Study design

A population-based, nested case-control study was performed using Clinical Practice Research Datalink (CPRD) GOLD (protocol 17_202), which uses data from practices using Vision® software. CPRD is a prospective longitudinal database that collects patient-level data from participating UK general practices [24]. Children under 18 years account for approximately 20% of the total cohort [25]. CPRD GOLD comprises of a cohort that is broadly similar to the UK general population in term of age, sex, and ethnicity composition [26, 27]. Ethical approval for use of CPRD data was approved by the Medicines and Healthcare Products Regulatory Agency Independent Scientific Advisory Committee in November 2017.

Study population

The study population consisted of CYP aged 0–21 years (inclusive) registered with a primary care practice that contributed data between January 1, 2000, and August 31, 2018, who had at least 30 days current registration prior to the index date, to ensure incident cases were captured. Cases were CYP with an initial medical code for severe stages of CKD recorded during the study period. It was hypothesised that at this level of kidney function, children could be symptomatic from a decline in kidney function to the extent of requiring medical review. Codes were included if they described CKD stages four or five (estimated glomerular filtration rate <30ml/min/1.73m²); established kidney failure; use, or complications of dialysis; kidney transplantation (S1 Table). The index date was the date of the first case-defining code in the medical record. Up to 20 controls who have never had a severe CKD code recorded in their medical record were selected for each case, matching on sex, age (+/-3 years), and deciles of practice kidney function testing rate, a proxy measure for access to investigations for kidney disease and referral at practice level. Each control inherited the index date of the case. Data for this study were extracted from the CPRD GOLD dataset on October 9, 2018, and accessed for research purposes until November 2021. The research team did not have access to information which could identify individual participants during or after data collection.

Exposure definitions

The presence of symptoms suggestive of CKD (yes/no) and consultation frequency in the 24 months prior to the index date were examined for cases and controls. A systematic approach to identifying and defining symptom variables was undertaken [28] (S1 File). This timeframe was chosen as it was hypothesised that CKD may have a relatively long latency period during which symptoms develop which could represent opportunities for earlier detection. Examination of clinical features during a shorter timeframe (6 months) was also performed to see whether this changed reported symptoms and effect estimates; we speculated that if a symptom variable were causally associated with CKD, any associations observed would strengthen as time to index date decreased.

In line with other work, a consultation was defined as any face-to-face medical consultation which occurred at the practice [29]. Consultations for immunisations were excluded. Sociodemographic reporting was determined using 2015 English Indices of Multiple Deprivation (IMD) scores, which were based on the practice postcode for patients whose practices consented to linkage. Derived from Census data, the IMD score is an ecological composite measure of relative deprivation based on multiple measures (e.g., housing, employment, crime) [30].

Analysis

Associations between symptoms and consultation frequency with severe CKD were examined using conditional logistic regression to estimate odds ratios (ORs), 95% confidence intervals (CI) and p-values. Use of a matched design does not completely control for confounding [31] and therefore matching variables were included, with age as a categorical variable (≤10; 11–17; 18–21 years). As sex and testing decile were exactly matched, these variables were omitted from the model. To assess the clinical utility of symptoms and consultation frequency, sensitivity, specificity, and positive likelihood ratios (LRs, sensitivity/1-specificity) were calculated [32, 33]. The positive predictive value (PPV) was estimated using Bayes’ theorem [34]; prevalence estimates of severe stages of CKD in children aged <18 years were obtained from the UK Renal Registry [35] and applied to the study population.

Prediction modelling using 24-month timeframe

Using age-adjusted conditional logistic regression, eight individual symptom variables were predictive of severe CKD in the 24 months before the index date: change in urine colour, cough, flank pain, oedema, urinary tract infection (UTI), vomiting and growth concerns (including weight loss), as well as generic ‘unwell’ codes (Table 2). The most sensitive symptoms were cough (23.4%), vomiting (11.9%) and UTI (9.2%). Cough was the least specific (82.6%), followed by vomiting (95.6%). In this timeframe, the highest positive likelihood ratios and predictive values were seen for flank pain (LR 16.8, 95% CI 1.1, 268.0; PPV 1.7 per 1000) and growth concerns (LR 7.1, 95% CI 3.6, 14.2; PPV 0.7 per 1000).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Frequency of symptoms predictive of severe CKD between cases and controls in the 24 and 6 months prior to the index date.

https://doi.org/10.1371/journal.pone.0314084.t002

Prediction modelling using 6-month timeframe

Within 6 months of the index date, codes describing a generally unwell child, growth failure, oedema, change in urine colour, UTI and vomiting remained positively associated with the outcome, while headache, dyspnoea/shortness of breath, thirst and visual problems were individual symptoms also predictive of severe CKD (Table 2). The most sensitive codes were vomiting (5.5%), headache (3.8%) and UTI (3.1%). All symptoms demonstrated high specificity (>98%). Growth concerns and change in urine colour had the highest LRs (17.7, 95% CI 5.7, 54.5 and 17.7, 95% CI 2.5, 125.0, respectively) and PPVs (1.8 per 1000 children e.g., of 1,000 children coded, 2 would be diagnosed with severe CKD in 6 months), followed by oedema (LR 9.8, 95% CI 3.3, 29.1; PPV 1.0 per 1000). Colour change was more frequently reported by cases but was not predictive of severe CKD in regression models; codes for oliguria and flank pain were seen among cases only.

To further investigate symptom reporting, associations with severe CKD were examined in the 0–6 and >6–24 months prior to diagnosis (S1 Fig). In the >6–24 months prior to diagnosis, growth concerns, vomiting, UTI, and oedema remained predictive of case status.

Sensitivity analyses

When children with evidence of known kidney disease were excluded, oedema, growth, vomiting, and vision independently remained strongly associated with case status in both 24- and 6-month timeframes (S2 Table). Similar findings were observed when the analysis was restricted to under 18-year-olds, although effect estimates were attenuated (S3 Table).

Consultation frequency

Higher odds of being a case were noted with increasing consultation frequency (Table 3). A similar pattern was seen when children with known kidney disease or ≥18-year-olds were excluded.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Associations between consultation frequency and severe CKD within the 24 and 6 months before the index date.

https://doi.org/10.1371/journal.pone.0314084.t003

Risk score development

The final multivariable model included codes for growth concerns, being ‘generally unwell’, oedema, vomiting, UTI, urine colour and consultation frequency as an ordinal variable. As the symptom ‘generally unwell’ was considered vague and difficult to define, it was excluded from the final model. There was evidence of effect modification (interaction) between age and growth reporting, with higher odds of severe CKD seen in younger children reporting growth concerns compared to older children. Sex was noted to modify the effect of oedema reporting on severe CKD, with higher odds of severe CKD among females with oedema present. Interaction terms for both were therefore included in the final model. Table 4 highlights findings from the prediction model, before and after internal validation, with mutual adjustment for included variables. A plot of the observed versus the expected probabilities demonstrated acceptable calibration, with a calibration slope of 0.91 (95% CI 0.00, 1.11) and calibration-in-the-large of -0.01 (95% CI -0.12, 0.12, see S2 Fig). The β-coefficients of the bootstrapped model were used assign a risk score and calculate predicted risk (equation shown in legend of Table 4). A risk threshold of 10% or more offered high positive likelihood ratios (S4 Table; LR at 10% predicted risk: 2.8, 95% CI 2.0, 3.7) and high specificity (94.7%, 95% CI 94.1, 95.3) although positive predictive values remained low due to the low background disease prevalence.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Multivariable prediction model for severe CKD based on presenting symptomatology to primary care in previous 6 months (n = 5,416, 290 cases).

https://doi.org/10.1371/journal.pone.0314084.t004

Comparison with existing literature

In our review of the literature, there is only one published risk score supporting CKD identification in children [23]. This identified similar symptoms, particularly growth concerns, oedema, and UTI, were predictive of disease compared to controls, despite differences in setting and study population. Our results are also consistent with observational studies of children referred to pediatric nephrology care, with several noting prolonged symptomatic periods before presentation [20, 39, 40]. From these findings however, it was not clear whether patients were previously seen by other healthcare providers and where interventions could be targeted. We confirm that symptoms suggestive of CKD are not only present before diagnosis but may be detectable for some time beforehand, highlighting potential opportunities for earlier diagnosis and intervention.

Strengths and limitations

This study was inclusive of all cases with severe stages of CKD identified in CPRD GOLD, a representative UK primary care database, which ensures results are generalisable to all children with CKD irrespective of underlying disorder as well as being applicable to other high-income countries with primary healthcare services. Symptom code lists were compiled using established methodology as has been employed in other studies [32, 41–43]. Furthermore, exposure status was assigned based on symptoms documented prospectively prior to coding of severe CKD which helps minimise information bias.

There are however limitations. As determination of symptom exposure was reliant on practitioners recording symptoms in the electronic patient record, under-coding may have led to underestimation of effect estimates, assuming that any misclassification is non-differential to future case status. Given similar predictive symptoms were noted when cases with pre-existing kidney codes were excluded, we are assured differential misclassification of symptom exposure did not occur; this finding also makes it less likely that reverse causality (children with advanced CKD are more likely to be coded for CKD-specific symptoms) has occurred. Due to the number of controls per case required, we were unable to match on more granular variables such as practice, raising the possibility of residual confounding.

Implications for research and/or practice

While there is increasing focus on early CKD recognition to prevent or delay kidney disease progression, it is not clear what the best strategy is to maximise detection whilst minimising harm, costs, and inappropriate use of resources. In adults, evidence suggests that a ‘case finding’ approach, screening patients known to be at higher risk of CKD due to coexisting conditions, is a more efficient option to population screening [18, 19]. For children, there are sparse data regarding who to test for CKD, despite the potential health and economic gains to be made across their lifecourse. In 2007, the American Academy of Pediatrics recommended discontinuation of routine urinalysis screening in children, a practice which Sekhar and colleagues estimated to be cost-ineffective [44]. Currently UK guidance from the National Institute for Health and Care Excellence advise consideration of blood and urine testing for children at risk of CKD development including those with diabetes, hypertension or known structural kidney disease, although acknowledge the limited evidence base [45]. In their 2021 evidence update, the national body stated research identifying prognostic factors for CKD development in children was a high national priority [46]. This study has demonstrated that pragmatic markers such as symptoms and consultation frequency can be used to identify children who go on to develop severe CKD. Incorporating a prediction model into clinical systems may support decision-making regarding testing and referral, which has helped shorten time to diagnosis for other diagnostically challenging and rare conditions such as childhood brain tumours [47]. Considering the low PPV for individual symptoms, it would be inappropriate for primary care clinicians to refer all children with these symptoms to specialist care for investigation, however the high specificity of symptoms, both individually and in combination is reassuring. For children with high predicted risk, urinalysis, or blood tests may help to confirm or refute a diagnosis of CKD. While this may increase costs associated with testing, identifying CKD in earlier stages may result in a more cost-effective pathway for future management and monitoring, thus representing appropriate use of healthcare resources [46]. Further work is needed to determine the economic impact on incorporating such a clinical tool into practice. In addition, work is needed to determine whether the discriminatory power of this model can be enhanced with inclusion of other variables readily accessible to primary healthcare teams before external validation [48]. Future mixed-methods research should explore the acceptability of a clinical prediction score as well as appropriate risk thresholds for invasive and non-invasive testing in children, from the perspectives of the child, family, and professionals.