T-CAIREM Trainee Rounds: Marco Istasy & Leen Alzebdeh
Join us for these presentations
Trainee Rounds Presentations (Session 1)
Marco Istasy
Marco V. Istasy is a fourth-year medical student and researcher at the Temerty Faculty of Medicine, University of Toronto and an incoming resident in the Division of Neurosurgery, University of Ottawa. Prior to that, he obtained his Honours Bachelor of Science with a specialization in Neuroscience and a Master of Applied Science with the Department of Biomedical Engineering, both at the University of Toronto. His research interests lie at the intersection of artificial intelligence and medical hardware innovation, with a primary focus on developing technological solutions to optimize clinical outcomes for neurosurgical patients. Most recently, his work was recognized with the Young Investigator Award from the Canadian Brain Tumour Consortium.
Abstract title
Machine learning identifies prognosticators of intracranial metastatic disease in patients with breast or lung cancer
Abstract
Background: Intracranial metastatic disease (IMD) is a severe complication of cancer associated with substantial morbidity and mortality. Patients with breast or lung cancer are at particularly elevated risk, yet over 80% present with multiple brain metastases at detection: a consequence of delayed identification in the absence of neurological symptoms. No validated tool currently exists for prospective IMD risk stratification, representing a critical gap in personalized cancer surveillance.
Methods: Using a population-based Ontario Cancer Registry cohort of 86,082 breast and 57,259 lung cancer patients diagnosed between 2010 and 2023, we developed interpretable machine-learning competing-risk models to estimate individual IMD risk. For each cancer type, cause-specific CatBoost Cox models for IMD and death without prior IMD were combined via the Aalen–Johansen estimator to produce absolute cumulative incidence estimates at one, three, and five years.
Results: In a held-out test set, models achieved high discrimination for IMD (Uno's C-index: breast 0.95, lung 0.88; AUROC(t) at 1/3/5 years: breast 0.96/0.97/0.96, lung 0.89/0.90/0.90) and favorable precision–recall performance (AUPRC(t) at 1/3/5 years: breast 0.17/0.53/0.63, lung 0.37/0.61/0.64). Discrimination improved by 15–17% over a full-covariate linear Cox model, and by 19–20% over an age-and-stage-only baseline. Model-derived risk groups showed monotonically ordered cumulative incidence with significant separation at all horizons (Gray's test, p<0.001). Interpretability analyses identified cancer stage as the dominant determinant in both cancers; triple-negative and HER2-positive subtypes contributed additional risk in breast cancer, while
histology and tumor size were prominent contributors in lung cancer.
Conclusions: Interpretable machine-learning competing-risk models demonstrate strong discriminative ability and clinical utility for IMD risk stratification in patients with breast or lung cancer, outperforming conventional approaches across multiple performance dimensions. These findings support integration of such models into personalized surveillance strategies to enable earlier detection of IMD.
Leen Alzebdeh
Leen Alzebdeh is a graduate student researcher at the University of Alberta. She is developing predictive ML models to predict blood glucose in hospitalized patients using electronic health records, with the goal of alerting staff to upcoming dysglycemia and enabling timely, personalized clinical decision support. She has hands-on experience with time-series prediction, handling missing and irregular clinical data, and “opening” black box AI for physicians using interpretability techniques such as SHAP. Building on this foundation, her collaborators are now extending their work to create an AI insulin recommender that utilizes the blood glucose predictor as a virtual patient simulator. This work aims to further automate insulin decision support, thereby reducing demand on healthcare staff, lowering hospital expenses, and improving patient outcomes. Beyond her work, she's interested in bed-to-bedside translation: interpretable, transparent, and trustworthy algorithms that aim to improve patient care. She is also interested in clinical decision support, time-series prediction, and human-centred AI.
Abstract title
Predicting Inpatient Blood Glucose for Cardiovascular Surgery Patients with T2D Using EHR Data
Abstract
Objective: To develop and validate a BG prediction model suitable for EHR-integrated decision support in hospitalized patients.
Methods: We analyzed 37,906 BG measurements from 1,279 cardiovascular surgery patients with type 2 diabetes. Our approach: (1) processes irregularly timed clinical data without imputation, (2) uses a unique cross-validation design ensuring patient homogeneity across folds, (3) generalizes without patient-specific customization, and (4) generates clinically relevant forecasts. Models were trained on 38 feature categories including BG history, demographics, and visit data.
Results: The best model, XGBoost, achieved a mean absolute error of 1.49 mmol/L (26.8 mg/dL), with 97.8% of predictions falling within Clarke Error (CE) Grid clinically acceptable zones A and B. It outperformed the naive baseline by 32.7% and achieved 99% per-encounter CE.
Conclusion: These results demonstrate strong predictive performance for inpatient decision support using routine EHR data. External validation across diverse cohorts and hospitals is the essential next step toward clinical deployment. In addition, the predictor can be integrated with an automated insulin recommender system to enable seamless decision support at the bedside.
Clinical Significance: By enabling accurate, real-time BG forecasting without requiring patient-specific calibration, this tool could help clinicians proactively adjust insulin therapy, reduce hypo/hyperglycemic events, and improve glycemic control in hospitalized high-risk populations—ultimately supporting safer, more personalized inpatient diabetes care and better patient outcomes.
