Peer-reviewed article
Causal estimation and machine learning methods for survival outcomes among AYA cancer patients: A scoping review
Nana Owusu Mensah Essel, Rui Liu, Liz Dennett, Paul Craig Nathan, Sumit Gupta et al. ·
Summary
Written by AI, not yet AICS's view · claude-sonnet-5-5This scoping review maps how causal inference and machine learning methods have been used to estimate or predict survival in adolescent and young adult (AYA) cancer patients. It included 104 studies, all retrospective cohorts, and most used US registry data, especially SEER. Propensity score matching was by far the most common causal method, and few studies specified a formal causal design or tested for unmeasured confounding. Machine learning prediction models were used in 19 studies, and only four were externally validated. The review is about research methods for survival data, not about children's well-being or an intervention for children.
Key points
- 104 studies met the inclusion criteria (99 peer-reviewed articles and 5 conference abstracts), and 68 (65.4%) were published from 2020 onward.
- All studies were retrospective cohorts, and 75 (72.1%) used US registry data; SEER alone accounted for 59 (56.7%).
- Of 87 studies applying causal inference methods, 75 (86.2%) used propensity score matching; only 3 (3.4%) were designed causal analyses, and 1 (1.1%) reported a sensitivity analysis for unmeasured confounding.
- Competing risks were mishandled in 22/99 (22.2%) peer-reviewed studies and model overfitting occurred in 15/99 (15.2%).
- ML prediction was used in 19 studies, most often random survival forests (12; 63.2%), and only 4 (21.1%) were externally validated.
Limitations
The review describes the methods used in published studies and does not assess the effect of any intervention on children. The document does not state the age range of the AYA population, which may include many adults, and the evidence is dominated by US registry data.
Abstract
Survivors of adolescent and young adult (AYA) cancer experience an elevated risk of premature mortality. While causal inference and machine learning (ML) methods are increasingly applied to observational survival data, the methodological landscape of this work remains unmapped. This scoping review mapped the extent, range, and nature of evidence using these methods to estimate or predict survival outcomes in this population. Following Joanna Briggs Institute (JBI) scoping review methodology, eight databases were searched from inception to November 2025. From 11,584 identified records, 5629 were screened after duplicate removal, of which 104 studies met the inclusion criteria (99 peer-reviewed journal articles and 5 conference abstracts). Of these, 68 (65.4%) were published from 2020 onward, 104 (100.0%) were retrospective cohorts, and most used US registry data (75 [72.1%]; SEER alone, 59 [56.7%]). Eighty-seven studies applied causal inference methods, among which propensity score matching was near-universal (75; 86.2%), followed by inverse probability of treatment weighting (9; 10.3%). None used targeted maximum likelihood estimation, marginal structural models for time-varying confounding, causal forests, or target trial emulation. Methods commonly associated with causal inference were widely used, but formally specified causal analyses were uncommon. Only three studies (3.4%) were designed causal analyses (regression discontinuity, causal mediation, and inverse probability weighted analysis [1 study each]), one (1.1%) provided a complete identification statement, and one (1.1%) reported a sensitivity analysis for unmeasured confounding. Competing risks were relevant where a quantity of specific event(s) was estimated and death from other causes could prevent the observation of specific events. Among the 48 applicable causal studies, 9 (18.8%) used a competing risks or relative survival method. ML prediction methods were used in nineteen studies, most often random survival forests (12; 63.2%); only four (21.1%) were externally validated. Among 99 peer-reviewed studies, the most common methodological issues were competing risks mishandling (22/99; 22.2% overall) and model overfitting (15/99; 15.2%). Research applying causal inference and ML methods to survival outcomes in AYA cancer patients is growing rapidly. However, the evidence draws on a narrow range of methods and data, dominated by propensity score matching using US registry data, with minimal documentation of identifying assumptions, rare probing of unmeasured confounding, and competing risks rarely addressed where the outcome made them relevant. ML models showed adequate discrimination, but a large proportion were not externally validated or calibrated, raising concerns about external validity and generalizability. Closing the gap requires adopting target trial-emulated causal designs with estimand-appropriate competing risk handling, together with externally validated, calibrated mortality prediction tools.