The application was thus nearly ten days ahead of the WHO. The algorithm used is able to analyse health newsletters or forums for weak signals. Based on its expertise (on other diseases including Zika), the engine established the starting point of contagion and a prediction of the cities that would then be reached.
An estimate of the global spread had also been calculated by the University of Southampton as early as 15th January.
Thus, Artificial Intelligence, without replacing human decisions, proved to be an effective decision-making tool. As the confidence index was not known, the Centres for Disease Control and Prevention and the WHO did not have confidence in this predictive model...
Complexity of modeling
At present, the world is living under lockdown and AI can help prepare for the announced release around mid-May.
Indeed, while it is "easy" to decide to confine a few billion individuals, it is infinitely more complex to determine how to allow a safe return to normality. It is likely that a schedule over several weeks or months will have to be worked out from 11th May onwards.
In the propagation mode, there are still many grey areas or parameters to be taken into account:
- effective duration of the contagiousness of the virus in the air or on inert surfaces,
- immunisation time of a cured patient
- exact proportion of asymptomatic patients and their impact in the spread of the virus
- the exact duration of a patient's contagion.
Moreover, the impossibility of testing all the inhabitants of a country introduces a degree of uncertainty into the models.
Are only cured patients allowed to be discharged (while incorporating the previous immunisation element)? Who should be tested and with what test?
PCR tests are considered to be the most effective at present. They make it possible to confirm that a patient is indeed infected with the coronavirus. Serological tests can be used to determine whether the patient has suffered from and been cured of an SARS-COV2 infection.
With this incomplete and time-varying information, it is necessary to provide the public authorities with elements that make it possible to determine whether deconfinement should be carried out by typology of cases (cured, not affected), by age group (the risk of death seems to vary according to age), by geographical area and by integrating travel patterns.
Is it, for example, possible to be "deconfined" in Normandy and go to work in Paris?
Finally, the projections must integrate the international dimension. The revival of transport requires that coordinated approaches be envisaged in all countries. However, each country is not affected in the same way and practices differ in terms of testing, containment and treatment.
At present, there are various databases concerning COVID-19, although we know relatively little about the real epidemiological incidence.
Currently the most studied epidemiological data concern :
- the rate of patients hospitalised for COVID-19 (PCR positive)
- the ratio of positivity of PCR rates achieved at certain well-identified sites.
- the number of deaths related to COVID-19
However, it is a pathology for which nearly 80% of the citizens affected do not use hospitals and do not, for the moment, benefit from tests to confirm the diagnosis.
In the absence of certainties as to the percentages of cases that will require hospital services, the most interesting data would come from the registers filled in by municipal medical services.
At present, however, there are few common databases for the identification of probable or suspected cases of COVID-19.
One of these is the Sentinelle network, composed of approximately 1,400 doctors (2.1% of general practitioners and 4.3% of paediatricians), spread throughout the territory, which, in order to meet the needs of the moment, records acute respiratory infections, all or part of which are linked to the COVID-19 epidemic. It is therefore neither exhaustive nor suitable for monitoring COVID-19 patients.
From the point of view of collecting data directly from the population, France has the GrippeNet registry, transformed into the COVIDnet network to meet current needs.
However, it is currently only includes 7,200 voluntary responses to the questionnaire.
The AI as back-up
The parameters for managing the situation related to the Coronavirus are particularly numerous, with uncertainties about certain factors, unknown data and others correlated to elements not under the control of a given country.
Consequently, it is extremely complex to imagine a secure exit model from lockdown that the public authorities could rely on.
On the other hand, Artificial Intelligence and digital technology can provide effective responses.
The information that such an AI system will need is obviously all the information that LRAs have on the dates of entry of patients, their evolution, the data from the PCR tests performed and the cures.
Serological tests will also provide a map of the patients cured and their geographical location.
However, the voluntary contribution of citizens will also be essential. Indeed, most of the proposed mobile applications are diagnostic applications that can be used to refer the patient to a doctor or not. Or for remote monitoring of patients who have previously consulted a practitioner. But they do not record the patient's situation.
The "Stop-Covid" application considered by the government would only provide a partial answer. According to current information, it would be based on the carrier's actual knowledge of his/her condition and the local risk of contamination he/she would run. However, this only includes a small proportion of citizens and does not allow for comprehensive mapping or planning. The StopCovid data, among others, would therefore feed an AI platform.
The ideal tool would consist of coupling an application for diagnostic purposes with an application for self-census purposes.
The probabilistic diagnosis will then be progressively refined (and confirmed or specified) by the evolution of the patient's state of health and by the addition of the results of complementary examinations (PCR, serology, chest CT scan >72H, ...).
This first step would provide a consistent database on the evolution of symptoms.
The self-census would provide a voluntary (and anonymous) mapping of citizens. Even partial, this tool will feed a predictive Artificial Intelligence model. The latest polls show that citizens are ready to accept this type of tool if they are at the service of the health of the entire population.
Moreover, such an application allows everyone to know whether an area is "high risk" or not. Such as can be found for the mapping of malaria, which allows for the adaptation of preventive drug treatment (prophylaxis).
Enriched by these elements, an Artificial Intelligence application would be able to develop several dynamic scenarios that can evolve as knowledge is acquired.
For example, it would be inadvisable, or even forbidden, for a subject who has not contracted COVID-19 (negative serology) to travel to an area where the epidemic is still very present. On the other hand, there is no reason (subject to the duration of the immunisation) to advise a patient who has recovered (positive serology) not to go there, especially to work.
The eligible technologies are numerous: for example, machine learning, Bayesian model, constraint propagation, as well as Graphe Database in order to have a model highlighting interconnections, flow management, risk areas, possible clusters, ...
The complexity of the tool to be built is not so much in the volume of data as in the variables that influence the result and their confidence range. The input values are based on all the factors previously mentioned in order to generate, on the one hand, a sequencing of the output of the containment (by city or by region), then the consequences of moving individual profiles (contaminated, healed, etc...) from one area to another).
The interest is thus to be able to have a real-time tool that can adapt to the different factors provided.
The reliability of the model therefore resides mainly in the self-census.
The integration of complementary information or actions such as the wearing of masks for all (proposed by the medical academy) can also make the model evolve and highlight different strategies for exiting lockdown.
Thus, by varying hypotheses or variations in stocks (masks, PCR, serology), the AI algorithm will make it possible to measure the risks of the different scenarios and will be able to justify the consequences of this or that decision (explicability).
Several doctors have recently expressed their confidence in an approach exploiting Artificial Intelligence. In France, we also have a recognised know-how and many companies capable of responding quickly to this emergency. The development of such a solution therefore seems essential and necessary, especially as the epidemic could return (rebound effects), if the immunisation rate is not reached or if immunisation is not complete and vaccines are slow to be administered.
Without the use of such tools, it will not be possible to produce an explainable predictive model to quickly exit lockdown. The challenge is therefore as much a health issue as an economic one.
However, the decision remains political for the choice of an application for the entire population.
Laurent Cervoni, AI Director at Talan
If you're interested in Artificial Intelligence, you should read this article about Power and creativity in AI.