engagement for AI response strategies. She also holds an Honours Bachelor of Science degree from the University of Toronto (Neuroscience, Psychology, and French Literature and Language Studies).

Abstract title

Bringing AI to Psychiatry: Developing and implementing an AI prediction and alerting tool for psychiatric adverse events.

Abstract

Patient deterioration in mental health settings can lead to a specific set of adverse events such as self-harm, increased violence, restraint, or seclusion, that can contribute to the worsening of a patient’s physical and mental well-being. Early warning systems and machine learning (ML)-based alerting systems are emerging tools that can be used to detect, predict, and prevent these mental-health specific adverse events. Despite the increase of ML adoption in many healthcare fields, the implementation of ML-based predictive tools remains limited across mental health settings, representing a gap in patient safety. The aim of this study is to develop and pilot a ML-based prediction tool for mental health-specific adverse events.

This retrospective cohort study sampled data from patients admitted to a large inpatient mental health hospital. Routinely collected data from the electronic medical record was used to train and test three different ML models (2020-2023 data). The best-performing model was selected for model evaluation with 2024 data. We named this model the Predictive Risk Identification for Mental Health Events (PRIME) tool, which was piloted in three hospital units to evaluate feasibility and safety of implementation.

The PRIME tool included 50 data elements, encompassing a range of static and dynamic routinely collected data. These best-performing model was a long short-term memory (LSTM) model, a time-sensitive machine learning model (AUC ROC = 0.83). The PRIME tool outperformed an existing validated risk assessment tool. During the pilot, physician-perceived accuracy aligned with model precision and physicians expressed high satisfaction with the alert frequency and the workflow integration. The PRIME tool demonstrated strong performance and feasibility in psychiatric settings, highlighting the importance of combining early warning systems and ML in mental health settings.