Each discipline combines scientific methods, measurable evaluation criteria and clinical expertise. We bring together the approaches most relevant to the objectives and challenges of each project.
Bias: types and methodologies
Biomedical datasets can contain different forms of bias that affect the validity, fairness and generalisability of AI models. We identify and quantify bias throughout the data lifecycle, from participant selection and data acquisition to measurement, missingness and model development. Our approach combines statistical assessment, targeted mitigation strategies and robustness evaluation.
Selection and representativeness bias — inverse probability weighting (IPW), stratified resampling and recruitment audits
Measurement and information bias — misclassification correction, sensitivity analysis and audit of labelling protocols
Demographic and equity bias — subgroup metrics (demographic parity, equalised odds), fairness audits and threshold review
Historical and shortcut bias — counterfactual tests, robustness to spurious correlates and provenance review
Causal inference: association, estimation and causal ML
Predictive association does not necessarily imply causation. We distinguish between identifying statistical relationships, estimating causal effects and investigating how these effects vary across individuals or populations. By combining causal inference frameworks with machine learning, we investigate complex biomedical relationships while explicitly addressing confounding and underlying assumptions.
Statistical association — correlation and predictive regression describe co-variation; they do not answer what happens under an intervention
Causal estimation — DAGs and the backdoor criterion, propensity scores, IPTW, doubly robust estimators, instrumental variables and difference-in-differences
Refutation — placebo tests, negative controls and sensitivity to unmeasured confounding (E-value)
Predictive AI: longitudinal and multimodal models
Biomedical conditions evolve over time and are rarely explained by a single source of information. We develop predictive models that capture temporal patterns and integrate heterogeneous data, including clinical records, laboratory measurements, medical imaging, biological measurements and physiological signals. Our focus is on predictive performance, temporal consistency and generalisation across datasets.
Longitudinal models — biomarker trajectories, mixed-effects models, temporal Transformers/LSTMs and survival models with time-dependent covariates
Multimodal fusion — imaging (CNN/ViT), tabular clinical data, genomics and omics, clinical text and wearable signals
Temporal and external validation — time-based splits and replication in independent cohorts
Explainability and white-box AI
Understanding how an AI model reaches its predictions is essential for scientific interpretation and clinical assessment. We combine explainable AI techniques with deeper investigations into internal model representations and computational mechanisms. We also explore interpretable reconstructions that make complex models easier to inspect, evaluate and audit.
Attribution — SHAP, Integrated Gradients and attention analysis at patient and cohort level
Imaging evidence — saliency maps that locate the regions supporting each finding
White-box reconstruction — interpretable surrogates, distillation, GAM/EBM and regularised trees
Mechanism audit — shortcut-learning detection and inspection of the features the model actually uses
Trustworthy AI: how reliability is measured
Predictive accuracy alone is not sufficient to establish the reliability of an AI system. We evaluate model performance through complementary measures of discrimination, calibration, robustness, fairness, uncertainty and generalisation. We also assess reproducibility and external validity to identify limitations and determine under which conditions model predictions remain dependable.
Discrimination — AUROC and AUPRC with confidence intervals, reported per subgroup
Robustness — external validation, stress tests and sensitivity to missingness and noise
Stability and equity — subgroup performance, data drift monitoring and post-deployment surveillance
Uncertainty — conformal prediction and intervals that tell the clinician when not to trust a case
Clinical knowledge integration
Biomedical AI should combine patterns learned from data with established clinical and biological knowledge. We integrate expert knowledge, clinical guidelines, physiological relationships, structured terminologies and causal information into model design, development and evaluation. This helps connect statistical predictions with clinically meaningful concepts and scientifically interpretable relationships.
Expert-driven design — variable definitions and DAGs built together with clinicians
Guideline constraints — regularisation and feature constraints aligned with clinical practice guidelines
Hybrid models — mechanistic and pharmacokinetic components combined with machine learning
Clinical review — case review with clinical committees and TRIPOD+AI style reporting
Collaboration
Let's build trustworthy AI together
Tell us the clinical or research problem and we will answer with a collaboration proposal.
