Clinical Trials: CLASSIF1 Standardized Individual Patient Classification

Summary: The CLASSIF1 percentile classifier provides standardized individualized patient classifications from experimental data of clinical trials, like the SAIS1 (Glimepiride versus Sitagliptin) or the PROLOGUE (Sitagliptin versus Diet/Training) trials, both available in EXCEL format.

Percentile directed classifications could equally individualize mass spectroscopy cytometry (CyTOF) heat map data being presently evaluated by sophisticated correlative patient group analysis that is typically unsuitable for individualized conclusions. Given a similar structure as data from RNA expression arrays, it would be advantageous to have them available in an exchangable format like EXCEL or comma separated variables (CSV) for further analysis.

1. Presence: Data of clinical trials with large patient groups are typically statistically evaluated to reveal advantages of new therapies. This is fundamental for medical progress, but ultimately individual patients are treated and not patient groups. Personalised, individualized, outcome oriented or precision medicine crucially depend of individually discriminating molecular parameter patterns for precise diagnosis and individualized outcome prediction to optimally apply therapies..

2. Standardized CLASSIF1 percentile classification uses only measured patient parameter values like weight, size, erythrocyte counts etc after transformation according to their position above (+), below (.) or in between (0) upper and lower percentile threshold pairs (5/95, 10/90, 15/85, 20/80, 25/75, 30/70%) of the value distributions for example of patients prior to therapy as reference. As parameter values increasse or decrease during therapy, more (+) or (-) values emerge in the triple matrix value representation. The classification program iteratively maximizes the sum of sensitivity plus specificity (fig.1) for the discrimination of diseases states by stepwise temporary removal of individual parameter columns of the database from consideration by the classification process. It is calculated whether the removal has improved or deteriorated the classification result, followed by reinsertion of the parameter column and continuation of the classification process without the values of the next parameter column until all parameter columns have been processed in this way.
- Only parameter columns having improved sensitivity and specificity of the classification above the initial level are retained at the end as inherently standardized classification masks, since the masks directly represent measured parameter values. Missing parameter values do not have to be reconstituted. Patients are classified with the available information until a predetermined cut-off level of less than 50% or 40% availibility of the required classification mask parameters. Although not required for classification, means, standard deviations, statistical significance and parameter correlations were calculated to better characterizate analyzed data sets.

In case the relevant percentile thresholds and classification masks are publicly available for example from clinical trial studies or other quality controlled sources, individual patients can be manually classified everywhere that means without computer for example in resource poor areas or countries. The number of provided data columns is in principle unlimited with data matrices of so far more than 50.000 analyzed parameter columns. Data pattern classification was in these cases superior in discriminiation to correlation or statistical analysis. In case of relative parameter values like fluorescence in flow cytometry, classifiers can be standardized on appropriate controls.

4. Example of Manual Individual Patient Classification: SAIS1

5. Future: As datasets of clinical trials become available in EXCEL or CSV (comma separated variables) format, molecular parameter patterns can be established to reveal molecular characteristics of individual patients in different diagnostic and therapeutic situations. They are a prerequisite for the development of molecular classifiers concerning disease diagnosis or individualized outcome predictions in case of a single or several applicable therapies. This applies also to the simplification of single cell mass spectroscopy (CyTOF) data evaluation (heat maps) in their presently considerable complecixity.