Prototypes and considerations for safety assessment

One of the main impacts of diagnostic tools based on imaging and artificial intelligence is the hit rate of the result. In other words, the possibility that the diagnosis is wrong.

According to research in the journal 'JAMA Dermatology', melanoma detection applications incorrectly diagnose more than 30% of cases, giving benign cases of skin cancer as benign when they really are not. When talking about diagnostic confidence, it is necessary to distinguish between two concepts:

  • False Positive: A false positive does not in itself represent any danger, apart from adding unnecessary stress, initial periods of shock, anger or a feeling of having wasted time and opportunities, aggravating the patient's psychic state. However, the danger is that the patient may take hasty action, such as expensive and invasive tests or treatments that may be harmful in the medium/long term.
  • False Negative: This case has a greater influence than the previous case, as a week's delay can decrease the survival rate by 30%, with possible fatal consequences. On the other hand, a false negative can lead to a much higher cost of treatment at the time of late detection than earlier detection. Therefore, care should be taken in case of doubt and more people/apps should be used to corroborate the opinion.

One of the main solutions proposed is that, through our app, patients, apart from receiving an assessment from Artificial Intelligence, will be able to contact their doctor directly, without any kind of intermediary. In this way, the doctor receives the photo that has been analysed and can give a second opinion or request a face-to-face consultation if they have any doubts. The aim of this is to achieve a higher success rate in the event of doubt, as the opinion of several people/apps on the same possible melanoma will be available.

The second solution proposed to solve the problem of misdiagnosis is to research and use advanced algorithm development techniques. The idea is to use a large database to develop code that correctly classifies the images, in combination with techniques that avoid over-fitting the results to the database used.