Heorfy
Scientific capabilities

Clinical validation of diagnostic tests and health technologies

A technology that works is not necessarily a technology with demonstrated clinical performance.

A diagnostic test, prediction algorithm or clinical decision-support tool may perform impressively during development and lose accuracy when applied to a different population, setting or care pathway. Validation starts by specifying what the technology must demonstrate, in whom, against which reference standard and under what conditions of use.

Heorfy provides research methodology and biostatistical expertise for the design and analysis of diagnostic performance and clinical validation studies. The purpose is to generate interpretable evidence about performance, limitations and the settings in which a technology could provide genuine clinical value.

SERVICES

How Heorfy can contribute

Specific scientific capabilities, tailored to each study design and project objective.

01

Validation study design

Intended use, study populations, comparators and reference standards.

02

Diagnostic performance

Sensitivity, specificity, predictive values and ROC analysis.

03

Prediction models and algorithms

Discrimination, calibration, clinical usefulness and uncertainty.

04

Internal and external validation

Resampling, transportability and performance across settings.

05

Subgroup analysis and bias

Spectrum effects, selection, heterogeneity and measurement error.

06

SAPs and validation reports

Predefined analyses and transparent reporting of strengths and limitations.

Explore the approach

Clinical validation starts with intended use

Before choosing a performance metric, it is important to define the intended patients, the point in the care pathway and the decision the tool aims to inform. A screening test, a confirmatory investigation and a prognostic model cannot be assessed as though they served the same purpose.

Study design must address the target population, reference standard, recruitment, blinding where appropriate, prevalence and the possibility of spectrum or verification bias. If a decision threshold is needed, choosing that threshold should be clearly distinguished from subsequently validating it.

Beyond discrimination: what the results mean in practice

Sensitivity, specificity, predictive values and likelihood ratios describe different features of diagnostic performance. ROC analysis and discrimination measures can inform comparisons, but no single number captures clinical performance in full. Confidence intervals, the consequences of false-positive and false-negative findings, and the context of use all matter.

For risk prediction models, discrimination should be considered alongside calibration: whether predicted probabilities are reasonably consistent with observed event frequencies. Clinical utility, subgroup performance, stability and changes in case mix between settings may also need to be examined.

Internal and external validation, with uncertainty made explicit

Rigorous evaluation distinguishes development, model tuning and validation. Internal validation methods can estimate optimism but do not replace testing in genuinely independent data when transportability or generalisability is at issue.

Heorfy can support study design, sample size planning, the SAP and statistical analysis of validation studies. Reports may include estimates with measures of uncertainty, sensitivity analyses, identified limitations and evidence-based recommendations for further investigation.

Scientific evidence for technologies in development

This work is particularly relevant when developing diagnostics, medical software, AI-based algorithms or devices that need to demonstrate performance in real clinical settings. The focus is the scientific and statistical quality of evidence; certification and specialised regulatory interpretation should be coordinated with appropriate regulatory professionals when the project requires them.

Clinical validation can also inform later health economic and value assessment, once sufficient evidence is available to model performance and its potential consequences.

Working on a scientific project?

Tell me about it ↗