Keywords: Synthetic Data, Rare Events, Tabular
Need: Machine learning models trained on NHS data frequently encounter rare but clinically critical presentations such as unusual drug interactions, atypical disease presentations, or uncommon patient demographics. Standard synthetic data generation methods replicate the central data distribution, leaving rare events underrepresented or absent. This class imbalance undermines model performance precisely where it matters most.
TailSynth will investigate methods for controllably synthesising rare or tail-of-distribution clinical events, developing approaches that allow practitioners to specify, amplify, and evaluate rare case generation for safer and more robust model training and testing.
Current Knowledge/Examples & Possible Techniques/Approaches: “SMOTE: Synthetic Minority Over-sampling Technique” (and Borderline-SMOTE), and CTGAN could act as baselines. ADASYN (He et al., 2008) adapts sampling density to local learning difficulty, generating more samples where the classifier struggles most. TabDDPM (Kotelnikov et al., ICML 2023) applies denoising diffusion to tabular data supporting class conditioning.
Evaluation for rare event synthesis may include Precision-Recall AUC, tail coverage, and nearest-neighbour adversarial accuracy (from SDMetrics) quantifies privacy risk, which is particularly important here as rare individuals are inherently more re-identifiable.
Related Previous Internship Projects: P61 - SynthVAE; NHS Synthetic Data Studio
Enables Future Work: Assurance of AI models, Equity testing during development
Outcome/Learning Objectives:
Datasets: NHS-specific rare event contexts include: sepsis onset in ICU (a canonical MIMIC-III/IV prediction task — sepsis affects a small fraction of stays but is highly lethal and well-studied); rare adverse drug events (ADEs occurring in <1% of prescriptions); atypical disease presentations in patients with uncommon comorbidity combinations; and rare disease patients catalogued via ORPHA codes (Orphanet maintains ~6,000 rare disease definitions). MIMIC-IV’s diagnostic codes span this rarity gradient naturally, making it an ideal benchmark dataset.
Desired skill set: When applying please highlight any experience around synthetic data, generative models, class imbalance methods (e.g. SMOTE, conditional GANs), python coding experience (including any coding in the open), and any other data science experience you feel relevant.
Return to list of all available projects.