There is certainly no shortage of interest in artificial intelligence in oncology. Hundreds of AI approaches are being developed and studied across the cancer continuum, from diagnosis to prognostic risk assessment to treatment selection, and more. Yet far fewer have transitioned into routine clinical practice.
Bringing AI from the lab into the clinic requires far more than building a technically sophisticated model. It takes rigorous evidence to show that the technology addresses a meaningful clinical need, performs reliably across diverse patient populations, and remains robust and reproducible under the realities of everyday clinical use. These challenges extend well beyond any one company or application. They represent a broader hurdle for healthcare as complex AI technologies move toward clinical adoption.
This is critically important in oncology, where AI has the potential to influence decisions with significant consequences for patients. The stakes are high, and so must be the bar for evidence to move these applications from technical promise to trusted clinical use.
What does better evidence look like?
Readiness for real-world use requires evidence that goes beyond whether an AI approach performs well in a technical setting. A clinically useful AI tool should meet the following criteria:
Addresses a meaningful clinical need. For AI to have a meaningful role in patient care, it must address a real clinical question and generate insights that have practical value for patients and the clinicians caring for them. The goal is not only to demonstrate that AI can perform a specific task or predict one particular outcome, but to show that it can meaningfully contribute to patient care and improve healthcare delivery.
Generalizes across diverse patient populations and clinical settings. A model that performs well in a narrowly defined evaluation cohort should not automatically be assumed to perform the same way across a broader patient population or clinical environment. Clinical validation must examine performance across independent cohorts and a range of patient characteristics and clinical circumstances, including different risk groups, ethnicities, geographies, disease presentations and stages, treatment approaches and clinical settings. This type of validation also helps identify potential sources of algorithmic bias and understand where a model’s performance may vary.
Produces reliable results under real-world conditions. An AI model may perform well in the controlled conditions in which it was developed, but real-world clinical practice introduces additional sources of variability. The technology needs to demonstrate that its results remain accurate and reproducible when those conditions change. This is where clinical and analytical validation play complementary roles.
Clinical validation assesses whether the model provides clinically meaningful information or accurately assesses the outcome it was designed to address. Analytical validation assesses whether the technology can reliably and accurately produce that result under the conditions in which it will actually be used.
This distinction is especially relevant in digital pathology, where factors such as the scanner used to digitize a slide, staining protocols, image resolution, and other technical variables can affect the images an AI model analyzes. A model can demonstrate strong clinical validation and still produce unreliable results when those underlying conditions change. Analytical validation therefore helps demonstrate that results remain reproducible across the range of equipment, processes, and other conditions that may come into clinical practice.
Withstands independent scientific and clinical scrutiny. Demonstrating that a technology works is not enough. Confidence also comes from having the evidence examined by those outside the organization that developed it.
Peer-reviewed research and independent validation studies allow third-party investigators to evaluate a technology’s performance and methodology. This can take different forms. Presentations at major oncology meetings such as the American Society of Clinical Oncology (ASCO) and the American Society for Radiation Oncology (ASTRO) provide additional opportunities for the clinical and scientific community to examine the evidence. Regulatory decisions, such as FDA clearances, and inclusion in evidence-based clinical guidelines, e.g., the National Comprehensive Cancer Network (NCCN), provide additional forms of external review and clinical recognition.
Taken together, these criteria define what a strong evidence base for clinical AI should demonstrate. But no single study or type of evidence can demonstrate all of these criteria, as different research approaches provide different insights. Building confidence in an AI-enabled technology therefore requires evidence to accumulate over time.
How do we build this evidence?
Different types of evidence answer different questions about an AI-enabled technology. Two valuable approaches are retrospective and prospective studies, each providing a different type of evidence. Neither should be viewed as a substitute for the other.
Retrospective studies use previously collected data to develop and evaluate AI models and can provide valuable evidence, especially when validation is conducted independently using data outside the model’s development set. Testing a model on new, unseen data helps assess whether its performance holds beyond the population on which it was developed. Independent retrospective validation can therefore provide meaningful evidence of how well a model generalizes to other patient populations. However, because these studies rely on data previously collected, they cannot fully answer how an AI-enabled technology will perform when introduced into clinical practice.
Prospective studies evaluate performance or clinical impact as patients are treated going forward. This provides insight into how the technology performs in the populations and workflows where it is intended to be used, and whether its information can meaningfully influence clinical decision-making.
Artera’s experience illustrates how different prospective study designs can contribute to this evidence base. PROSTATE-IQ is a prospective Phase 3 randomized clinical trial evaluating treatment strategies in patients with prostate cancer recurrence after prostatectomy, with patients risk-stratified using the ArteraAI Prostate Test. The ASTuTE trial is a prospective multicenter implementation study examining whether Artera results change shared decisions about short-term androgen deprivation therapy in men receiving prostate radiotherapy.
These studies do not answer the same question, which is exactly the point. Evidence is not a single checkbox. Different forms of evidence build on one another to show not only that a technology works, but how and where it can be used responsibly.
Prospective clinical trials, particularly those designed to evaluate long-term outcomes, however, can take years to read out. Waiting a decade for evidence of an outcome before acting would not be practical for many clinical decisions. Rather than relying on a single study or waiting years for the outcome, researchers can build evidence incrementally, using validated measures and established reference standards to demonstrate clinical performance and utility along the way.
Recent recommendations from the Digital Pathology Association are designed to support researchers, clinicians, payers, and regulators in understanding how to best deploy AI in an effective and safe manner. The framework distinguishes among clinical validation, analytical validation, and clinical utility, providing a structured approach to evaluating AI in clinical practice.
A roadmap for assessing clinical AI
The goal isn’t to slow the field down; it is to establish a roadmap for navigating new and evolving territory. As an early leader in bringing AI-enabled cancer tests into clinical use, Artera has seen firsthand the importance of building an evidence base that can withstand scientific, clinical, and regulatory scrutiny. These learned lessons can help inform an evidentiary framework for the broader field, so that other AI technologies can follow a clearer path toward responsible clinical adoption.
Healthcare brings together many stakeholders, each with a different role in moving technologies toward patient care. Laboratories, hospitals, clinicians, and payers may approach AI from different perspectives. Still, they ultimately share the same destination: ensuring that AI can be used with confidence to deliver better care and outcomes for patients
For AI in cancer care, the need for a shared roadmap will grow as the technology evolves and new applications move toward clinical use. The evidence generated to date is beginning to reveal the challenges, detours, and potholes along the way, helping pave a clearer path for what comes next. The opportunity is not to choose between innovation and rigor, but to build the evidence and infrastructure that allow both to move forward together.
