As the use of Artificial Intelligence (AI) and machine learning methods expand in medical devices and HealthIT software, an oft asked question is whether the data sets used for training should be retained as part of the design history file (DHF) or other long term storage mechanisms.  SoftwareCPR partners Alan Kusinitz, Sherman Eagles, John Murray, and Brian Pate recently met to discuss this topic and arrived at several guiding principles that may be useful to manufacturers as they consider specific policy with regard to retaining training data sets.

We went into our roundtable discussion making the following assumptions:

1. A trained model or algorithm represents a design output – produced by the activities and tasks of a development team and subject to Design Controls.
2. Training set data is the method (or a portion of the overall method) by which the development team used to create the model or algorithm.
3. Design input included the required patient population distribution (e.g., age range, sex, skin pigmentation, etc.) and quantity (number of data items), minimum accuracy (e.g., sensitivity, specificity), and other user controllable factors (e.g., if imaging, resolution, etc.)

One could consider model training to be a research activity – and thus, retain very little information in the DHF.  However, this would likely create an impediment to on-going development and improvements to the medical device or HealthIT system since the team would be “blind” to previous work.  So this leads to the question:  what would a “new” development team need from the previous development team to orchestrate further development and improvements to the system?  This question illustrates precisely one of the key purposes of a DHF.

Lean Product Development

If we approach the question from a lean product development viewpoint, we might re-frame the question as: what is the minimum amount of information a “new” development team would need from the previous development team to orchestrate further development and improvements to the system?  We considered this question at the roundtable and we arrived at this list:

1. Source(s) of data items
2. The number of data items
3. How “ground truth” is annotated or associated with data items
4. Patient population distribution
5. Validation records

The assumption is that with this information, the manufacturer could re-create a new model or algorithm with equivalent performance as the original model, where equivalent performance is defined in design validation terms from the design input assumptions above.  By this approach and argument, we could envision an approach of not retaining the actual training set data in its original form.

We hopes this provides useful input to your planning for your AI/ML products.