Need for Quality Auditing for Screening Computational Methods in Clinical Data Analysis, Including Revise PRISMA Protocols for Cross-Disciplinary Literature Reviews

Abstract

Deep learning (DL) is a leading paradigm in ML, which recently has brought huge improvements in benchmarks and provided principally new functionalities. The shift towards the deep extends the horizons in seemingly every field of clinical and bioinformatics analysis. Computational platform are exposed to a great volume of new methods promising improvements. Yet, there is a trade-off between the number of man/hours and the degree to which cutting edge advances in methodology are integrated into the routine procedure. Understanding why many of the new shiny methods published in the CS literature are not suitable to be applied in clinical research and making an explicit checklist would be of practical help. For example, when it comes to survival analysis for omics and clinicopathological variables, despite a rapidly growing number of architectures recently proposed, if one excludes image processing, the gain in efficiency and general benefits are somewhat unclear, recent reviews do not make a great emphasis on the deep paradigm either, and clinicians hardly ever use those. The consequences of these misunderstandings, which affects a number of published articles, results in the fact that the proposed methods are not attractive enough to enter applications. The example with the survival analysis motivates the need for computational platforms to work on the recommendations regarding (1) which methods should be considered as apt for a consideration to be integrated into the analysis practice for primary research articles, and (2) which literature reviews on cross-disciplinary topics are worth considering.