Development of a primary care screening algorithm for the early detection of patients at risk of primary antibody deficiency

The hallmark clinical features of PAD are recurrent respiratory tract infections (RTI), including otitis media, sinusitis and pneumonia [12, 21, 30]. Gastro-intestinal complaints, such as chronic diarrhea and parasitic infections, also occur frequently in certain PAD [12, 31‐34]. In addition, it is estimated that around 30% of PAD patients develop auto-immune disorders, e.g. idiopathic thrombocytopenic purpura, auto-immune hemolytic anemia or rheumatoid arthritis [35‐37]. Finally, certain PAD can also present with lymphoproliferative symptoms and are associated with an increased risk of non-Hodgkin lymphoma and gastric cancer [38‐41]. All diagnostic ICPC codes related to these presenting symptoms of PAD were considered as candidate components of the algorithm.

Since PAD patients often present with recurrent infections that require antibiotic treatment, we also considered the type and number of antibiotic prescriptions for inclusion in the algorithm. We considered antibiotics registered in the Netherlands for the treatment of respiratory tract or parasitic intestinal infections [42]. As antibiotics for upper RTIs are frequently prescribed in the early years of childhood, we only took prescriptions from the age of 6 years into account [43].

As PAD, by definition, can be characterized by a reduced level of one or more (subtypes of) immunoglobulins, we considered these laboratory results for inclusion in the algorithm. Since immunoglobulin diagnostics may not be requested commonly in primary care, we also considered calculated globulin. Calculated globulin is a measure based on total protein and albumin levels, which can indicate hypogammaglobulinemia [44, 45].

A previous study reported that a higher number of consultations at the GP, as well as a higher number of requests for lung function tests and blood tests for infection, were statistically significantly associated with an increased odds of CVID diagnosis [46]. Therefore, we also considered these factors for inclusion in the algorithm.

In contrast to PAD, secondary antibody deficiencies are the consequence of an underlying disease, such as hematological malignancy. Diagnoses known to be causes of secondary antibody deficiencies (e.g. leukemia, multiple myeloma, HIV) were added as exclusion criteria [47‐49]. In the case that a diagnosis could be both a cause of secondary antibody deficiency as well a complication of PAD (e.g. non-Hodgkin lymphoma), it was labelled as ‘‘ambiguous diagnosis’’. For these diagnoses, the EHRs were screened up to the moment of the ambiguous diagnosis. This approach allows detection of patients for whom the ambiguous diagnosis is the consequence of an underlying PAD, whilst still excluding patients with a secondary antibody deficiency. Immunosuppressant medication was not added as an exclusion criterion, as these may also be prescribed for auto-immune symptoms that are caused by an underlying PAD. Table 1 shows an overview of the candidate components of the algorithm.

To gain insight in the prevalence of the candidate algorithm components among PAD patients and control groups, we used two data sources. First, we collected the primary care EHR data of 30 patients from different age groups with a variety of PAD diagnoses from an academic hospital in the Netherlands (the University Medical Center Utrecht). Patients signed informed consent and research agreements were signed by their GPs. Second, we collected primary care EHR data from the pseudonymized database of the Julius General Practitioner Network (JHN), a research collaboration of GPs in the region of Utrecht, Netherlands [50]. Owing to privacy regulations from the JHN database, we only had access to the metadata (e.g. means, medians and proportions) concerning the candidate components of the algorithm for several patient groups of interest, defined based on registered diagnostic ICPC codes. The groups of interest included PAD patients as well as several control groups. Because no specific ICPC code for PAD exists, we selected the first group based on the code for immunodeficiency (T99.01). This group approximates PAD patients, as causes of secondary immunodeficiencies were excluded based on the exclusion criteria, and because PAD represents the majority of the remaining PIDs. Therefore, this group of ‘primary care immunodeficiency’ patients was considered to be of interest, in addition to the group of confirmed PAD-patients from the academic hospital.

The selected control groups were ‘patients from the general GP population’, ‘patients with upper RTIs’, ‘patients with chronic obstructive pulmonary disease (COPD) or asthma’, ‘patients with inflammatory bowel disease (IBD)’ and ‘patients with a malignancy’. These groups were selected as they were expected to have overlapping presenting symptoms with PAD patients, but should be distinguished from potential PAD patients by the algorithm in order to prevent false-positives. For all control groups we applied the exclusion criteria, and excluded patients that had an ICPC code for immunodeficiency (T99.01). The meta-data of all available patients in the JHN-database that met the criteria for the control groups were utilized.

From both the academic hospital and the primary care database we selected patients aged 12–70 years, as PAD most commonly presents in the second to fourth decade of life [9‐12]. In addition, as recurrent infections and antibiotic prescriptions are common in younger children, other cut-off values are to be expected for this age group. Lastly, our algorithm was developed in line with the study design of a subsequent external validation study, where high-risk patients will be invited for laboratory analysis of immunoglobulins. Considering the medical ethical guidelines, this validation study is only feasible in patients aged ≥ 12 years.

A distinct ‘censoring date’ was selected for different patient groups. All available EHR data before the censoring date were extracted. The censoring date indicates the date up to which time the EHR data were extracted. For most patients, this was the date of data-extraction, November 18, 2021. For the PAD and primary care immunodeficiency patients, we extracted the EHR data both pre- and post-diagnosis. The censoring date could therefore be the date of PAD diagnosis (pre-diagnosis), or the date of data-extraction (post-diagnosis). When deciding which components were to be maintained in the algorithm, pre-diagnosis data were used because our aim is to identify patients before diagnosis. For patients with an ambiguous diagnosis, the date of this specific diagnosis was defined as their censoring date.

To determine the validity of the diagnostic ICPC codes, we compared the presenting symptoms registered as diagnostic ICPC codes in the primary care EHR of the PAD patients with the presenting symptoms registered as free text in the secondary care EHR of the academic hospital and determined their concordance.

The decision to include a candidate component in the algorithm and the corresponding weight was based on the results of analyses of the corresponding EHR data in combination with clinical rationale. To estimate the ability of the presence of individual diagnostic ICPC codes to differentiate PAD patients from the control group, we calculated the Youden’s index. Youden’s index is a summary measure of diagnostic quality based on sensitivity and specificity, where a higher index indicates that the component has a better discriminatory ability [51]. The sensitivity of the presence of a component was calculated for the PAD patients and primary care immunodeficiency patients, and the mean of the sensitivities the was used to determine Youden’s index. Specificity was calculated from the general GP population control group. The weight of the diagnostic ICPC codes was based on the discriminative value as expressed by Youden’s index in combination with clinical rationale.

The diagnostic ICPC codes included in the algorithm were grouped according to the following categories: ‘respiratory tract infections’, ‘gastro-intestinal complaints’, ‘other infections’, ‘auto-immune symptoms’ and ‘malignancies, lymphoproliferative and other symptoms’. For each of these groups, we determined whether the entire EHR before the censoring date should be screened for these diagnostic codes, or only the 10 years before the censoring date. This decision was based on clinical rationale and on the discriminative value as expressed by Youden’s index.

For the ATC codes, we calculated the mean number of antibiotic prescriptions per year in the 10 years before the censoring date. The decision on the inclusion and weight of individual antibiotics was based on both clinical rationale and the difference in the mean number of prescriptions between the control groups and immunodeficiency patients.

For immunoglobulin levels and calculated globulin, we determined the mean laboratory values in the EHR in the 10 years before the censoring date. Lastly, regarding the number of consultations at the GP and the number of requests for lung function tests/blood tests for infection, we calculated the mean and median number of visits and requests in the year before the censoring date for both the PAD/primary care immunodeficiency groups and the control groups. To determine the optimal cut-off point for the number of lung function tests, blood tests for infection and number of visits to the GP, we calculated the Youden’s index for several cut-off points spread around the median values in PAD patients.

Ethical approval for this study was received from the Medical Research Ethics Committee Utrecht, protocol number 19–748. Statistical analysis was done with the base package in R version 4.0.2.

Of the 30 included PAD patients from the academic hospital, 12 were diagnosed with CVID, 8 with IgG subclass deficiency, 4 with unclassified antibody deficiency, 3 with IgA and IgG subclass deficiency, 2 with selective IgA deficiency and 1 patient was diagnosed with SpAD. The mean age at the time of data extraction was 29.8 years, with a large variation (standard deviation of 14). The PAD patients thus represented a wide range of diagnoses and ages. The mean delay from symptom onset to diagnosis of these patients was 12.4 years [SD 12.2, median 9.5 (3.0–19.5)]. The patient characteristics of the PAD patients from the academic hospital, of the primary care immunodeficiency patients and the control groups (all aged 12–70 years) are shown in Table 2.

Table 3 shows the concordance between the presenting symptoms registered as diagnostic ICPC codes in the primary care EHRs and the symptoms registered in free text in the secondary care EHRs. Concordance was assessed by calculating the percentage of patients for whom the presence or non-presence of a symptom was equal between the diagnostic codes in the primary care EHR and the free text in the secondary care EHR. For almost all presenting symptoms, apart from gastro-intestinal complaints, concordance between the diagnostic primary care ICPC codes and the free text in secondary care EHRs was high.

The selection of diagnostic codes and their attributed weight was based on the Youden’s index (Additional file 1: Table S1) in combination with evidence from literature and clinical expertise of PAD specialists in the University Medical Center Utrecht. The diagnostic codes included in the final algorithm are shown in Table 5.

Figure 2A shows the mean number of antibiotic prescriptions per year in the 10 years before the censoring date for the different patient groups. PAD patients and primary care immunodeficiency patients had a higher mean number of antibiotic prescriptions than all control groups. In the pre-diagnosis groups this effect appeared to be most pronounced from 4 years before the diagnosis. Figure 2B shows the mean total number of antibiotic prescriptions in the 4 years before the censoring date, where the immunodeficiency and PAD groups had a higher mean compared to the control groups (5.14 in PAD vs. 0.48 in the general population). It was therefore decided to include antibiotic prescriptions from the past 4 years in the algorithm, see Table 5. The mean number of antibiotic prescriptions per individual antibiotic is shown in Additional file 2: Table S2. The weight of the individual antibiotics was based on the difference in means between immunodeficiency and control groups in combination with clinical rationale. For example, if the difference in means was < 0.05, we generally attributed a weight of 1, rather than a weight of 2, for each antibiotic prescription.

Table 4 shows the mean laboratory values of the immunoglobulins and calculated globulin in the EHR in the 10 years before the censoring date. Immunoglobulins were only requested for a very small proportion of patients from the control groups. Calculated globulin was requested more often than immunoglobulins in these control groups, although still for a very small proportion. For example, calculated globulin was requested only for 0.38% of the general GP population. For the PAD and primary care immunodeficiency patients, immunoglobulins and calculated globulin were not requested pre-diagnosis. Although these parameters are thus not likely to apply to many patients, a reduced level of immunoglobulins or calculated globulin was still added to the algorithm based on the strong clinical relevance. Although it could be expected that a PAD diagnosis is quickly made after immunoglobulin testing, patients may be missed by the GP, thus making this a relevant component to add to the screening algorithm.

Based on previous research by Ilkjær et al. [46] we also considered the number of requests of leukocytes, C-reactive protein and lung function tests by the GP. These however did not appear to have a discriminatory value when comparing the general population with immunodeficiency patients, see Additional file 3: Fig. S1. These components were thus not added to the algorithm.

Figure 3 shows the median number of visits to the GP in the year before the censoring date. The immunodeficiency groups showed a higher median number of visits to the GP compared to the control groups. For example, PAD patients pre-diagnosis had a median of 6 visits per year, compared to 2 visits per year in the general population. These data were not available for the PAD patients from the academic hospital. Based on Youden’s index the cut-off value for the algorithm was set to ≥ 6 visits to the GP in a year.

The components and corresponding weights of the final PAD screening algorithm are shown in Table 5. The categories of the algorithm include ‘Antibiotics’, ‘Respiratory tract infections’, ‘Gastro-intestinal complaints’, ‘Other infections’, ‘Auto-immune symptoms’, ‘Malignancies, lymphoproliferative and other symptoms’, ‘Laboratory values’ and ‘Visits to general practitioner’.