Increasing Safety of Neural Networks in Medical Devices

Neural networks are now widely used in industry for applications such as data analysis and pattern recognition. In the medical devices domain neural networks are used to detect certain medical/decease indications. For example, a potential imminent asthma insult is detected based e.g. on breathing pattern, heart rate, and a few optional additional parameters. The patient receives a warning message and can either change his behavior and/or take some medicine in order to avoid the insult. This directly increases the patient's quality of life. Although, currently medical devices mostly use neural networks to provide some guidance information or to propose some treatment or change of settings, safety and reliability of the neural network are paramount. Internal errors or influences from the environment can cause wrong inferences. This paper will describe the experiences we made and the ways we used in order to both increase safety and reliability of a neural network in a medical device. We use a combination of online and offline tests to detect undesired behavior. Online tests are performed in regular intervals during therapy and offline tests are performed when the device is not performing therapy.