Lung cancer as a disease group is the leading cause of cancer incidence and mortality worldwide, according to World Health Organization data. Non-small cell lung cancer (NSCLC) makes up the majority of lung cancer incidence and treatments depend on the specific clinical factors and molecular biomarkers associated with the disease and its subtype.

This means, in short, that treatment is highly personalised.

Traditional clinical trials continue to define standards of care, yet they do not always capture the diversity and pace of real clinical practice. Real-world data (RWD) offers a complementary lens, capturing how patients are diagnosed, treated and managed outside controlled settings.

Artificial intelligence and machine learning (AI/ML) offer a powerful opportunity to extract insights from messy, complex RWD if applied with the right structure to produce credible, reproducible evidence.

Real-world data in NSCLC includes electronic health records, disease registries and genomic testing results. These sources hold rich information on biomarkers such as EGFR, ALK, KRAS and PD-L1, which increasingly guide treatment selection.

AI can integrate this data at scale to detect patient subgroup patterns, predict treatment response outcomes and support more personalised therapies.
For example, AI might predict which patients respond better to a certain drug based on their genetic markers and history.

However, regulators and health technology assessment frameworks from the FDA, NICE and ISPOR/ISPE expect pre-specified study designs, transparent reporting of methods and assumptions, and demonstrable reproducibility.
Without clear, rigorous frameworks, AI tools may produce results that are unreliable or difficult to interpret or reproduce.

A robust approach to applying AI in NSCLC real-world data rests on five core principles.

Pre-specification of analyses: define cohorts, variables and endpoints before analysis begins. This aligns with trial-like discipline and reduces analytical bias.

Data provenance and quality: document data sources, curation steps and validation checks. Confidence in outputs depends on clarity at input.

Bias mitigation: Apply methods to address confounding and missing data. This ensures that observed effects reflect true clinical relationships rather than artefacts.

Model development and validation: use established algorithms such as gradient boosting or regularised regression with rigorous internal validation. Prioritise interpretability to support clinical trust.

Transparent reporting: provide detailed documentation of workflows, diagnostics and performance metrics. This enables independent verification and reuse.

Together, these elements align with established frameworks from regulatory and methodological bodies, creating a common language for evidence generation.
 
Potential outputs include discrimination and calibration metrics, model diagnostics and visualisations showing how clinical and biomarker data could inform treatment stratification.

When applied within this structure, AI models can achieve strong performance. Metrics such as discrimination and calibration demonstrate their ability to both distinguish outcomes and reflect real-world probabilities.

Integrated visualisations and diagnostics help translate complex outputs into clinically meaningful insights, showing how biomarker profiles and patient characteristics interact to influence treatment outcomes.

AI-enabled real-world evidence reflects broader patient populations and adapts to evolving standards of care. AI/ML RWE allows pharmaceutical sponsors to complement and refine evidence strategies across the product life cycle and demonstrate value in diverse populations.

A rigorous methodology allows AI to turn messy real-world NSCLC data into regulator-ready evidence that helps doctors choose the right treatment for the right patient.