In order to be as transparent as possible Bingli has adopted a new rule to broadcast the inner working of the model and all changes to this program on an info page.
22/3/2020: NSAID is not being used in the calculations anymore
23/3/2020: ANOSMIA/AGEUSIA has been added
24/3/2020: Hightened impact of combination MRC grade 3, MRC grade 4 in combination with worsening conditions
24/3/2020: Worsening conditions of chronic pulmonary diseases ignores long onset
Bingli’s COVID 19 questionnaire is more than just a decision tree model. These have often proven to be too rigid to encompass the whole scope of all possible symptom combinations a patient can present themselves with. Instead, we have chosen for a broader perspective: symptoms, risk-groups and age categories, each with their own strategy.
• Probabilistic logic:
The chance of being infected with COVID19 is calculated by using conditional dependencies. COVID19 symptoms are given a likely-hood-ratio depicting the strength of their relationship with COVID19. The combination of all contributing symptoms will result in a posterior probability. At the moment there isn’t any exact data available (eg. Sensitivity/Specificity) on how all symptoms present themselves in COVID19 patients so approximations for these values are made. Our current model is based on literature combined with expert input.
One important notice: almost all of the symptoms are common to most infectious pulmonary diseases such as pneumonia and the flu. This model, just like every other model currently existing, is not designed to make this distinction, or even to predict your chance of having COVID19. It does, however, give a solid base to start working on predicting the patient’s overall condition , linking it to specific triage recommendations.
• Risk assessment:
Many conditions do not directly influence the chance of having the infection. They are however extremely important since they influence the disease outcome. Seeing that COVID19 information is constantly evolving, this list is appended accordingly. Destabilizing conditions are given a score based on these conditions and their presentation. Each age range has its own policy, ensuring proper care for the elder patients without diminishing the effect for young ages.
• Many-valued logic:
These three strategies are combined in a “fuzzy logic model”. This model was chosen to accommodate the uncertainty in the patient’s answer or disease presentation. Fuzzy models or sets are mathematical means of representing vagueness and imprecise information (hence the term fuzzy). One of the major strengths of the model is its flexibility. When new information of the coronavirus is discovered, this can be implemented by just adding the info in the right category, taking no more than 5 minutes. A tree model often has to be completely revisioned.
• Model optimization:
Each stage of the model has certain parameters that can be tweaked to better predict the outcome. With our validated cases, we were able to do this tweaking using extensive parameter testing, resulting in an improved model. This model remains entirely white box after this process is finished, meaning we can always trace back how the model responds to certain input symptoms.
In order to improve and validate our model, we need patient data. Sadly this is unobtainable at the moment (however we are working on it). To fill this gap we have made a Random Simulated Case (RSC) program to build simulated COVID19 patients with the risks and symptoms described above. These are being presented to doctors as we speak and results are flooding in.
• The doctor then gives us his clinical assessment. With the results of the RSC we are able to improve and finetune our model.
• The final validation will come with the injection of real life cases, something we are working on now with academic and peripheral hospitals.
If, as an healthcare professional, you want to help us improve our model, contact us at firstname.lastname@example.org