BROWSE
Detail
Journal Article
Use of a Theory of Expected Information for Sparse Data and Adverse Events in Clinical Trials and Other Biomedical Studies
Cited 0 time in 
Share 
Share
- Authors
- B. Robson, OK Baek
- Issue Date
- 2024-10
- Citation
- Information Sciences, v.680, pp.1-19
- ISSN
- 0020-0255
- Publisher
- Elsevier BV
- Language
- English
- Type
- Journal Article
- Abstract
- Objective: We argue for a novel contemporary information-theoretic approach to supplement current statistical and AI methods for design and analysis of biomedical studies such as stage I-III clinical trials or cohort studies with limited sample sizes and the potential for adverse events, and as a basis for decision support systems that address these issues. Current methodologies in trials have faced criticism, and the growing volume of available digital medical records suggests a fresh perspective. The wealth of high-dimensional data offers new opportunities. To tackle them, we wish to draw attention to multiple uses of the zeta function. The several practical advantages of this unusual feature are not evident a priori, so they are extensively discussed. Method: To tackle the challenges posed by sparse data, rare event analysis, we propose the combination of work of Black Swan Event (BSE) Theorists with extension of the Theory of Expected Information (TEI), from the 1970s, which is still employed in certain widely used bioinformatics algorithms and data mining practices. Our extension incorporates expectations of information derived from finite data, integrating over degrees of belief about physical probabilities. It is extended to predictive methods for selecting patients for clinical trials, which are “Glass Box”. Results: This framework leads to a set of methods utilizing the incomplete zeta function ς(s,n) summed over n observations, where s can take various meaningful values. Our findings indicate that formulations and algorithms built around such zeta functions can replace many statistics and computations in biomedical studies. Whenever we see a count n of something, possibly including a knowledge or degree of belief represented by a virtual count, it may appear as ς(s,n) in an appropriate new formulae. The “Glass Box” methods based on this often outperformed “Black Box” Machine Learning. Conclusions: These methods align with both frequentist and Bayesian approaches but also accommodate sparse data and justify intuitive rules-of-thumb (e.g., α = 0.05 significance threshold and the “rule of three” in trials). Furthermore, they provide “Glass Box” “explainable” AI, enhancing transparency and interpretability. This project received support from a South Korean government grant, specifically the Electronics and Telecommunications Research Institute (ETRI) grant 23ZS1100, focused on advancing Core Technology Research for Self-Improving Integrated Artificial Intelligence Systems.
- KSP Keywords
- Adverse events, Bayesian approaches, Bioinformatics algorithms, Black box, Black swan event, Clinical trials, Cohort studies, Data mining(DM), Degree of belief, High-dimensional data, Medical records