Comment on “Toward clinically actionable explainable AI in pulmonary arterial hypertension: Endpoints, calibration, and external validation”

Dear Editor, We read with great interest the recent contribution proposing a machine learning model (XGBoost), developed using registry data, to estimate mortality risk in adult patients with pulmonary arterial hypertension (PAH), and incorporating an SHAP-based interpretability strategy to clarify, both globally and at the individual level, the determinants of the prediction. This approach is particularly valuable because it supports a clinically credible use of artificial intelligence: not a “black box” producing a score, but rather a tool capable of substantiating its output and thus engaging with medical reasoning and clinical communication.1 Nevertheless, several methodological and translational aspects deserve consideration, as they may influence real-world implementation. First, the selected endpoint (“death by the next follow-up visit”), in the presence of nonuniform follow-up intervals, may introduce a component driven by healthcare organization in addition to intrinsic prognosis; in this regard, adopting predefined time horizons (e.g. 6 to 12 months) or a time-to-event framework could improve comparability and clinical usefulness.1 Supporting this view, prognostic analyses in populations with pulmonary hypertension associated with chronic respiratory diseases are commonly conducted using survival models, identifying independent predictors related both to hemodynamic severity and functional measures, thereby providing a natural methodological reference for temporally standardized endpoints.2 Second, while prioritizing high sensitivity is understandable in a setting in which the clinical cost of false negatives is substantial, a non-negligible proportion of false positives (with consequent reduced precision) makes it essential to define the intended use case and downstream “action pathway” explicitly: which interventions should follow an “high-risk” classification and what balance is expected between clinical benefit and healthcare burden. A similar message arises in PAH screening among high-risk populations, in which algorithms with high sensitivity and negative predictive value can reduce missed diagnoses but require structured diagnostic-decision pathways to manage lower specificity and the risk of overtesting appropriately.3 Finally, for practical applicability, beyond discrimination metrics it would be desirable to report calibration comprehensively and to discuss the decision threshold transparently, as well-calibrated probabilities and clinically justified thresholds are pivotal for turning predictions into decisions.1