Taguchi Inner and Outer Array Design for Medical Device Testing

A

Alvin Chen

Hello everyone,

We are trying to design an experiment for performance testing on a medical robotics device. Our goal is to optimize across environment variability, specifically the simulated variability of human skin. My questions relate to the Taguchi approach.

1. Is it poor practice to have a much larger outer array than inner? For example, an L16 (five factors, four levels) outer array and an L4 or L9 inner array? If not, what is the reason for this? Would a single array approach be better than the traditional inner-outer cross array?

2. Is it poor practice to combine completely independent influences within one experiment? In our case, the inner array consists of control factors for optical/imaging performance, as well as control factors for mechanical/robotic performance. Similarly, the outer array consists of both optical variables and mechanical variables (optical and mechanical properties of skin). I am fairly certain that the optical control factors will influence only the optical noise factors, and likewise for the mechanical factors. Given this, should I run two separate experiments, or is it ok to run a single larger grid?

Thank you in advance for your time.

Alvin Chen
 

Miner

Forum Moderator
Leader
Admin
The quick answer is no, it is not poor practice. Now lets discuss your options:
  • Option 1: Traditional Taguchi approach. This is the approach that you describe. It has the advantage of being easy to understand and easy to analyze. Although I advise against using the S/N ratio and just analze the mean and standard deviation.
  • Option 2: Classical approach. This involves creating a split-plot design with the control factors in the whole plot and the noise factors in the sub plot. This is the approach recommended by George E. P. Box (of Box plot fame). This results in a smaller experiment than the traditional Taguchi approach. The analysis is more difficult because there are different error structures between the whole and sub plots. In addition, the interpretation for reducing sensitivity to noise is somewhat indirect and can be somewhat confusing.
  • Option 3: Modified Taguchi approach. Use a classical fractional factorial for the inner array. Condense the outer array into a single compound variable with two levels, then run and analyze as a standard Taguchi. This has all of the advantages of the traditional Taguchi design, yet has the fewest total experimental runs. Create the compound variable as follows:
    • Set the levels of all noise factors to drive the response lower. Run the inner array.
    • Set the levels of all noise factors to drive the response higher. Run the inner array again.
I recommend option 3.

DISCLAIMER: I am qualified to discuss Designed Experiments, including Taguchi methods in non-regulated industries. However, I am not qualified to discuss Medical Device testing. If the requirements of the medical device industry impose constraints on your options, my advice may not apply.
 
Last edited:
Top Bottom