To our knowledge, this is the first population-based study that applied statistical models and machine learning algorithms to identify a combination of predictive variables for the early diagnosis of XLH using a pediatric general practitioners database.

The primary endpoint of the study, that is the annual prevalence of XLH cases in the PediaNet database, has been estimated to be 1.78 cases per 100,000 registered patients. XLH epidemiologic data reported in literature are poor and sometimes contradictory depending on the location of the epidemiologic evaluation; in Europe according to a Danish, a Norwegian and a French investigations, XLH prevalence can range from 1.07 to 4.8 cases per 100,000 inhabitants [4, 5, 28]; these data are in line with prevalence PediaNet findings.

The three models used to predict XLH diagnosis achieved a fair diagnostic accuracy, with random forest yielding the highest AUC values in both the one-year and the two-years prior to ID analyses. Overall, considering the one-year and the two-years prior to ID analyses together, six diagnosis predictors were selected by the three predictive algorithms. The selected predictors of XLH diagnosis were factors associated with the number of vitamin D prescriptions, mainly consisting of vitamin D and analogs, the number of recorded diagnoses for acute respiratory infections, the number of antihistamines for systemic use prescriptions, and the number of allergology visit prescriptions; other selected predictors of XLH diagnosis were specific exams or diagnostic tests (i.e., the number of X-rays of the lower limbs and pelvis, and the number of examinations for the body composition assessment). Regarding the number of vitamin D prescriptions, mainly consisting of vitamin D and analogs, it is reasonable that patients affected by hypophosphatemic hereditary disorders are usually treated with conventional therapy including active vitamin D and analogs to sustain inappropriately low vitamin D (1,25OH)2D3 levels. To date, the evidence concerning the potential association between vitamin D deficiency and acute respiratory infections is controversial. A large body of evidence suggests that vitamin D deficiency may be associated with increased autoimmunity and increased susceptibility to acute respiratory infections, including bronchiolitis and pneumonia, especially among children with vitamin D deficiency rickets [29,30,31,32,33]. In our study acute respiratory infections included acute nasopharyngitis, acute bronchitis and bronchiolitis, acute upper respiratory infections of multiple or unspecified sites, acute pharyngitis, acute bronchospasm, acute laryngitis and tracheitis, acute sinusitis and acute tonsillitis. In particular, a meta-analysis of eight observational studies exploring the association between serum vitamin D levels and community-acquired pneumonia (CAP) found that patients with vitamin D deficiency are at increased risk of CAP as compared to patients with normal vitamin D serum levels [34]. According to this meta-analysis, vitamin D is involved in the onset of CAP due to the binding of the active form of vitamin D (1,25OH)2D3, which is deficient in patients with XLH, to its receptor (VDR), which stimulates the expression of antibacterial peptides resisting bacterial and viral infections; moreover, it has been demonstrated that in in vivo experimental models, vitamin D deficiency causes the reduction of the level of VDR, thus damaging the epithelia of the respiratory tract mucous membranes with uncontrolled inflammatory reactions [34]. However, a clinical trial evaluating the association between vitamin D supplements and tuberculosis prevention did not find any benefit of vitamin D supplementation for the prevention of acute respiratory infection in children harboring severe vitamin D deficiency [35]. Therefore, patients included in our study may have an increased susceptibility to respiratory infections irrespective of vitamin D levels, as observed in other rare bone diseases such as the type 1 osteogenesis imperfecta [36,37,38].

Findings from our study showed that the number of allergology visit prescriptions as well as the number of antihistamines for systemic use prescriptions are two potential predictors of XLH diagnosis. This may be explained by the increasing, but still controversial, evidence demonstrating that low vitamin D levels are associated with an increased incidence of allergy, such as food allergy [39, 40] and allergic asthma [41, 42]. A recently published systematic review and meta-analysis of observational studies comparing vitamin D levels between children with food allergy and healthy controls reported that the first were found to have a 68% increased probability to experiment a food allergy episode, particularly in their second year of life, as well as a 56% increased probability of developing food sensitization [39].

It is indeed reported that vitamin D plays a crucial role in the modulation of the innate immune response, mainly by increasing the production of antimicrobial peptides (e.g., cathelicidin and β-defensin), as well as the adaptive immune response [43]. Since immune cells such as B cells, T cells, macrophages and dendritic cells express vitamin D receptors on their surfaces and they are capable of synthesizing active vitamin D, they can rapidly increase local levels of vitamin D, thus modulating adaptive immune responses [44]. Vitamin D inhibits B-cell proliferation, blocks B-cell differentiation and immunoglobulin secretion, suppresses T-cell proliferation and affects T-cell maturation, decreasing inflammatory cytokines such as interleukin (IL)-17 and IL-21 and increasing anti-inflammatory ones, such as IL-10 [43]. Moreover, vitamin D reduces mast cell activation, eosinophil count and infiltration of lung disease. These mechanisms, by reducing inflammation and modulating immune cells that are involved in the pathogenesis of allergic asthma, could explain the increased incidence of this pathology in rickets affected patients with inappropriately low active vitamin D levels.

The typical presentation of patients with XLH includes deformities of the lower limbs, bone pain, stunted growth, physical dysfunction and an increased prevalence of overweight/obesity in comparison to general population [10, 25]. Consequently, as expected, the other two XLH diagnosis predictors were related to XLH diagnostic procedures, i.e., the number of X-rays of the lower limbs and pelvis and the number of examinations for the body composition assessment.

One of the major strengths of this study is that data were collected from PediaNet, which is a large and validated research database managed by a national network of FPs, containing information on diagnoses, prescriptions, and outcomes for more than 430,000 children aged from 0 to 14 years of age during the study period. The size of the database and the long follow-up time are particularly relevant for rare diseases research, considering that the number of affected patients is very small. An additional strength is that cases were validated through manual search for FPs diagnosis in the clinical charts.

However, this study also has some limitations. First, a limited number of patients affected by XLH was found in PediaNet, and some of them did not have a sufficiently long database history, thus reducing the statistical power of the analyses. Second, many variables such as weight, height, head circumference, specialist visit (e.g., dentistry data) code and related free text report, but mostly clinical laboratory DMR alphanumeric code and analyte exam results, e.g., phosphate or vitamin D levels, were missing and information on hospitalizations, immunizations and privately purchased medications might be underreported, thus globally affecting the prediction analysis. Third, as the date of the first XLH diagnosis registered in PediaNet may not exactly coincide with the actual onset of the disease, the index date used to set the timeframes for diagnostic prediction might have been misclassified. Fourth, the exact indication of use of prescribed drugs, including vitamin D, was not available in PediaNet DB. Fifth, when available, anthropometric data may be affected by a measurement bias. Lastly, the predictor-diagnosis relationships assessed using data driven approaches, such as machine learning algorithms, do not always imply a causal relationship.