The Role of Biostatistics in Clinical-Translational Studies

Translating novel biomarkers or AI models into clinical practice requires precise data analysis to establish analytical and clinical validity. Our biostatisticians understand the specific statistical methods needed to interpret complex translational data, such as handling low variant allele frequencies (VAF) in liquid biopsy, evaluating turnaround times (TAT) in POC testing, or assessing model overfitting and calibration in machine learning algorithms. Partnering with an independent statistical team provides objective validation, lending credibility to the clinical-translational study data.

Clinical-Translational Study Design for IVDs and AI Algorithms

Diagnostic devices and their underlying AI/ML models require tailored study designs to evaluate performance against a reference standard. Study protocols are developed in collaboration with your team, aligned with CLSI guidelines, the FDA AI/ML SaMD action plan, and MHRA expectations. The approach includes:

Retrospective Biobank and Prospective Cohorts

Utilising retrospective biobank specimens for early biomarker discovery and algorithm training, followed by prospective translational cohorts to validate clinical performance and lock the algorithm.

Reference Standard Selection

Establishing the appropriate composite reference standard (CRS) for comparison. In liquid biopsy validation, where tissue biopsies may suffer from tumour heterogeneity, latent class analysis is implemented to account for an imperfect gold standard, with extended models utilised when conditional independence assumptions are violated.

Bias Mitigation

Designing studies to avoid spectrum bias and verification bias. For AI models, selection bias in training and tuning datasets is minimised to support algorithmic generalisability across diverse demographic subgroups.

Data Partitioning for AI

Establishing predefined training, validation, and independent test (holdout) datasets. Strict separation of these datasets is maintained to prevent data leakage, a critical requirement for validating locked algorithms and supporting Predetermined Change Control Plans (PCCPs) that govern bounded, pre-specified post-market modifications.

Resource Efficiency

For diagnostic and AI validation studies, precision-based sample size calculations are often more relevant than power-based calculations. Sample sizes are calculated to achieve narrow 95% confidence intervals for sensitivity and specificity, preventing underpowered studies or unnecessary patient enrolment.

Demonstrating Clinical Utility in Translational Research

Moving beyond analytical and clinical validity requires demonstrating clinical utility. Decision Curve Analysis (DCA) is utilised to quantify the net benefit of the diagnostic or prognostic algorithm across different threshold probabilities. Additionally, Net Reclassification Improvement (NRI) and Integrated Discrimination Improvement (IDI) are calculated to demonstrate how the novel algorithm improves risk stratification compared to standard clinical models. This provides complementary reclassification metrics to support the real-world impact of the device for payers and providers.

Real-World Evidence and Post-Market Surveillance

Demonstrating long-term clinical utility requires data beyond the controlled clinical trial setting. Analysis of real-world evidence (RWE) shows how the device or algorithm performs across diverse patient populations and healthcare settings. By analysing observational data and post-market surveillance metrics, the impact on patient management is defined. This includes calculating Number Needed to Screen (NNS), evaluating shifts in PPV/NPV based on real-world disease prevalence, and monitoring for algorithmic performance drift over time.

Regulatory Compliance and Submissions

Meeting regulatory requirements for IVDs and AI SaMD requires precise statistical documentation. The statistical sections for regulatory submissions are prepared, including Clinical Evaluation Reports (CER) for EU MDR, Performance Evaluation Reports (PER) for EU IVDR, as well as 510(k) and De Novo applications for the FDA.

Statistical Analysis Plans (SAPs)

Detailed SAPs are drafted to lock the analysis methodology before unblinding the data.

Agency Interactions

Direct statistical support is provided during FDA Pre-Submission (Pre-Sub) meetings or Notified Body reviews, defending methodology choices and addressing reviewer queries.

Data Quality and Interpretation

Clear separation between the sponsor and independent analysts is maintained during data analysis to prevent confirmation bias, while collaborating closely on study design and SAP finalisation prior to unblinding. Data collection is structured using CDASH and SDTM standards, and ALCOA+ principles are implemented for data management. Handling indeterminate or invalid results is a critical component of diagnostic study analysis. Predefined statistical rules are applied for handling these missing or uninterpretable data points without introducing bias, providing clear data visualisation and clinical study reports (CSRs) that translate complex statistical findings into actionable clinical insights.