Data mining limitations

Since cluster analysis methods are of limited nature, and MVA methods are not really data-driven, we must conclude that these methods as well as the metric MDS methods are not qualified as maximally data-driven data-mining aids. Therefore, our only hope is to promote the methods broadly called nonmetric MDS (nMDS) as a versatile general reduction methodology for relational data. 

Phase IV studies can also take the form of retrospective pooled analyses which are designed to reflect the totality of clinical research experience with a new drug which is usually obtained via analysis of pooled clinical trial databases. Such data-mining efforts should not be considered inferior to data obtained from the conduct of an individual clinical trial. In fact, there are substantial benefits of such an approach, as the results of a given trial, especially if the end point is not prespecified to be primary because of power limitations, can be a function of chance. To assure that the results obtained from pooled analyses are not biased, a prespecified data analysis plan is often formulated as the first step, outlining the clear goals of the proposed analysis as well as its methodology. Key to this type of analysis is the definition of the outcome measure. Both efficacy and safety measures can be the focus of these types of analyses. 

In this case example, the FDA s SRS + AERS database, through the end of the second quarter of 2005, was data mined to determine the lower 95% confidence interval limit of the EBGM scores (denoted as EB05), a measure of disproportional-ity, for rhabdomyolysis associated with the use of statins. The drugs of interest were atorvastatin, cerivastatin, fluvastatin, lovastatin, pravastatin, rosuvastatin and simvastatin. The event of interest was rhabdomyolysis. 

The biggest limitation with data mining is the quality of the data. Simply put, the results of the analyses are only as good as the data from which they are derived. The best databases are those that are relevant, complete, have rich-quality data, are large and get updated frequently. Unfortunately, many databases are designed for purposes entirely different than what they are being used for, when they are data mined. 

The interpretations of data mining results need the expertise of safety reviews and medical officers to analyze and interpret data appropriately. Data-mining signals by themselves are not indicators of problems, but are indicators of possible problems. Moreover, caution must be exercised with any comparison of disproportionality ratios across different products, for example comparison with competitor drugs because of the various limitations that exist when making these kinds of comparisons. 

Next to ADME phenomena, recent data mining studies also focused on the development or improvement of models predicting physicochemical properties relevant to the field of ADME. Examples are Henry s law constant [92], polar surface area [93], and log P [94]. These models try to overcome limitations of already existing models, see for example SlogP [94] vs. Clogp [95], or aqueous solubility [96], The latter study used more than 2000 compounds selected from the AQUASOL [97] and PHYSOPROP [98] databases. Comparison with a multilinear regression showed clear preference for the neural network. 

---