Page 116 - AIH-1-3

Artificial Intelligence in Health ISM: A new multi-view space-learning model

In this approach, sets of views with enough common and versatility of ILSM, extending far beyond the scope of
observations are integrated with ISM separately. Using the multi-view data analysis.
model parameters, the transformation into the latent ISM
space can be expanded to all views over all observations 5. Conclusion
in the set, resulting in much larger transformed views The proof-of-concept analysis results provide strong
than the original intersection would allow. This expansion preliminary support for the proposed new method. As a
process enables the integration of the ISM-transformed next step, we will perform a comprehensive comparison of
data from the different view sets, again using the ISM. ISM with state-of-the-art alternative methods, including
Interestingly, a similar integrated latent space approach those considered in this article, and report the findings in
has already been proposed to study the influence of social a follow-up article.
networks on human behavior. After masking a large
47
number of views, the dataset of UCI Digits dataset was To further illustrate ISM’s key benefits and broad
analyzed using this approach. A more detailed description applicability, we will conclude by presenting some potential
of the expansion process (Workflow S1, Figure S1) applications currently under evaluation, with results to be
and preliminary results (Figure S2) can be found in the published in future articles.
Supplementary Materials. In longitudinal clinical studies, where participants are
Important issues such as the handling of highly dynamic followed up later, the ISM model can be trained at baseline
or rapidly updating datasets have not yet been investigated. and applied to subsequent data to calculate meta-scores.
This will be addressed in a future article. The interpretability of the associated components makes
ISM meta-scores more appealing to clinicians compared
It is worth noting that by replacing NMF with NTF in to the mixed-sign latent factors from other factorization
the initialization unit of the ISM workflow, ISM can be methods.
easily extended to multi-view data where the views are
themselves tensors of order three or higher, provided that Consider complex multidimensional multi-omics data
all dimensions except the attribute dimension are shared from one and the same set of cells (single-cell technology).
between the views. Interesting applications include the There is a growing amount of single-cell data corresponding
analysis of longitudinal multi-view data or the integration to different molecular layers of the same cell. Data
of multiple X-ray views. These topics will be the subject of integration is a challenge as each modality can provide a
dedicated articles. different clustering stemming from a specific biological
signal. Therefore, data integration and its projection into
Finally, the extension of ISM to the ILSM approach, as a space must: (i) preserve the consensus between two
described in the methods section (Section 2), is achieved clusterings and (ii) highlight the differences each modality
by a simple chained matrix multiplication – an example of may bring. ISM view loadings can address these two key
ISM inheriting the simplicity and compactness of the NTF requirements: components with similar contributions from
model, made possible by embedding views in a 3D array. each molecular layer highlight a consensus that can be
This has important advantages: inferred from clustering based on the ISM meta-scores of
i. Performance such components. In contrast, components with differing
• Independent view factorizations can be achieved contributions from each molecular layer highlight each
using parallel computing.
• The number of attributes in each transformed modality’s specificities, which can be inferred from clustering
based on the ISM meta-scores of such components.
view is reduced to its factorization rank, allowing
ISM to be performed on a much smaller dataset. The area of spatial mapping, including spatial
ii. Versatility imaging and spatial transcriptomics, is expanding at an
• ILSM can be applied to compute NMF on big unprecedented pace. An effective method for integrating
data in a federated or distributed way. To this different levels of information, such as gene or protein
end, smaller slices are constructed at random, expression and spatial organization of cell phenotypes, is
with each slice considered a particular view an unmet methodological need. We believe that ISM can
that is submitted to ISM. Preliminary results integrate these different levels of information, as shown
indicate significant performance improvements in the analysis of the UCI Digits data, to capture the
(Workflow S2 and example in the Supplementary constituents that allow spatial patterns to be distinguished
Materials). across all levels.
While ILSM does not claim to outperform all alternative The identification of new chemotypes with
approaches in every context, this illustrates the scalability biological activity similar to that of a known active

Volume 1 Issue 3 (2024) 110 doi: 10.36922/aih.3427