T-CAIREM Trainee Rounds: Elnaz Ziad & Ganlin Feng

Join us for these presentations

Trainee Rounds Presentations (Session 3)

Elnaz Ziad

Elnaz Ziad is a PhD candidate in Mechanical and Industrial Engineering at the University of Toronto and a researcher at Princess Margaret Cancer Centre. Her research aims to improve care for older adults with cancer. Specifically, she is developing a machine learning–based model to predict treatment-induced toxicities for this vulnerable population at the time of their first treatment. By enabling accurate and scalable automated prediction, her work seeks to address important limitations of current toxicity prediction tools. In addition to her research, she pursues her passion for teaching as a Teaching Assistant, supporting students throughout their learning journey. Her other interests include community building, event organization, mentorship, and creating supportive spaces for learning and connection.

Abstract title

Geriatric oncology toxicity risk estimation after treatment (GO-TREAT): Machine learning models to predict adverse events in older adults with cancer.

Abstract

Background: The Cancer and Aging Research Group (CARG) score can predict cancer treatment toxicities in older adults but requires manual administration, limiting its use. The objective of this study was to develop a scalable machine-learning (ML) model to predict toxicity endpoints in older adults with cancer and to compare its performance to the CARG score.

Methods: We performed a retrospective cohort study of patients aged ≥65 years treated at Princess Margaret Cancer Centre (Toronto, Canada). CARG scores were available for a subset assessed in the Older Adults with Cancer Clinic (OACC). Predictors included demographics, diagnoses, labs, treatments, and patient-reported symptoms. Outcomes were ≥3-point deterioration in any Edmonton Symptom Assessment System (ESAS) symptom (10-point scale), grade ≥3 hematologic toxicity, acute care use, and mortality. We used a temporal train–test split (training n=8,725; test n=20,405),with the split date defined as the date the OACC was established. We trained multiple classification models on the training set, using a 20% internal validation split for hyperparameter tuning and model selection, and evaluated performance of the best model on the full test set as well as on the OACC subset (n = 184). Performance was evaluated using the area under the receiver operating characteristic curve (AUROC) and average precision (AP).

Results: Median age was 71.3 years (IQR 67.8–75.6); 54% were female; 63% received palliative-intent treatment. In the full test set, GO-TREAT achieved AUROC/AP of 0.72/0.48 (90-day acute care use), 0.73/0.51 (90-day grade ≥3 hematologic toxicity), 0.81/0.64 (365-day mortality), and 0.60/0.70 (90-day ESAS deterioration). In the OACC subset, GO-TREAT outperformed CARG across outcomes (ΔAUROC 0.10–0.22), largest for ESAS deterioration (0.22). 

Conclusions: GO-TREAT provides scalable prediction of multiple adverse outcomes in older adults with cancer and outperforms CARG in the OACC subset.

Ganlin Feng

Ganlin Feng is a second-year PhD student in Computer Science at Western University under the supervision of Dr. Pingzhao Hu, and holds a BSc in Computer Science from the University of Waterloo. Her research lies at the intersection of computer vision and medical AI, with a focus on rare disease recognition from facial images, a setting characterized by extreme data scarcity and high clinical variability. Her work involves building reproducible benchmark datasets, as well as developing deep learning approaches and synthetic data generation for fine-grained classification tasks. More recently, she has been exploring agentic AI systems for medical reasoning, where multiple agents collaborate to analyze visual and clinical signals and generate structured diagnostic hypotheses. By combining vision-centric approaches with learning-based and agent-based methods, she aims to improve the reliability and interpretability of AI-assisted diagnosis in the field of medical imaging.

Abstract title

Adaptive Agentic AI to Enhance Clinical Diagnosis in Paediatric Rare Diseases

Abstract

Introduction: Rare diseases (RDs) affect millions of children worldwide and often remain undiagnosed for years due to clinical heterogeneity and limited specialist expertise. Current diagnostic workflows primarily rely on clinical evaluation and genetic testing, which can be costly, time-consuming, and difficult to access during early screening. Many RDs present craniofacial phenotypes observable from facial photographs, providing a non-invasive signal for computational analysis. Recent advances in deep learning have explored automated facial analysis for RD diagnosis; however, existing approaches often struggle with little training data and limited interpretability. To address these challenges, we propose a curated pediatric RD facial dataset and an adaptive agentic AI system to enhance clinical support.

Methods: We curated a pediatric RD facial dataset of 456 images covering 103 disease categories under ultra-low-sample settings, supplementing training with phenotype-aware synthetic images. Image analysis models first examine facial photographs and generate a short list of candidate diseases for each image input. Building on this initial screening step, a multi-stage agentic AI reasoning system then evaluates each candidate by assessing its consistency with observed facial characteristics and phenotypic features to produce interpretable explanations. A final decision module then compares the evidence across all candidates and outputs the most likely differential diagnosis.

Results: The image analysis model alone achieved 25.66% Top-1 accuracy and included the correct disease within the top five suggestions in 43.36% of cases. A baseline vision-language model provided only marginal improvement (best Top-1 accuracy 26.55%). In contrast, our reasoning system improved Top-1 accuracy to 34.51% (+8.85% over the baseline), recovering approximately 79% of cases where the correct disease appeared in the candidate list.

Conclusion: By combining a curated pediatric RD dataset with adaptive agentic AI, our system advances AI-assisted RD diagnosis toward more transparent and clinically meaningful decision support in ultra-low-resource settings.