Almost every daily activity of the diabetologist has to deal with digitalization, from electronic medical records to imaging diagnostics, from laboratory references to the various software programs for administrative practices and certifications, and from glucose level data downloaded from glucometers to sensors for continuous glucose monitoring and insulin pumps. Through the analysis of these data, the diabetologists make therapeutic decisions, and the more complete and numerous the data, the more they need informatics tools to guide them in the analysis, help them identify specific patterns for detecting glycemic abnormalities, understand the possible causes, and adopt the appropriate therapeutic strategies to correct these abnormalities [9,10].

In the early days, data collection and management for each individual patient occurred manually at the hospital by physically connecting cables or wireless devices to the local software/network. Currently, thanks to the technological evolution, these data can be transmitted automatically from the glucometers via “cloud” or on data integration platforms that collect elements from different devices and provide standardized reports [11,12]. If not properly managed, this large amount of data is likely to overwhelm both the patient and the health care professional, and the professionals’ experience in reading these reports is certainly the added value that increases their potential. Advanced informatics tools can simplify the analysis and provide suggestions to guide the clinical decisions of the physician [13]. With experience and technology coming together, the analysis time can be increasingly reduced and more appropriate and correct data-driven therapeutic decisions can be made, thanks to the greater customization strategies of the therapy [14-21].

In this context, the use of AI, and in particular, the use of machine learning allows an important step forward compared with the traditional data analysis techniques (ie, graphs representing a picture of reality, which are very useful and precise but are static and outdated). Through the automatic identification of specific patterns within the data and through inductive reasoning typical of the human mind, machine learning can highlight correlations that lead to “predictions,” without being programmed in advance to carry out this activity [19].

In a not too distant future, AI, thanks to algorithms that can enable learning and improving the machine’s own abilities independently, will offer effective solutions to satisfy the most disparate needs and will be able to deal with problems that today may seem insurmountable obstacles, for the benefit of the community [22]. In the field of clinical diabetology, these tools could have multiple potential implications, including the identification of new risk factors for the onset of diabetes (through the evaluation of large databases related to the general population) or revealing unsuspected subjects who are at a high risk of complications (by cross-checking clinical and administrative records of patients with diabetes) and identifying the behavioral and therapeutic variables that are most closely related to the progression of a specific complication.

There are many examples of collaboration between pharmaceutical companies, information technology companies, scientific institutes, and universities that exploit large complex datasets (the so-called big data) with the aim of improving the treatment of type 2 diabetes mellitus (T2DM) and of unraveling physiopathological mechanisms through the integration of data from biological, demographic, clinical, environmental, and genomic sources [23].

The ability to elaborate a large and heterogeneous amount of data, even in real time, through increasingly powerful algorithms, allows us to extract knowledge and exponentially make predictive assessments on the behavior of individuals as well as, more generally, assume decisions for the entire community [24,25]. This means that the use of AI will make it possible to transform the immediate data (descriptive) into knowledge of the factors that “condition” behavior and correlations (predictive) [25-27] up to the identification of the key factors that can facilitate an improvement in the expected results (prescriptive) [28,29]. In this scenario, one of the most interesting applications would be to identify the variables that could be related to greater responsiveness to a specific drug because this approach would open the door to a truly personalized medicine that uses the right drug for the right person, with obviously greater efficacy, improved outcomes, and containment of costs. Furthermore, predictive analysis techniques based on AI could be used to identify which group of patients would require more attention and which strategies would be more effective, depending on the individual patient, thereby allowing more efficient methods of personal care with lower costs and better outcomes. This concept of risk stratification, which considers all the individual characteristics of the patient (ie, clinical aspects, genetic data, lifestyle, environmental factors), is the basis of the modern model of clinical governance called “population health management.” Today, the risk stratification of the population results from the extraction of historical expenditure data (eg, inpatient and outpatient admissions, diagnosis, pharmaceutical expenditure) and this presents the following series of biases: weak clinical validity; temporal misalignment between the extraction, the analysis, and the health status; and absence of information on the real socioeconomic condition and behavioral health. In this way, the data represent “satisfied demand” rather than “real needs.” The new challenge is the use of big data and business analytics systems to activate risk stratification models based on real health status, integrated use of multiple sources, and collaboration between professionals (ie, care team and data manager in a data mining process) [30,31].

Another aspect that characterizes the ongoing medical revolution is the development and increasing importance of biotechnologies used in genetics and in the so-called “omics” sciences. These new techniques have been showing that the complex pathophysiological processes underlying type 1 diabetes mellitus (T1DM), T2DM, and gestational diabetes mellitus are caused by disturbances in the gene expression that lead to alterations in the processes within the organs involved in glucose homeostasis [32]. The complexity of the system is exacerbated by the fact that the relative contribution of each component is highly individual. Understanding the molecular mechanisms underlying these interactions is crucial for developing new personalized prevention and treatment strategies [32,33].

Finally, the ever-increasing use of social networks and apps has major implications in the diabetes care. On one hand, health professionals must change the way they communicate and keep up with the times by always acquiring new technological and communication skills to deal with and manage the changes taking place, while on the other hand, the new technologies could be used as motivational and coaching tools in support of the traditional educational activities and as an alternative data source. Notably, the areas that influence the state of health are only 10% due to the medical care and health status, whereas the major role is instead played by lifestyle and behaviors as well as by genetics [34-36]. For this reason, to obtain a more precise picture of the state of health of the general population or of specific subgroups, we cannot limit ourselves to collecting and analyzing only health data, but we must consider what comes from the world of social networking and smartphone apps, despite all the known limitations and criticalities.

Cases of diabetes have been identified within large heterogeneous databases, and new risk factors for diabetes have been identified. An interesting multi-database retrospective study on the identification of cases of T2DM [17] strategically used the European Medical Information Framework Project database, a European project for the efficient reuse of health data for epidemiological research [38]. This database collects health information of about 52 million European citizens by using heterogeneous sources and by acting as a support for the execution of high-quality multinational observational studies, based on populations with large sample sizes and otherwise inconceivable follow-ups. Subjects with T2DM were identified using a complex algorithm strategy in 8 different health data sources, and the strengths and limitations of each data source were revealed during the creation of a model that ensures the interoperability of systems of heterogeneous electronic medical records. These efforts represent a methodological advancement for carrying out studies of multinational and multi-data sources, thereby providing sufficient information for the contextualization and correct interpretation of the results and generating transparent and reusable documentation.

Additionally, a study conducted by researchers from the University of New York and Philadelphia described a new “data-driven” approach of population health management based on the use of machine learning techniques to develop predictive models and identify risk factors for the onset of T2DM [39]. They based this model on administrative data concerning health services, pharmaceutical records, insurance databases, citizens’ access to the different care facilities, and laboratory results from 4.1 million individuals over a period of 4 years and a total of 42,000 variables. This model has been able to identify new risk factors for the appearance of diabetes with a predictive probability of at least 50% higher than the models based on the known traditional risk factors, thereby avoiding the costs of performing a screening. The identification through new predictive models based on machine learning of the part of the population that is at the highest risk of developing diabetes will be able to generate clinical hypotheses for the identification of new risk factors and to implement more targeted cost-effective interventions.

A study of 65,000 newly diagnosed patients with T2DM [40] estimated that 10% of the subjects absorb 68% of the resources. The task of predictive analytical methods based on big data is precisely that of identifying the 10% of the subjects most at risk, on whom the treatment needs to be intensified in order to improve the health outcomes at lower costs.

An interesting Japanese study [41] has developed a practical framework for phenotyping T2DM by using both specialized knowledge and an approach based on machine learning (in particular, the support vector machine) to develop 2 phenotyping algorithms with data extracted from electronic medical records: one with high sensitivity for screening and the other to identify the subjects for research. Both algorithms showed superior performance compared to the basic algorithms, thus suggesting the possibility of using the proposed framework to conduct appropriate research depending on the objective.

The use of predictive models based on AI has demonstrated the feasibility of identifying individuals with the highest probability of having undiagnosed diabetes through clinical data that can be easily obtained from different databases [42], thereby exploiting the potential of machine learning algorithms, including neural networks, as tools for diagnosing diabetes [43]. Machine learning and pattern recognition are tools of enormous interest as they are promising in improving the sensitivity and specificity of disease detection and diagnosis, and they appear to be able to reduce the potential for human error in decision-making. An exemplar application is the creation of surveillance algorithms that are able to detect diabetes and, in particular, to distinguish T1DM from T2DM by using structured electronic medical records [14]. The extraction of data from the electronic database of a 4-year long, multi-sectorial, and multi-specialist outpatient practice allowed the inclusion of approximately 700,000 patients. Possible cases of diabetes were reported using laboratory test results, diagnostic codes, and prescriptions. More cases of diabetes were captured by taking advantage of the entire range of data acquired from the records compared to that captured by the analysis of only administrative databases, thus increasing the sensitivity of the method. The application of these algorithms to electronic folders has the potential to provide timely and clinically detailed information on large numbers at low cost and nearly in real time. Electronic records will probably become increasingly important sources for the surveillance of public health and for the definition of more targeted health policies [15,44].

Diabetic retinopathy, in particular, is a chronic complication of diabetes. The automated classification of diabetic retinopathy has potential benefits such as increasing the efficiency, reproducibility, and coverage of screening programs; reducing obstacles to accessing and improving results; and providing early diagnosis and treatment. To maximize the clinical utility of the automated classification, an algorithm designed to detect specific lesions or to predict the presence of any level of diabetic retinopathy was recently developed [45]. In this study, deep learning, together with visual and pattern recognition techniques, has allowed the identification of the desired features with the highest predictive value directly from the images on a large set of labeled example data. These results show that deep neural networks can be “trained” by using large data sets and without having to specify lesion-based features to identify diabetic retinopathy or diabetic macular edema in fundus images with high sensitivity and high specificity.

Cost-effectiveness studies of health interventions have been performed previously. An Italian group performed a retrospective analysis through the cross-examination of large clinical and administrative databases with the aim of quantifying the relationship between health care costs attributable to diabetes and the level of glucose control [46]. The results indicated that glycemic control (expressed by hemoglobin A_1c [HbA_1c] levels) is a useful surrogate not only to estimate the odds of developing diabetes-related complications but also to estimate the costs associated with health care. The integration of administrative and clinical databases seems to be suitable to show that an appropriate management of diabetes can allow a better allocation of resources.

The ongoing medical revolution is strongly linked to the spread of digital health, the new software for AI, the use of big data to make more appropriate data-driven decisions, and the even greater focus on predictive, preventive, personalized, and participatory medicine [6]. Each of these elements has important repercussions in the management of complex and widespread chronic diseases such as diabetes. Table 1 summarizes the advantages and weaknesses attributed to each area of application, especially considering the challenges that diabetology faces, according to our professional experience and opinion.

The AMD Annals represent a periodic publication that has allowed, annually since 2006, to assess the care profiles of people with T1DM and T2DM who have been admitted in Italian diabetes care centers. A wide network of diabetology services equipped with a computerized clinical record system, used for the normal management of patients in charge, has a software program provided by AMD that allows the extraction of a standardized set of clinical information (ie, Data File AMD). The database obtained is used to calculate the quality of the care indicators, both centrally and locally. In this way, a benchmarking activity is promoted based on the comparison of one’s own performance with that recorded at the national level (best performers approach). This activity reflects the professional performance of the diabetologists—it is perfectly in line with the recommendations of the National Diabetes Plan. It is very cost-effective and has produced a systematic improvement over the years of all the indicators considered [52].

In addition, the AMD Annals database represents a valuable source of observational research data. In fact, it has allowed the investigation of many key aspects over time, such as the care of the elderly patients, gender medicine, cardiovascular, renal, and hepatic issues, the appropriateness of the use of drugs, and glycemic self-control, thereby providing a realistic picture of the work of the diabetology services. Notably, the analysis of the Annals database has allowed the identification of critical areas and, therefore, the timely activation of processes of improvement, in a logic of continuous quality enhancement, that is, a process of periodical performance assessment of diabetes centers on data collection and quality of care according to standardized indicators [53-57]. It is easy, at this point, to visualize that the Annals project “photographs” the history of the evolution of clinical diabetology and has allowed, for its own conception and structuring, each individual center to self-evaluate and improve in real time. Precisely, this careful measurement of behaviors and results is capable of triggering a dynamic evolution of an entire class of professionals, thus representing an irreplaceable benefit.

With innovative and far-sighted thinking, AMD has created this database, which today represents an unparalleled heritage in a cultural civilization that is increasingly aware of the importance of information and the value of a large and accurate collection of data. In this new era, the world is being organized to take advantage of the ever-increasing databases to rely on technology and use them as assets to be interpreted to facilitate and accelerate important decisions in every field. It is clear that it is increasingly important to have a quality data collection that is increasingly “clean,” and it is fundamental to highlight the need for all professionals to be trained in the culture of big data and its correct recording. In fact, given that decisions can be made on the results of data processing, it is essential to raise the problem of the truthfulness of the data on which the analyses are based, to have a measure of the reliability of the results, and to be aware of them. Further, in this area, AMD has invested time and resources in activating a process—born more than 10 years ago—of culture on the quality of the data, which involved many partners who participated in the Annals project [53-57]. This is why AMD feels ready to investigate the topic, as it has an advantageous background, and it wants to project itself in a competent and proactive way in the world of big data and AIs that represent a new cultural challenge in the scientific field.

The DIA&INT (Diabetes Intelligence) is an AMD project that rises in the setting of exploring new methods of analysis of unconventional data, and it can be considered as the first experiment in analyzing unstructured data by using the business intelligence method (a precursor of machine learning systems).

In a comprehensive vision of clinical activity, AMD focuses on the value of the skills that identify a diabetologist and make it a decisive tool in the care process. To achieve this objective, AMD implemented a business intelligence project called DIA&INT [47]. This project was mainly aimed at encouraging the implementation of the CCM in an “evidence-based” way by highlighting the direct link between the activities carried out by the diabetologist and their expected outcomes in order to highlight those activities that are essential to obtain the best outcomes in modern diabetology. Furthermore, this method may contribute to extrapolating the actions that could optimize the scarce resources and it may represent a valid support for the institutional choices in the current health system revision.

Admittedly, DIA&INT has been designed to respond to the need for establishing the qualitative dimension of the diabetologist’s performance and the specific contribution of the many factors participating in the clinical decision in the real world. To do this, AMD has chosen to use the following accredited tools.

The resulting “program” has been structured to standardize the information and define activities and competences, as implicitly described in the official guidelines, which are measurable and comparable, with processing methods different from those of classical epidemiology, but necessary to perform predictive and prescriptive assessments. The methodology used [47] exploited the collective intelligence of the diabetologists who participated in a survey, expressing their opinion on a complex node such as the definition of the integrated benefit of certain activities in a personalized and weighted way. The intention was to “display” how diabetologists think when carrying out clinical judgment.

Through this approach, 2 main concepts were outlined: the requirements (priority, specificity, frequency, and multidimensional analysis) and the decision-making elements for “if…then” choices based on a priori knowledge (medical, regulatory, ethical, psychosocial, etc) and the ability to propose a dynamic personalized treatment project to the person with diabetes that does not yet involve healing and that requires an active personal effort. By doing so, DIA&INT has selected the necessary skills and activities in medical practice, which have an impact on the evolution of the quality of the “health product” and the outcomes. DIA&INT produced the Core Competence curriculum of the diabetologist [58] and measured the impact of the activities on the experiences [47].