The integration of Artificial Intelligence (AI) from research environments into clinical oncology workflows represents a major development in cancer care, but real-world adoption is shaped by complex non-technical variables. To map these dynamics systematically, the Patient AI Treatment Hub (PATH) project conducted a landscape analysis of procedures, barriers, enablers, and gaps across six core dimensions: legal, ethical, technical, social, economic, and care delivery.
The PATH Screenathon
To map the real-world barriers and enablers of AI in clinical oncology, the PATH project utilised an innovative, human-AI collaborative “Screenathon“. Over two intensive days, experts from the PATH consortium used an active-learning AI tool (ASReview) to rapidly screen over 12,000 policy and grey literature documents. This crowdsourced effort efficiently distilled the vast database into the “PATH Knowledge Warehouse,” a curated collection of 351 highly relevant policies and strategies that grounds our evidence base.
Key insights across six dimensions
1. Legal: data access and liability
Rules on who may access health data and under what conditions vary between countries. Rules are interpreted differently, sometimes even within the same country. Hospital administrations and software developers absorb the cost of that inconsistency, usually as delay. The European Health Data Space (EHDSA) should improve matters, providing a unified statutory framework to harmonise cross-border data exchange and streamline standardised data permits. However, a gap persists: the absence of updated liability frameworks designed for machine-driven errors and continuous software updates, leaving clinicians and developers with unclear lines of accountability.
2. Ethical: bias and transparency
Ethical challenges in clinical AI centre on algorithmic biases that can harm minority or underserved populations. Models trained on data that under-represents certain groups tend to perform worse for those groups, so the effect falls on populations that are already underserved. The second is “black box” model opacity that obscures the clinical rationale behind automated recommendations. When a system cannot show why it reached a recommendation, clinicians have no way of judging whether the reasoning holds.
These issues are addressed by adopting “equity-by-design” principles from the ground up, such as actively including marginalised groups to balance training data and implementing governance models, such as the EU AI Act. Despite these emerging regulations, there remains a lack of standardised certification processes and independent oversight bodies to address ethical accountability when an automated clinical system fails.
3. Technical: infrastructure and standards
Hospital implementation is constrained by older IT infrastructure than the AI being proposed for it and many systems run on outdated clinical hardware. Data flow can be improved by adopting open APIs and common data communication standards, specifically HL7, FHIR, and DICOM for medical imaging. A technical gap remains the lack of post-deployment monitoring protocols to detect real-world model drift and performance degradation over time.
4. Social: trust and digital literacy
This dimension highlights a lack of trust and digital literacy gaps running through clinical staff and patients. Trust can be built by explicitly designing AI as a supportive “co-pilot” that enhances clinical judgment rather than replacing it. For patients with limited health or digital literacy or poor digital access, community patient navigators have proved a useful intermediary. Neither approach is yet standard practice, so the amount of trust built still depends heavily on the individual institution.
5. Economic: costs and reimbursement
Economic adoption is hindered because traditional reimbursement systems are structured around volume-based, fee-for-service models that do not account for AI-driven efficiency, leaving hospitals to absorb upfront integration and subscription licensing costs. This financial pressure can be managed by adopting flexible public procurement templates, targeted research funding and shared data infrastructure like the EHDS. However, decision-makers are held back by a lack of empirical economic evaluations, and evidence of long-term financial impact is scarcer still.
6. Care Delivery: fitting into clinical workflow
Training datasets that do not reflect the actual patient population reproduce existing disparities rather than correcting them. Interfaces that sit outside the clinical workflow add to the documentation burden instead of easing it, which could result in staff less willing to use them. Enablers include workflow-friendly AI applications, such as ambient dictation and scribing tools, which automatically draft clinical notes and return time to healthcare practitioners. A clinical gap is the “pilot-to-practice” bottleneck, where algorithms stall before scaling because they were engineered in isolation without frontline clinical input or real-world testing.