This study by Ali et al. ^{Reference Ali, Ayub and Naveed1} makes a valuable contribution to intracerebral hemorrhage (ICH) prognostication by validating machine learning (ML) models across two genuinely distinct cohorts. We would like to reflect on the following points.

First, the cross-cohort transportability design is a methodologically rigorous test of external validation and generalizability. The authors successfully bridged two highly divergent datasets: the unselected Qatari registry, which included large hemorrhages and normotensive presentations, and the restricted ATACH-2 clinical trial cohort. Demonstrating robust discrimination (area under curve 0.86–0.92) across such differing baseline severity distributions supports the authors’ premise that multi-cohort training captures highly transportable clinical patterns.

Second, incorporating inpatient variables (e.g., intubation status, hematoma expansion) meaningfully improved model performance. While clinically intuitive, this raises an important tension: models weighed heavily toward in-hospital events risk reproducing a self-fulfilling prophecy. These variables are often downstream of clinical decisions, such as overall treatment intensity or the withdrawal of care, rather than independent of them. Standardized timeline metrics, as proposed by the Code-ICH framework, ^{Reference Li, Yakhkind and Alexandrov2} will be essential to untangle true prognostic value from decision-consequence.

Finally, a critical consideration for clinical deployment is the rapidly evolving surgical landscape. With recent trials demonstrating the benefit of minimally invasive hematoma evacuation, ^{Reference Pradilla, Ratcliff and Hall3} ICH treatment paradigms are fundamentally shifting. Because these ML models were developed using historical cohorts managed prior to the widespread adoption of these techniques, their prognostic estimates may not seamlessly generalize to patients undergoing minimally invasive surgery today. Future model iterations will likely require retraining or recalibration as these advanced procedures become integrated into standard clinical practice.

Overall, this work significantly advances our understanding of how ML can leverage complex clinical data across diverse populations, and the transparent TRIPOD-AI reporting sets a commendable standard for future prognostic tools.