Democratising Clinical AI through Dataset Condensation for Classical Clinical Models
Authors: Anshul Thakur, Soheila Molaei, Pafue Christy Nganjimi, Joshua Fieggen, Andrew A. S. Soltan, Danielle Belgrave, Lei Clifton, David A. Clifton
Organizations: University of Oxford, Oxford, UK. · Oxford University Hospitals NHS Foundation Trust, Oxford, UK. · GlaxoSmithKline, London, UK · Nuffield Department of Primary Care Health Sciences, University of Oxford, Oxford, UK. · Oxford-Suzhou Centre for Advanced Research, Suzhou, China.
Dataset condensation (DC) learns a compact synthetic dataset that enables models to match the performance of full-data training, prioritising utility over distributional fidelity. While typically explored for computational efficiency, DC also holds promise for healthcare data democratisation, especially when paired with differential privacy, allowing synthetic data to serve as a safe alternative to real records. However, existing DC methods rely on differentiable neural networks, limiting their compatibility with widely used clinical models such as decision trees and Cox regression. We address this gap using a differentially private, zero-order optimisation framework that extends DC to non-differentiable models using only function evaluations. Empirical results across six datasets, including both classification and survival tasks, show that the proposed method produces condensed datasets that preserve model utility while providing effective differential privacy guarantees - enabling model-agnostic data sharing for clinical prediction tasks without exposing sensitive patient information.
Figures & tables
Figure 1: Overview of dataset condensation (DC) workflow and structure of synthetic data. (A) Schematic showing how DC integrates into clinical ML pipelines. (B) t-SNE projections (left) and nearest-neighbour distance distributions (right) for real and synthetic samples across PUH, OUH, and UK Biobank Proteomics datasets. Distributions compare distances from real to real and synthetic to real samples.
Datasets
Samples
Task
Portsmouth University Hospitals (PUH)
38,717
COVID-19 Prediction
Oxford University Hospitals (OUH)
161,955
University Hospitals Birmingham (UHB)
95,236
UK Biobank –- Proteomics
52,999
Myeloma Prediction
SEER (Breast Cancer Subset)
4,024
Cancer Survival Analysis
UK Biobank -– Diabetes
474,886
Diabetes Survival Analysis
Table 1: Summary of datasets used for evaluation, including sample sizes and associated prediction tasks. The SEER dataset refers to a curated breast cancer subset commonly used in survival modeling.
Dataset: Diabetes (UK Biobank)
Model
Metric
Full
IPC 50
IPC 100
IPC 500
IPC 1000
COX
C-index
0.79 ± 0.002
0.787 ± 0.004
0.79 ± 0.003
0.788 ± 0.002
0.789 ± 0.003
ε
—
1.91
1.937
2.442
2.54
XGB
C-index
0.802 ± 0.003
0.762 ± 0.004
0.775 ± 0.003
0.786 ± 0.004
0.788 ± 0.002
ε
—
0.491
0.560
0.736
1.146
Dataset: SEER
Table 3: Evaluation of survival model performance (C-index with 95% CI) and privacy budget ( ε ) across IPC levels over the UK Biobank diabetes and SEER datasets. δ is fixed at 10−5 .
Figure 2: Kaplan–Meier (KM) curves for models trained on real and condensed datasets. The first row shows KM curves from A) XGBoost and B) Cox models trained on the Diabetes (UK Biobank) dataset and their respective best-performing condensed datasets. The second row shows corresponding KM curves from C) XGBoost and D) Cox models trained on the SEER dataset.
Model
PUH → UHB
OUH → UHB
AUROC
Sensitivity
Specificity
PPV
NPV
AUROC
Sensitivity
Specificity
PPV
NPV
Full
0.830 ± 0.037
0.631 ± 0.074
0.866 ± 0.005
0.038 ± 0.005
0.996 ± 0.001
0.842 ± 0.040
0.611 ± 0.077
0.917 ± 0.004
0.058 ± 0.008
0.996 ± 0.001
IPC 50
0.855 ± 0.033
0.758 ± 0.064
0.847 ± 0.006
0.040 ± 0.004
0.998 ± 0.001
0.825 ± 0.037
0.711 ± 0.074
0.814 ± 0.006
0.031 ± 0.003
0.997 ± 0.001
IPC 100
0.860 ± 0.035
0.765 ± 0.070
0.840 ± 0.004
0.039 ± 0.003
0.998 ± 0.001
0.831 ± 0.035
0.691 ± 0.074
0.863 ± 0.005
0.041 ± 0.004
0.997 ± 0.001
IPC 500
0.847 ± 0.038
0.698 ± 0.074
0.836 ± 0.003
0.035 ± 0.004
0.997 ± 0.001
0.864 ± 0.033
0.772 ± 0.064
0.833 ± 0.006
0.037 ± 0.003
0.998 ± 0.001
IPC 1000
0.861 ± 0.032
0.805 ± 0.064
0.845 ± 0.005
0.042 ± 0.003
0.998 ± 0.001
0.844 ± 0.035
0.745 ± 0.070
0.862 ± 0.005
0.043 ± 0.004
0.998 ± 0.001
Table 4: Cross-site generalisation performance across four train–test pairs. Each row shows the performance of models trained on condensed datasets generated from one hospital site and evaluated on a held-out test set from another site. The first row reports results on the UHB test set using models trained on data condensed from PUH and OUH. The second row reports results on the OUH test set using models trained on data condensed from PUH and UHB.
Figure 3: Cross-model evaluation of XGBoost-derived condensed data. Performance of support vector machine (SVM), Random Forest, and Logistic Regression models trained on condensed data generated using XGBoost, across four datasets: (A) PUH, (B) OUH, (C) UHB, and (D) Proteomics.
Table 5: Top 15 covariates ranked by hazard ratio (HR) with 95% confidence intervals from Cox models trained on real and condensed variants of (a) the Diabetes and (b) SEER datasets.
Figure 4: SHAP-based feature attribution comparison between models trained on real and condensed datasets. Top row: feature attributions from XGBoost models trained for COVID-19 prediction on (a) PUH and (b) OUH datasets. Bottom row: feature attributions from XGBoost-ARF models trained for survival prediction on (c) the UK Biobank diabetes dataset and (d) the SEER dataset. Each panel shows results using the best-performing condensed dataset.
Figure 5: Membership and attribute inference attacks on condensed data. A) Performance of a distance-based white-box membership inference attack, evaluated using AUROC, membership advantage, and true positive rate at a false positive rate of 0.1. B) Performance of the attribute inference attack, reported as R2 scores for top attributes.