Call - Special issue: Learning from Multiple Data Sources for Decision Making in Health Care

The increasing availability of digital data, along with recent developments in Artificial Intelligence, especially in the Machine Learning and Deep Learning fields, led the scientific community to debate whether data alone is sufficient for decision making and scientific exploration. We focus the attention on the healthcare domain, where peculiar issues affect data: indeed, data are usually collected under heterogeneous conditions (i.e., different populations, regimes, and sampling methods), suffer missingness – very often not at random – and their use is strongly constrained by privacy issues. In such a complex setting, this special issue challenges computer scientists to contribute to the above debate by designing and developing innovative methodological approaches, for solving complex decision-making problems in health care, leveraging on observational data.

Topics of interest include, but are not limited to, the following with an emphasis on novel generalizable methods applied to the healthcare domain:

• Causal discovery from multiple data sets.
• Federated causal discovery.
• Causal discovery from heterogeneous data sets.
• Transportability of causal models and inference.
• Neuro-symbolic approaches to learn from heterogeneous data sources.
• Continual learning on streams from multiple data sources.
• Computational intelligent strategies to support causal inference.
• Edge computing for decision making in healthcare.
• Integrative AI methodologies.
• Distributed inference methods.
• Continual Learning.
• Knowledge Discovery and Integration.
• Combination of deductive approaches with ML models.
• Combination of ontologies and/or knowledge-bases with ML to support decision making.

Peer Review Process:

All submitted papers will undergo a rigorous peer-review process featuring at least two reviewers. All submissions should follow the guidelines for authors available at the Journal of Biomedical Informatics website (http://www.elsevier.com/locate/yjbin). JBI’s editorial policy outlined on that page will be strictly enforced by special issue reviewers.

Note that JBI emphasizes the publication of papers that introduce innovative and generalizable methods of interest to the informatics community. Specific applications can be described to motivate the methodology being introduced, but papers that focus solely on a specific application are not suitable. A few examples of papers focused on methods previously published in JBI include: Kyrimi, et al. [1], Huang, et al. [2], Kocbek et al. [3], Houston et al. [4], García Del Valle et al. [5], Graudenzi et al. [6] and Sims et al. [7].

In particular, the authors of [1] showed the relevance of causal models and expert knowledge to develop credible models, i.e., capable of achieving good predictive performances when transported from the study cohort to the target population. Furthermore, [2] tackles the relevant issue of partially overlapping variables when data are collected from multiple data sources. This problem is extremely relevant both in theoretical and practical terms for decision making in the healthcare sector.

The contribution provided in [3] stressed the importance of working in a multi-source context by demonstrating how the linking of different repositories can improve the overall understanding of patients' conditions. Similarly, in [4] the authors extended this concept by introducing a methodology to evaluate to audit the data quality of the sources exploited by healthcare information systems. Then, in [5] the multi-source concept is transferred within the multi-modal environment and the authors surveyed the importance of considering different modalities to obtain a better disease understanding.

The works in [6] and [7] focuses on the importance of data. In [6] a data integration framework is defined for characterizing the metabolic deregulations that distinguish cancer phenotypes, by projecting RNA-seq data onto metabolic networks without the need for metabolic measurements; in [7] a biomedical informatics method is introduced that uses multiple public health data sources to perform surveillance of methadone-related adverse drug events. Interestingly, even if patient data are not linked between different data sources, results show that the integration of multiple public data sources can capture more cases and provide more clinical details than individual data sources alone.

Key requirements for JBI ML papers in addition to presenting novel methods (not simply application of existing methods to a new healthcare domain) are as follows: 1) projects must have clinicians involved in research question/problem formulation, defining input data, and assessing the results. 2) An explanation (with clinicians) of how the proposed method would fit into the clinical workflow is expected. It must be translational to practice. 3) Data sets should preferably be collected from hospitals after the research question was formulated, thus avoiding the use of available data (MIMIC) to define a very wide research problem that could potentially be answered with available open datasets (as an example: detecting if someone has COVID from Chest X-Rays would not be acceptable, as the gold standard test is the laboratory test). 4) As for explainability, SHAP values and related diagrams would not be enough: the paper should clearly describe and explain how clinicians use the visualization to make decisions. For further details please refer to https://www.sciencedirect.com/journal/journal-of-biomedical-informatics/publish/guide-for-authors.

Further details available -

Submission process, Questions, and References 

My source: 

https://aixia.it/en/gruppi/hc/