In today’s healthcare environment, both physicians and patients are the beneficiaries of great scientific and technologic advances. However, modern medicine remains dependent on the accurate evaluation of a patient’s health state, recognizing that disease is a process that evolves over time and interacts with many factors unique to that patient. As a result, the most critical factors that must be determined are as follows: (1) what is the disease path/process that a patient is exhibiting, i.e., the diagnosis, and (2) how far along that path has the patient progressed, i.e., disease staging. The entire premise of precision medicine requires the ability to determine both diagnosis and disease stage with a high degree of accuracy so that the appropriate selection of treatment pathway can be applied to optimize outcome (1).

In spite of the increased access to clinical data resources, there remains a lack of standardization of definitions and formats that negatively affects interoperability of health records. This further limits the development of comprehensive and universally adopted standards for diagnosis and treatment using evidence-based medicine and reduces the accuracy in establishing a patient’s true health condition. The US Institute of Medicine estimates that approximately 10% of all diagnoses are incorrect. This estimate includes both type 1 errors (false-positive diagnoses) and type 2 errors (false-negative diagnoses) and leads to unnecessary patient deaths (estimated at 40,000–80,000 per year in the United States) with an estimated cost of $750 billion per year in the United States. Death due to medical errors has been estimated to be between 200,000 and 400,000 per year (the United States) and places this cause just behind heart disease and cancer with misdiagnosis and missed diagnosis representing between 15 and 25% of all medical errors (2). In addition, these numbers significantly underestimate healthcare costs associated with non-fatal misdiagnoses and missed diagnoses that result in inappropriate but non-fatal treatment of patients (3).

Heart failure (HF) is one of the most important public health problems encountered in the developed world. Despite the introduction of new treatments, HF is associated with high costs and poor outcomes. Many of the patients exhibit non-specific symptoms, which makes it difficult to identify HF and distinguish it from other conditions. Thus, many patients may have undiagnosed HF, or even when diagnosed, other undiagnosed concomitant conditions, as diabetes that is common in patients with acute HF, may confound the HF diagnosis. It is important to identify these patients and provide access to appropriate treatment to reduce mortality, improve healthcare, and reduce costs derived from undiagnosed disease (4, 5).

The development of rule-based systems for organizing and presenting information extracted from multiple sources of standardized data related to large patient populations that facilitate problem solving and decision making in the clinical field is the first step toward establishing Precision Public Health (6).

In Italy, common and interoperable syntaxes have been developed for the National Health System (NHS) subcomponents (hospitalization, outpatient specialized treatment, and others). These enable collection, in specific data flows, of all information (individual characteristics, treatments/drugs, tariffs, co-pay fee exemption) generated when citizens interact with the NHS. The use of these standardized information models facilitates their integration from different data sources. This enables the development of more precise diagnostic criteria; moreover performing the analysis on large databases that include many different healthcare processes and events, enhances patient phenotype identification/assignment.

Although the primary purpose of these sources of data is to collect, store, and exchange patient information to allow administrative management of the NHS (7), in this manner, they may be also used to identify patient cohorts by first deriving rule-based phenotyping from the combination of standards, archetypes, ontology, and reasoning (8) and then application of these algorithms to screen all patients in the database for potential candidates with as yet undiagnosed disease.

This effort to develop phenotyping algorithms requires consistency in data (archetypes level) and knowledge level (ontology) simultaneously because diagnostic criteria (reasoning) are calculated from raw data but also need derived variables not directly available in original data (8, 9).

Furthermore, administrative data represent not only a patients’ clinical characteristics but also the manner in which information is collected and recorded during the healthcare process events, ranging from inpatient admission, inpatient discharge, outpatient visit, emergency department visit, and ambulatory surgery (10). The propensity of the variables for clustering based on their association with specific healthcare process events results in better derivation of phenotyping algorithms for cohort/classification (10).

This study was undertaken to establish and validate the CARPEDIEM approach and algorithm for identifying those patients presenting with or at risk for HF to streamline the development of patient registries for supporting the proactive approach inspired by the Chronic Care Model, developed by Wagner in the late 1990s (11). This model assumes that improvement in care requires an approach that incorporates patient, provider, and system level interventions and that better coordination of care may also reduce medical expenditures, especially for persons with multiple chronic conditions (12).

Heart failure happens when the heart cannot pump enough blood and oxygen to support other organs in the body. It is a complex condition that can result from coronary artery disease, heart attack, cardiomyopathy, high blood pressure, diabetes, thyroid disease, kidney disease, valve disease, etc. and requires a specialist for both accurate diagnosis and therapeutic intervention. Globally it remains the single largest cause of death. The prevalence of this condition in Italy is about 1.25% [95% confidence interval (CI), 1.23–1.27], and the incidence rate was 1.99 per 1000 person-years (95% CI, 1.81–2.08) that is expected to grow (20).

As with most chronic diseases, in the absence of effective prevention, early detection and appropriate treatment are critical for patient management and quality of life.

The estimated cost for HF in Italy is $39.2 billion annually (21).

Because of the variability in cause, in presentation, in patterns of comorbidities, etc. and especially in the differences in presentation of symptoms between males and females, it is critical that patients be identified who are at risk or in early stages of this disease so that proper diagnosis and treatment can be obtained. The complexity associated with HF makes it significantly susceptible to the errors of misdiagnosis and missed diagnosis, especially in patients who are not engaged with a cardiac specialist.

Understanding the complexity associated with the healthcare process is vital for achieving scalable and high-throughput phenotyping algorithms that might operate on the administrative data contained in the NHS data repositories (Figure 4).

A significant aspect of the CARPEDIEM approach is that it is based on the data that exist with patient information flow databases and is not reliant on the electronic health record to accomplish its goals.

The nature of HF, not considered to be a discrete condition, but a “complex clinical syndrome” (7, 22) characterized by high comorbidity burdens, makes the integration among different healthcare process events mandatory in defining HF cases.

Moreover, recent literature stated that most if not all age-related diseases share low-grade chronic inflammation termed “Inflammaging,” which is a highly significant risk factor for both morbidity and mortality in elderly people (23).

Chronic inflammation can be prevented and cured (24), but there is an urgent need to better understand the links between molecular features and phenotype to develop strategies to deal with their implications for the individual patient (25).

Currently used organizational structure of the taxonomy of diseases, not taking into account that multiple different diseases share a common molecular cause, is so rigid as to preclude description of the complex interrelationships that link diseases to each other and to the vast array of causative factors.

This confirms the hypothesis that combining different variables from many sources of administrative data (hospitalization, drugs, ambulatory care record) is the best strategy in identifying cases of specific disease basing on their association with specific healthcare process events (10).

For instance, while HF may contribute to the need of hospitalization caused by other health problems, the HF diagnosis may not be entered on the discharge record. Considering diagnosis codes from hospital discharge as a sole criterion to identify cases of HF, leads to underestimate the prevalence of HF (26).

In integrating different sources of data, it is mandatory to consider that the inclusion of drug prescription codes in the rule-based identification system could increase the HF false-positive rate, because among clinically similar or related diseases, many medications overlap (27). Otherwise it is also true that the pharmacological treatment of patients with chronic HF seems to be acceptably adherent to the recommendations (28).

In this study, we specify, apply, and validate the most suitable algorithm to identify those patients with or at risk to develop HF and we set this as the first step for improving the Precision Public Health approach.

In particular, we tested two different combinations of diagnostic and therapeutic codes used for identifying individuals with HF from the administrative data of the general population.

The broader combination was based on the inclusive disjunction (OR connective) among drug, diagnoses, and exemption codes, while in the narrow combination, the CARPEDIEM algorithm, drug codes were grouped in some possible therapeutic strategies for HF. These subgroups were taken in inclusive disjunction (OR connective) with diagnoses and exemption codes.

The HF prevalence rate obtained by running the broader combination is four times higher than the value estimated by GPs, while the prevalence from the CARPEDIEM algorithm is only slightly higher (2.3%; 2.3–2.4%).

The accuracy and the predictive values for the CARPEDIEM algorithm with respect to the GPs and the clinical standards are highly consistent with those from other studies in terms of validity of HF diagnoses in administrative databases. In particular, the sensitivity of the CARPEDIEM algorithm compared to the clinical standard ranges from 72 to 92% according to the increase of age, with an average value of 88%, which is consistent with those values estimated among the high-quality studies selected by McCormick et al. (29).

The sensitivity of the CARPEDIEM algorithm compared to the GPs standard is a bit lower (76.1%; IC 95%: 75.3–76.2), but it varies from 34.1 to 81.3% according to the complexity of the clinical status, expressed as the number of comorbidities belonging to different categories of disease (disability, psychiatric disorders, chronic renal insufficiency, transplantation, neoplasms, cardiovascular diseases, chronic obstructive pulmonary disease, gastroenteropathy, neuropathy, autoimmune, endocrine and metabolic diseases, diabetes, rare diseases, pregnancy, and other health conditions that cannot be classified as chronic disease or pregnancy).

Evidence of the relationship among clinical status severity and the CARPEDIEM algorithm validity also comes from the extent of agreement with the clinical scores assessed within the validation cohort (NHYA and HF severity), which increases from less severe conditions to the worst (Cohen’s Kappa from 0.44 to 0.70).

These results suggested that severe cases of HF may be recorded more often in administrative databases than mild severity ones.

PPV is moderate (65.1%) when the final rules have been compared to the clinical standard and very low when compared to the GPs standard (13.6%; IC 95%: 13.4–13.8). On the contrary, the best NPV refers to the GPs standard (99 vs 65%).

This means that the rule-based system is more consistent with a clinical evaluation performed in a specialist setting while it has a more intense ability in ruling out the false negative HF cases within a more general clinical setting (General Medicine). This could be critical to add in the early detection of patients either at risk for HF or in early stages of HF and enable their referral to specialists who can further optimize their cardiovascular care.

This could reflect that PPV and NPV are both dependent on the prevalence of the condition in the reference population. As a consequence, PPV will be higher in a specialist setting where the prevalence is higher, while the NPV will be better in a general setting were the prevalence is lower.

Further support about the rule-based system validity respect to the general setting was provided by the LR values estimated through the comparison with GPs standard. In particular, LR (+) is greater than 10 in both geographic areas, and this means that an individual classified for HF by the CARPEDIEM algorithm is more than 10 times likely to have had a HF diagnosed by GPs with respect to a not classified individual.

In summary, CARPEDIEM establishes the potential that an algorithmic approach, based on integrating the administrative data with other public data sources, can present the opportunity to develop low-cost and high-value population-based evaluations for improving public health and impacting public health policies. This suggests that the CARPEDIEM approach can become a significant contributor to the ultimate goal of Personalized Medicine by providing a reliable and reproducible component of disease diagnosis. Moreover, the core of the algorithm can be easily customized to process data sources other than those we used in our study, provided that they contain standardized clinical information. This represents a potential for evolution and improvement of predictive levels of the CARPEDIEM algorithm not only in the heart diseases.

Further research will be carried out by using alternative methodologies such as the Social Network Analysis, a specific tool for enhancing hypothesis generation and with a particular ability in identifying complex and hidden relationships among the data, exposing them for further appropriate statistical verification and supporting complex investigation about undiagnosed HF and gender differences in the clinical presentation of HF.

Thus, the ability to examine readily accessible data, such as the administrative data contained in the National Health Service database, can be a critical component of a broad population-focused surveillance program in HF and also lead to its potential application in other chronic and complex disorders, and CARPEDIEM provides an initial prototype for such development.

This study was carried out in accordance with Italian Legislative Decree 196/2003 on privacy, as stated by the ethical committee of “Area Vasta Nord Ovest PISA” (approval document no. 539/2015, on February 20, 2015). Written informed consent was given by the patients for their information to be stored in the databases used in this study. Written consent follows the criteria established in the Arts. 13 and 22 of the Italian Legislative Decree 196/2003 on privacy. GPs collect written consent only once, at the time of taking in charge of each patient. GPs usually store written consents in their digital archives. Written consent is mandatory whenever patient and NHS interact. All the written consents also regards statistical processing of the data for research purposes. In this case, data must be anonymous, and the patient must not be identifiable in any way. The linkage procedure among the databases was performed using encrypted unique identifier codes created for each patient by the Local Administrative Authority, which is the health provider legally authorized for the processing of sensitive data.

MF provided substantial contributions to the epidemiological conception and design of the work, assisted with acquisition of data and algorithm identification, completed all analysis, led interpretation of data, drafted the work, and revised it critically. SP provided relevant technical contributions to the design of the work, assisted with acquisition of data and algorithm identification, led interpretation of data, and assisted in drafting and revising the work. CP and ME provided substantial contribution to the clinical conception of the work and algorithm identification, acquired and provided clinical data, and participated to the final approval of the version to be published and ensured that questions related to the clinical accuracy or integrity of any part of the work are appropriately investigated and resolved. SM assisted in designing the work, drafting the paper, and revising it critically. SM also provided the final approval of the version to be published and ensured that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.