Exploratory data analysis & analytics on complex data:
Identifying risks to project data and to subject safety

This presentation will have 2 sections:

• Past experience
• A self-developed scenario

Past Experience

Reviewing oral corticosteroid (OCS) doses in an asthma phase III trial

• Critical to Quality (CtQ) study data
  • Asthma exacerbations
  • OCS dose titrations
• OCS dose titration rules

B. Cooper & R. Patterson, “The Corticosteroid Graph”, J. Allergy Clin. Immunol. 1976, Vol 8, Nr.6

A typical OCS Dose Graph

How was it done in real-life

• Stage 1: each CRA developed their own tool

• Stage 2
  • Tool developed only provided information on status of CRF completion
  • Issue: a significant amount of data was not SDV’ed (many patients were still under screening)
  • Lack of adequate visual dashboards –that would allow to quickly understand the OCS dose graph for a given pt., the current OCS dose status & whether the data being reviewed was confirmed data (SDV’ed data) or not.

A way to do this visually:

• Line plot for each individual patient
• With colour coding or similar to indicate SDV’ed or not
• Could also be Shiny app for CRA use –where CRA can mark if the data was SDV’ed or not
• Could also take the form of an interactive table (format could resemble the table in the previous slide)

• Reviewing asthma exacerbations in the same phase III trial
  • databases requiring review: CRF and patient ePRO input
  • review using CRF web platform & Excel outputs from the ePRO data
  • data analysis tool could combine data from both databases
• Analysing protocol deviations using Spotfire

• Communication of results
  • Rmarkdown presentations is an option
  • Include dataset used -guarantees results are reproducible

Self-developed Scenario

The question that we have asked ourselves:

• Can we identify patients that may potentially have diabetes, from an unknown dataset?

The dataset contains healthy and diabetic female subjects (source: https://data.world/data-society/pima-indians-diabetes-database)

Cases: 786

Variables: 10

The 2 “label” variables (that indicate whether each subject has diabetes or not) were removed from the dataset.

First quick look at the dataset structure:

• All variables are numeric
• And there are no missing values in the dataset (NAs)

##  Pregnancies      Glucose        BloodPressure   SkinThickness  
##  Mode :logical   Mode :logical   Mode :logical   Mode :logical  
##  FALSE:768       FALSE:768       FALSE:768       FALSE:768      
##   Insulin           BMI             Age         
##  Mode :logical   Mode :logical   Mode :logical  
##  FALSE:768       FALSE:768       FALSE:768

We then perform a first visual review of the dataset, and a descriptive statistics review:

First insight:

• INSULIN is unevenly spread with many high outliers.

##   Pregnancies        Glucose      BloodPressure    SkinThickness  
##  Min.   : 0.000   Min.   :  0.0   Min.   :  0.00   Min.   : 0.00  
##  1st Qu.: 1.000   1st Qu.: 99.0   1st Qu.: 62.00   1st Qu.: 0.00  
##  Median : 3.000   Median :117.0   Median : 72.00   Median :23.00  
##  Mean   : 3.845   Mean   :120.9   Mean   : 69.11   Mean   :20.54  
##  3rd Qu.: 6.000   3rd Qu.:140.2   3rd Qu.: 80.00   3rd Qu.:32.00  
##  Max.   :17.000   Max.   :199.0   Max.   :122.00   Max.   :99.00  
##     Insulin           BMI             Age       
##  Min.   :  0.0   Min.   : 0.00   Min.   :21.00  
##  1st Qu.:  0.0   1st Qu.:27.30   1st Qu.:24.00  
##  Median : 30.5   Median :32.00   Median :29.00  
##  Mean   : 79.8   Mean   :31.99   Mean   :33.24  
##  3rd Qu.:127.2   3rd Qu.:36.60   3rd Qu.:41.00  
##  Max.   :846.0   Max.   :67.10   Max.   :81.00

• BMI & Blood Pressure have several incorrect records/datapoints –their values cannot be “0”

• Data imputation is not recommended in this case (not realistic/convenient)

• The outliers present in the INSULIN variable are removed
• As well as the “0” values for all variables where this is applicable

AGGREGATING/CLUSTERING THE PATIENTS

Since all variables are numerical, we will try to separate the healthy from the diabetic patients using numeric cluster analysis.

The k-means algorithm seems the best choice in this case –given that there are only 2 groups expected (healthy vs. diabetic). Hierarchical clustering is better for a bigger n of clusters

Here, k-means results in a clearer visualization

After reviewing the ranges of the dataset variables, scaling of the dataset was performed –this allows obtaining a clustering that is more accurate.

Three (3) k-mean clustering runs are generated for the dataset and plotted below:

Reviewing soundness of clustering obtained

• Plotting dataset variable relations and obtaining insights:

The k-means algorithm separates patients with lower Glucose & BMI, from those that have higher values of both variables

Similarly with Blood Pressure & Skin Thickness:

Worth further questioning & exploring

Further steps: ?

Reviewing safety signals

• Select outliers visually
• Or via a protocol-specific algorithm, or using standard algorithm libraries (Hy’s law)
• Level of data aggregation: study, country, site
• Review the combined outliers and make a decision: query/don’t query, further multivariate analysis, or other analysis