The Klinik AI uses Bayesian probabilistic models for generating inference on probable diagnoses and a suggested level of urgency. Bayesian logic is the method that actual human clinicians use in their clinical decision making; to build the differential of diagnoses and decide on actions based on probability thresholds – a form of probabilistic and statistical calculations, in fact. This Bayesian inference logic is currently considered to be the golden standard in digital triaging and medical AI, as it is the safest way to build a vast range intelligence engine that can carry out the needed inference while each and every outcome can be traced back and proven to be both clinically and functionally appropriate. One more advantage of Bayesian inference is that it can be controlled and further developed by clinicians.

The system uses four urgency levels, which utilised by the professionals:

• Suitable for self-care: the patient does primarily not need a GP appointment, but has symptoms that could benefit from symptomatic treatment
• Non-urgent: a routine appointment is suggested (non-serious/chronic/subchronic issue)
• Urgent: appointment during next few days is suggested
• Emergency: asap evaluation by a GP/A&E is suggested

The system as a healthtech software and a medical device has been evaluated throughout for safety performance. As per our DCB0129 Safety Case the solution has been assessed to be safe for use. In our Clinical Evaluation for EU CE I marking, the performance has been accepted  and the product is safe to use in a clinical setting.

In a recent, scientific clinical study we were able to show that the system was in A&E/OOH setting regarding urgency sensitivity comparable to that of A&E clinicians, has a good diagnostic performance (60 % exact or close match) and very high usability (highest quartile using SUS scale).

Through our continuous monitoring from our data ops and by clinicians, we can see that < 2 % of emergency cases are caught undetected. This reflects the current safety margin which has been built in, and is generally much safer than triage actions carried out on the telephone. The distribution of urgent/non-urgent cases reflects very closely to that of the clinicians that provide feedback on system performance. The current distribution of urgencies is as follows:

1. Suitable for self-care: 14 %
2. Non-urgent: 27 %
3. Urgent: 42 %
4. Emergency: 16 %

Our current diagnostic range is among the leading ones on a global scale, especially in a primary care setting. The system uses and recognises over 1000 symptoms, diagnoses and other medical concepts and is actually currently the most widely used when it comes to geographical coverage. This makes it even more robust and advanced in medical and clinical terms.

 To be more specific, our AI uses Bayesian probability based models for inferring probable diagnoses using likelihood ratios and post-probability calculations. Through an algorithmic approach, this is used conjointly with a vast medical database that we’ve created and we continuously develop, to create dynamic history-taking forms which enable patient issues to be assessed in a reductionary way, mimicking the methodology of a doctor. 

The AI keeps track of probable conditions and asks for subsequent information, until triage is inferred from the differential and potential red flags. The AI alerts the patient if emergency classification of urgency is detected.

This is currently state-of-the-art as is described in our Clinical Evaluation for our CE I marking. We are continuously investigating possibilities of enhancing the system using machine learning techniques and generative AI.

Calculation of conditions 

The Klinik AI calculates using Bayesian logic and probabilistic modelling a list of most probable conditions based on the patient inputs (symptoms selected etc.). This is presented in the Dashboard as a probability ordered differential diagnosis list. A cut-off threshold of over 15 % probability is used for selection into the differential. The case urgency is determined from the differential as the highest priority of the conditions, which can be raised by any selected red flags.