Predicting a Dichotomous Health Outcome to a Certainty Using Many Attributes

Abstract

The issue of risk/prognostic prediction has long troubled clinicians and medical researchers alike. Unless the cause-and-effect relation is deterministic, a correct prediction for a subject has never been easy. A novel prediction method, the "prediction using many attributes" (PUMA), will be proposed. Under the homogeneity assumption that every subject of the same outcome has the same distribution of attributes, PUMA will always (with probability one) lead to a correct prediction, if the following three conditions are met: (1) the number of attributes that are collected tends to infinity; (2) the proportion of the collected attributes that have some (albeit only very trivial) predictive power for the outcome does not tend to zero; and (3) among the collected attributes, an attribute can correlate with no more than a finite number of other attributes. In this bio-informatics age, it is of interest to test-run a PUMA, using the now readily available genomic/proteomic data.