No abstract

If modern artificial intelligence (AI) comes often misunderstood, this is mainly due to the fact that, historically, it is solely tied to the way human brains work and think. New machine learning (ML) algorithms, instead, learn now by processing massive piles of data. This process enables machines to adapt to real-world situations, as well as to propose suggestions on how to classify and interpret a variety of different real phenomena. Simply speaking, the deployment of modern ML systems into critical applications is directly influenced by the way training data are organized and modeled [1-3]. Hence, while those modern algorithms rapidly sift through huge datasets, loaded with millions of information, a thoughtfully designed AI, beyond its ML-based core, should never disregard the fact that algorithms that learn are, for now, just another form of machine instruction, still guided and influenced by the potential and the limitations that training data carry with them. In other words, even when we train algorithms to learn basic associations that can then be used to approximate,

On 21 February 2020, a violent COVID-19 outbreak, which was initially concentrated in Lombardy before infecting some surrounding regions exploded in Italy. Shortly after, on 9 March, the Italian Government imposed severe restrictions on its citizens, including a ban on traveling to other parts of the country. No travel, no virus spread. Many regions, such as those in southern Italy, were spared. Then, in June 2020, under pressure for the economy to reopen, many lockdown measures were relaxed, including the ban on interregional travel. As a result, the virus traveled for hundreds of kilometers, from north to south, with the effect that areas without infections, receiving visitors from infected areas, became infected. This resulted in a sharp increase in the number of infected people; i.e., the daily count of new positive cases, when comparing measurements from the beginning of July to those from at the middle of September, rose significantly in almost all the Italian regions. Upon confirmation of the effect of Italian domestic tourism on the virus spread, three computational models of increasing complexity (linear, negative binomial regression, and cognitive) have been compared in this study, with the aim of identifying the one that better correlates the relationship between Italian tourist flows during the summer of 2020 and the resurgence of COVID-19 cases across the country. Results show that the cognitive model has more potential than the others, yet has relevant limitations. The models should be considered as a relevant starting point for the study of this phenomenon, even if there is still room to further develop them up to a point where they become able to capture all the various and complex spread patterns of this disease.

ObjectivesCOVID-19’s second wave started a debate on the potential role of schools as a primary factor in the contagion resurgence. Two opposite positions appeared: those convinced that schools played a major role in spreading SARS-CoV-2 infections and those who were not. We studied the growth rate of the total number of SARS-CoV-2 infections in all the Italian regions, before and after the school reopening (September–October 2020), investigating the hypothesis of an association between schools and the resurgence of the virus.MethodsUsing a Bayesian piecewise linear regression to scrutinise the number of daily SARS-CoV-2 infections in each region, we looked for an estimate of a changepoint in the growth rate of those confirmed cases. We compared the changepoints with the school opening dates, for each Italian region. The regression allows to discuss the change in steepness of the infection curve, before and after the changepoint.ResultsIn 15 out of 21 Italian regions (71%), an estimated change in the rate of growth of the total number of daily SARS-CoV-2 infection cases occurred after an average of 16.66 days (95% CI 14.47 to 18.73) since the school reopening. The number of days required for the SARS-CoV-2 daily cases to double went from an average of 47.50 days (95% CI 37.18 to 57.61) before the changepoint to an average of 7.72 days (95% CI 7.00 to 8.48) after it.ConclusionStudying the rate of growth of daily SARS-CoV-2 cases in all the regions provides some evidence in favour of a link between school reopening and the resurgence of the virus. The number of factors that could have played a role is too many to give a definitive answer. Still, the temporal correspondence warrants further systematic experiments to investigate on potential confounders that could clarify how much reopening schools mattered.

Deep learning models are tools for data analysis suitable for approximating (non-linear) relationships among variables for the best prediction of an outcome. While these models can be used to answer many important questions, their utility is still harshly criticized, being extremely challenging to identify which data descriptors are the most adequate to represent a given specific phenomenon of interest. With a recent experience in the development of a deep learning model designed to detect failures in mechanical water meter devices, we have learnt that a sensible deterioration of the prediction accuracy can occur if one tries to train a deep learning model by adding specific device descriptors, based on categorical data. This can happen because of an excessive increase in the dimensions of the data, with a correspondent loss of statistical significance. After several unsuccessful experiments conducted with alternative methodologies that either permit to reduce the data space dimensionality or employ more traditional machine learning algorithms, we changed the training strategy, reconsidering that categorical data, in the light of a Pareto analysis. In essence, we used those categorical descriptors, not as an input on which to train our deep learning model, but as a tool to give a new shape to the dataset, based on the Pareto rule. With this data adjustment, we trained a more performative deep learning model able to detect defective water meter devices with a prediction accuracy in the range 87–90%, even in the presence of categorical descriptors.