DISCLAIMER:

You may not need this template. It's aimed companies that develop devices including ML algorithms. Why do
they need this template?

Due to the "blackbox" nature of many of those algorithms, their functionalities typically can't be verified
in the same sense as code reviews can verify correct coding in the world of old-school software
development. That's why they can only be validated by subjecting the model to rigorous testing on separate
test datasets. This report is the place to document all the results (e.g. sensitivity, specificity,
accuracy, ROC curve values, discussion of outlayers, edge cases, etc.). Side note: there's a disconnect
between what data science and regulatory affairs understands as validation and verification. Don't get
confused!

1. General Information

This document describes the specification of our machine learning (algorithm) model and its testing strategy.

Keep in mind that the results of the algorithm validation should also serve as input to and be referenced in
your clinical evaluation / performance evaluation report.

Regulatory references:

Relevant other documentation:

• SOP Software Development
• Software Development and Maintenance Plan
• SOUP List

2. Development Resources

For more context information, refer to the device’s software development and maintenance plan.

2.1. Developer Team

2.2. SOUP / Frameworks

2.3. Data

2.4. Development Planning

Optional: remove or add to the following sub-paragraphs as appropriate.

Intended Purpose

Reference intended use; describe intended task the model should solve.

(...)

Clinical Environment

Reference clinical evaluation plan; describe clinical context in which the model will be used.

(...)

Software Architecture

Reference software architecture (diagram); describe how the model will be embedded in the overall system and
how it will interact with other software components.

(...)

Quality Requirements

Reference clinical evaluation plan; describe quality criteria that shall apply to the model, for example,
quantify values for sensitivity / specificity, AUC, positive/negative predictive value, accuracy,
repeatability. Also, provide justification for why certain parameters may not apply and compare your
approach to the current state of the art / technology.

(...)

3. Data Management

3.1. Data Acquisition

Describe how you acquired your data (sources, inclusion / exclusion criteria, possible biases, data
protection measures, etc). Give a rough estimate of what size of datasets are required.

If applicable, reference relevant QMS processes.

(...)

3.2. Data Annotation

Describe how data is labeled and how you determined the ground truth. Further, you can describe: labeling
requirements (quality criteria), qualification / training of staff involved, quality assurance for correct
labeling. If applicable, reference relevant QMS processes.

(...)

3.3. Data Pre-Processing

Describe how you pre-processed data. Possible steps may include: anonymization / pseudonymization, removing
data elements, converting data formats or units, handling missing values or data outliers. How did you
divide data into training, validation and test data. If applicable, reference relevant QMS processes.

(...)

4. Algorithm Model Training and Description

Describe your algorithm model. For example: which type of model (e.g. CNN) is used? Optionally, add a graph
or diagram to explain the model architecture. Describe model training; e.g. selection of features, (most
important) hyperparameters, etc.

(...)

5. Algorithm Model Evaluation

Describe how testing is done: describe your test data set, which metrics are used for evaluation and which
values are deemed acceptable.

(...)

6. Conclusion

Discuss your evaluation results: why do they meet predefined quality criteria, why is the model fit for
purpose in accordance with the medical device’s intended use? Take into account possible risks, biases and
limitations.

(...)

Template Copyright openregulatory.com. See template
license.

Please don't remove this notice even if you've modified contents of this template.