Clinical prediction models using machine learning in oncology: challenges and recommendations
Gary S. Collins, Mae Chester‐Jones, Stephen Gerry, Jie Ma, João Matos, Jyoti Sehjal, Biruk Tsegaye, Paula Dhiman
Abstract
Clinical prediction models are widely developed in the field of oncology, providing individualised risk estimates to aid diagnosis and prognosis. Machine learning methods are increasingly being used to develop prediction models, yet many suffer from methodological flaws limiting clinical implementation. This review outlines key considerations for developing robust, equitable prediction models in cancer care. Critical steps include systematic review of existing models, protocol development, registration, end-user engagement, sample size calculations and ensuring data representativeness across target populations. Technical challenges encompass handling missing data, addressing fairness across demographic groups and managing complex data structures, including censored observations, competing risks or clustering effects. Comprehensive internal and external evaluation requires assessment of both statistical performance (discrimination and calibration) and clinical utility. Implementation barriers include limited stakeholder engagement, insufficient clinical utility evidence, a lack of consideration of workflow integration and the absence of post-deployment monitoring plans. Despite significant potential for personalising cancer care, most prediction models remain unimplemented due to these methodological and translational challenges. Addressing these considerations from study design through post implementation monitoring is essential for developing trustworthy tools that bridge the gap between model development and clinical practice in oncology.