Page 23 - AIH-2-4

Artificial Intelligence in Health AI editorial policy ethics

Table 1. Summary of critiques and editorial responses across two case studies
Dimension Case study 1 (Haghish, 2025) 11 Case study 2 (Ding et al., 2025) 12
Data imbalance Overreliance on SMOTE without evaluation of generalizability. Not applicable.
Model interpretability Lacked SHAP-based or interpretable mechanisms. Feature attribution unclear; 178 features used without ranking.
Generalizability Single-site data; no external validation or transfer learning. No discussion of generalizability beyond one speech corpus.
Temporal modeling Not applicable. Fixed 5-s windows insufficient for dynamic emotional
variance.
Silence removal Not applicable. Silences removed, obscuring emotional/psychological cues.
Model architecture Not discussed. Bi-LSTM used despite limitation; transformers not explored.
Multimodal design Not applicable. Speech-only; no integration of text, physiology, or behavioral
context.
Proposed improvement SHAP, transfer learning, balanced evaluation. Transformer models, QPSO, SHAP, multimodal fusion.
Potential consequence Misclassification in adolescent self-harm; clinical misapplication. Missed suicide risk signals; failure in real-world crisis detection.
Editorial justification “Outside scope.” “Overly technical.”
Abbreviations: Bi-LSTM: Bidirectional Long Short-Term Memory; QPSO: Quantum-behaved particle swarm optimization; SHAP: SHapley Additive
exPlanations; SMOTE: Synthetic minority over-sampling technique.

Figure 1. Editorial gatekeeping in artificial intelligence health research

implicit biases, and institutional priorities that may limit Journals, including digital-only platforms, must often
the diversity of published viewpoints. 32-38 Because of space balance the volume of valid commentaries they receive
limitations and journal formatting constraints, the scope against practical considerations, such as editorial resources
of arguments permissible within letters to the editor is and thematic coherence, making it unrealistic to publish all
32
frequently restricted; moreover, although such letters serve submissions regardless of their merit. Furthermore, the
as a platform for academic discourse, their acceptance sensitive nature of mental health data imposes significant
remains contingent upon editorial discretion and alignment privacy constraints that restrict the open sharing of patient-
with the journal’s thematic priorities. 32,37 The visibility level information. Ethical and legal obligations to protect
of alternative frameworks within scholarly publishing is participant confidentiality limit access to raw datasets,
influenced by citation networks, funding availability, and which complicates reproducibility and external validation
institutional affiliations, affecting the accessibility of critical efforts – challenges well documented in AI healthcare
perspectives outside dominant paradigms. 19,32-38 research. 39-43 These factors underscore the need for adaptive

Volume 2 Issue 4 (2025) 17 doi: 10.36922/AIH025210049