Update section 7 draft#6
Conversation
hjung-stat
changed the title
|
Nice job! This section is well structured and clearly tied to the main themes of the paper. One suggestion would be to make each subsection slightly more concrete by including at least one specific biomedical example. For instance, in the “Robustness Under Distribution Shift and System-Level Constraints” subsection, it may be useful to elaborate on a concrete form of distribution shift, such as hospital-to-hospital or scanner-to-scanner variation, perhaps drawing a bit more detail from the cited reference. |
|
|
||
| Key unresolved questions and research opportunities that will shape the development and deployment of foundation models in biomedicine. | ||
| Despite substantial progress in biomedical foundation models, several fundamental challenges remain unresolved. Building on the central themes of this review—namely the tension between domain-specific and general models, the importance of system-level integration, and the need for reliable evaluation—we highlight three core open problems that are likely to shape the future of the field. These challenges are deeply interconnected and reflect limitations not only in model design, but also in data, training objectives, and deployment environments. | ||
|
|
There was a problem hiding this comment.
The open problems listed here are great.
Under domain specific, is there value in mentioning explicitly the difference between learning ICD codes vs diagnosis.
Another thing is do we want a single multimodal model vs modality specific models that are aggregated by a head model that solves a downstream task. Maybe the multi-model approach is more "intepretable".
Interpretability of models can be absolutely essential in biomedical settings. Should this be added.
There was a problem hiding this comment.
Thank you for the helpful suggestions!
I revised the section by:
- clarifying the difference between ICD code prediction and clinical diagnosis,
- adding a brief discussion of unified multimodal models versus modular, modality-specific approaches, and
- highlighting the role of interpretability in real-world biomedical deployment.
[ci skip] This build is based on 9fdf1d6. This commit was created by the following CI build and job: https://github.com/AdaptInfer/fm-survey/commit/9fdf1d68928dae77ce92ec6a69d7b9d8616d9a2d/checks https://github.com/AdaptInfer/fm-survey/actions/runs/25328637370
[ci skip] This build is based on 9fdf1d6. This commit was created by the following CI build and job: https://github.com/AdaptInfer/fm-survey/commit/9fdf1d68928dae77ce92ec6a69d7b9d8616d9a2d/checks https://github.com/AdaptInfer/fm-survey/actions/runs/25328637370
4 participants
No description provided.