TABA Seminar and Networking Event
Location: Astrazeneca, 1004 Middlegate Rd, Mississauga, ON
Time: 14:00 PM EST, light snacks and refreshments provided
Cost: This event is free of charge, kindly sponsored by AstraZeneca
Register here

Agenda:

Presenter: Jianyue Bai, University of Toronto
Title: Leveraging Neural Networks to Predict Abscess Formation in Acute Appendicitis Pediatric Patients
Abstract: We performed a retrospective chart review of patients undergoing appendectomy from 2018–2025 at a single institution. Twenty-eight literature-supported pre-and intra-operative risk factors were used as model inputs, with post-operative IAA serving as the binary outcome. We developed a five-step attentive neural network that mimics sequential clinical reasoning using gated linear unit blocks and dynamic attention. Five-fold nested cross-validation (a rotating 3:3:1 train/validation/test split) was used to prevent selection bias, and an ensemble of 50 runs ensured robust estimation. Final performance was evaluated using out-of-fold (OOF) testing, with 95% CIs derived via bootstrapping, and benchmarked against logistic regression, multilayer perceptron, and CatBoost. From 2,850 appendectomies, 1,462 reviewed, n=977 met inclusion criteria: pediatric patients with acute complicated appendicitis (median age 9 years, IQR 7-12), undergoing primary laparoscopic appendectomy. The observed IAA rate was 14% (IAA n=139, no IAA n=838). In OOF testing, our attentive model achieved the best discrimination (Ensemble ROC AUC = 0.77, 95% CI 0.73–0.81). With sensitivity of 0.91 (95% CI 0.85–0.95) and negative predictive value of 0.96 (95% CI 0.94–0.98), it functions as an effecti screening tool. Monotonically increasing observed IAA rates across predicted risk strata validate its utility for risk stratification. We conclude IAA risk can be predicted using pre-and intra-operative factors with good discrimination, enabling screening and risk stratification in pediatric complicated appendicitis.

Presenter: Luke Bai, Statistician, Roche
Title: Applying a Bayesian Predictive Probability of Success Framework with Mixture Priors in Phase III Clinical Trials
Abstract: In clinical trials targeting accelerated approval, interim analyses often evaluate early surrogate endpoints to forecast long-term confirmatory outcomes. Standard Predictive Probability of Success (PPoS) frameworks depend heavily on historical relationships between these endpoints, which can lead to overestimated success if the novel treatment exhibits prior-data conflict. Building on recent advancements in robust mixture priors, this presentation walks through an adapted Bayesian PPoS framework that incorporates partial confirmatory data alongside the interim surrogate endpoint. We implement a mixture prior that blends an informative historical surrogate prior with a vague, skeptical component. This allows the model to adapt, naturally placing less weight on historical assumptions if the interim surrogate results conflict with the early confirmatory data we’ve already collected. By using this methodology, we can clearly predict long-term clinical benefit while protecting against prior-data conflict, demonstrating the real-world utility of Bayesian designs.

Presenter: Mahmood Gohari, HSIF post-doc, Public Health Ontario
Title: High-Dimensional Genomic Prediction of Antimicrobial Resistance: An Interpretable Machine Learning Framework 
Abstract: Predicting antimicrobial resistance from whole-genome sequencing data presents a challenging high-dimensional statistical problem, where the number of genomic features far exceeds the number of samples and strong correlations are pervasive. We develop an interpretable machine learning framework to predict ciprofloxacin resistance in Shigella using genomic features. Beyond classification accuracy, we emphasize probabilistic prediction and model calibration, enabling reliable estimation of resistance risk. We evaluate performance using temporally independent data to assess generalization under potential distribution shift. Additionally, we incorporate SHAP-based methods to decompose predictions into feature-level contributions, linking statistical signals to known biological mechanisms.