Background: Cancer prediction AI models hold the potential for precision oncology but reinforce inequity with lifecycle bias. This narrative review summarizes recent findings on the sources of bias and the effectiveness of mitigation to suggest equitable biostatistical frameworks for oncology analytics. Methods: This current paper reviewed 19 primary studies (2020-2025), identified through PubMed, Scopus, ResearchGate, and Google Scholar, that address AI analyses in oncology with fairness reporting. Thematic synthesis was applied across stages of the AI lifecycle, from conception to post-deployment. Results: Bias was found to be a multi-faceted, lifecycle issue created due to the formulation of the problem, generation of data, and proxy outcomes. The proportion of models utilizing demographically diverse cohorts was only 22% with reports of race present in just 5% of papers, which comprised 88% white participants, where stated. In 11% of the models, environmental/life-course variables emerged. AI performance decreased from training (AUROC=0.93) and evaluation (AUROC=0.76) with uniform demographic disparities. Breast cancer staging models performed best in White compared to non-White patients, with older adults (≥70 years) having the worst performance despite having the highest cancer burden. Using the HEAL framework, fairness-health burden alignment was absent across age (0.0%), minimal across race/ethnicity (80.5%), and low in intersectional analyses (17.0%). Mitigation methods including resampling, adversarial learning and post-processing showed mixed and inconsistent effectiveness, with no method significantly improving fairness across all datasets. Conclusion: AI in oncology perpetuates structural inequity through upstream design decisions and insufficient representativeness. Fair biostatistical models require burden-concordant fairness assessment, integration of social determinants, participatory formation of problems, and lifecycle regulation and constant sub-group tracking.