Wikipedia, “Curse of Dimensionality,” available at <https://en.wikipedia.org/wiki/Curse_of_dimensionality> (last visited July 1, 2019).+(last+visited+July+1,+2019).>Google Scholar

Newell, A. and Simon, H. A., “Computer Science as Empirical Inquiry: Symbols and Search,” Communications of the ACM 19, no. 3 (1976): 113-126; Gardner, A., “Search: An Overview,” AI Magazine 2, no. 1 (1981): 2-23.CrossRefGoogle Scholar

Simon, H. A., Langley, P., and Bradshaw, G. L., “Scientific Discovery as Problem Solving,” Synthese 47 (1981): 1-27; Shrager, J. and Langley, P., Eds., Computational Models of Scientific Discovery and Theory Formation (San Mateo, CA: Morgan Kaufmann, 1990).Google Scholar

“Treatments” are, in reality, usually complex treatment plans, including activities such as cycles of drug infusion, monitoring, additional tests, and so on. Here we will simply call these collectively “treatments.”Google Scholar

Pharmaceutical Research and Manufacturers of America (PhRMA), “Biopharmaceutical Industry-Sponsored Clinical Trials: Impact on State Economies,” 2015; Martinez, J. “Driving Drug Innovation and Market Access: Part 1-Clinical Trial Cost Breakdown,” 2016, available at <https://www.centerpointclinicalservices.com/blog-posts/driving-drive-drug-innovation-and-market-access-part-1-clinical-trial-cost-breakdown/> (last visited July 1, 2019).+(last+visited+July+1,+2019).>Google Scholar

Kramer, A. D. I., Guillory, J. E., and Hancock, J. T., “Experimental Evidence of Massive-Scale Emotional Contagion through Social Networks,” PNAS 111, no. 24 (2014): 8788-8790.CrossRefGoogle Scholar

Two senses of the term “big” are often conflated. The sense of “big data” where, for example, Google and Facebook have enormous datasets, might more accurately be called “tall, narrow data,” wherein there are many independent observations of a few features. Put in terms of dimensionality and sample size, from earlier in the paper, there are many independent samples (“n”) for relatively few dimensions of feature space. This “tall narrow data” might be more correctly called “large n, low (or moderate) dimensionality data.” Medical data, especially at the molecular level, might better be called “short wide data,” with relatively few independent observations over a very large number of features, or, in the above parlance: “small n, high dimensionality data.” The “big data” problems that have seen success are the tall narrow ones. The short wide ones remain out of reach of current technology.Google Scholar

Karp, P. D., Billington, R., Caspi, R., Fulcher, C. A., Latendresse, M., Kothari, A., Keseler, I. M., Krummenacker, M., Midford, P. E., Ong, Q., Ong, W. K., Paley, S. M., and Subhraveti, P., “The BioCyc Collection of Microbial Genomes and Metabolic Pathways,” Briefings in Bioinformatics (2017): bbx085, doi: http://dx.doi.org/10.1093/bib/bbx085.Google Scholar

Shrager, J., “The Fiction of Function,” Bioinformatics 19, no. 15 (2003): 1934–1936, doi: http://dx.doi.org/10.1093/bioinformatics/btg277.CrossRefGoogle Scholar

Sweetnam, C., Mocellin, S., Krauthammer, M., Knopf, N., Baertsch, R., and Shrager, J., “Prototyping a Precision Oncology 3.0 Rapid Learning Platform,” BMC Bioinformatics 19 (2018): 341, doi: http://dx.doi.org/10.1186/s12859-018-2374-0.Google Scholar

Sweetnam et al. note that TRs should be captured not only for the recommended treatments (tests, etc.), but also for those actions that are considered but rejected, because they are incorrect, undesirable, or infeasible. These “contra treatment rationales” (“contra-TRs”) can carry as much, or in some cases more, information than the rationale supporting the final recommendation; often the final recommendation is a safe or possible choice, whereas physicians might like to do something that may be more effective if practical barriers, such as cost or side effects, could be surmounted. Contra TRs may also represent new treatment hypotheses, possibly worthy of testing.Google Scholar

Even though reasoning based on these partial explanations would likely be imperfect, these can serve as hypothetical “creases” where the space could fold, as depicted in Figure 2. (Unless they are analytic [for example, based upon mathematical formulae], such hypothetical folds need to be treated with statistical care regarding Type I Error [False Discovery]). The present hypothesis is that such hints do more to reduce the dimensionality of the problem than they do to increase the false discovery rate.Google Scholar

Hey, S.P. and Kesselheim, A. S., “Countering Imprecision in Precision Medicine,” Science, July 29, 2016.CrossRefGoogle Scholar

Shrager, J., “Precision Medicine: Fantasy Meets Reality,” Letters, Science 353, no. 6305 (2016).Google Scholar

Id.Google Scholar

Shrager, J., “Theoretical Issues for Global Cumulative Treatment Analysis (GCTA),” arXiv: 1308.1066v1 [stat.AP] (2013); Shrager, J. and Tenenbaum, J. M., “Rapid Learning Precision Oncology,” Nature Reviews Clinical Oncology 11 (2014): 109-118.CrossRefGoogle Scholar