Predicting microbiologically defined infection in febrile neutropenic episodes in children: global individual participant data multivariable meta-analysis.

BACKGROUND: Risk-stratified management of fever with neutropenia (FN), allows intensive management of high-risk cases and early discharge of low-risk cases. No single, internationally validated, prediction model of the risk of adverse outcomes exists for children and young people. An individual patient data (IPD) meta-analysis was undertaken to devise one.
METHODS: The 'Predicting Infectious Complications in Children with Cancer' (PICNICC) collaboration was formed by parent representatives, international clinical and methodological experts. Univariable and multivariable analyses, using random effects logistic regression, were undertaken to derive and internally validate a risk-prediction model for outcomes of episodes of FN based on clinical and laboratory data at presentation.
RESULTS: Data came from 22 different study groups from 15 countries, of 5127 episodes of FN in 3504 patients. There were 1070 episodes in 616 patients from seven studies available for multivariable analysis. Univariable analyses showed associations with microbiologically defined infection (MDI) in many items, including higher temperature, lower white cell counts and acute myeloid leukaemia, but not age. Patients with osteosarcoma/Ewings sarcoma and those with more severe mucositis were associated with a decreased risk of MDI. The predictive model included: malignancy type, temperature, clinically 'severely unwell', haemoglobin, white cell count and absolute monocyte count. It showed moderate discrimination (AUROC 0.723, 95% confidence interval 0.711-0.759) and good calibration (calibration slope 0.95). The model was robust to bootstrap and cross-validation sensitivity analyses.
CONCLUSIONS: This new prediction model for risk of MDI appears accurate. It requires prospective studies assessing implementation to assist clinicians and parents/patients in individualised decision making.

Children developing cancer in North America and Europe have an 80% chance of cure (Pritchard-Jones ; Ward ). The cost of this cure, in terms of intensity of therapy and recurrent admissions with toxic effects, is a considerable burden upon children, young people and their families (Chisholm and Dommett, 2006). One such toxicity, fever with neutropenia (FN), also known as ‘febrile neutropenia' or ‘neutropenic sepsis', is the clinical dilemma of potentially severe infection in an immunocompromised child or a young person. Its management requires balancing competing risks from overtreatment against undertreatment, and personalising care in FN by accurately differentiating those at higher or lower risk of significant infection, to determine who is eligible for alternative treatment approaches (Lehrnbecher ).

Current practice in managing febrile neutropenia in paediatric oncology is variable (Chamberlain ; Boragina and Patel, 2007; Stabell ; Phillips ). Some centres use a risk-stratified reduced intensity approach directed by clinical decision rules, whereas others treat all children with aggressive, in-patient delivered antibiotic therapy.

Systematic reviews (Haeusler , 2013b; Phillips , 2012a, c) show that many clinical decision rules have been proposed to predict which children will have poor outcomes during episodes of febrile neutropenia. These identified several difficulties including small numbers of patients, analysis techniques which lead to optimistic estimates of predictive power, and lack of validation in different geographies and care systems (Phillips ).

Individual patient data (IPD) meta-analysis of data from therapeutic studies has been developed over two decades to improve the precision and reliability of answers to questions of treatment, and provides a means of overcoming many such problems (Stewart and Tierney, 2002; Riley ). More recently, the approach has been promoted for developing and validating risk-prediction models from observational data, to improve the quality of answers to important prognostic questions (Altman and Riley, 2005) and matters of diagnostic accuracy (Leeflang ), and to establish more robust clinical decision rules. Meta-analysis of IPD from multiple studies allows harmonisation of study data sets, standardisation of end points and analysis methods, as well as greater statistical power to identify more complex statistical relationships (Stewart and Tierney, 2002).

To maximise the value of the information already collected by previous studies and cohorts of children with febrile neutropenia, a global collaboration (The ‘Predicting Infectious Complications in Children with Cancer': PICNICC) was established to facilitate an IPD meta-analysis that aimed to robustly develop and externally validate a new prediction model for the risk of microbiologically defined infection (MDI) in febrile neutropenic children (Phillips ).

Methods

A detailed protocol for the PICNICC IPD study was developed, registered and published before commencement of the analysis (Phillips ). The primary aim of the IPD analysis was at the onset of an episode of FN to quantify the risk of adverse clinical outcomes according to clinical variables recorded in children and young people aged 0–24 years old undergoing treatment for malignant disease at the onset of an episode of FN, and to develop a new risk-prediction model.

Collaboration development

The PICNICC collaboration comprises international clinical and methodological experts, parent representatives and health-care researchers. Clinical and research experts were identified from trials identified during preliminary systematic reviews of aggregate data (Phillips , 2012c), in response to oral presentations at paediatric oncology conferences, and web-based invitations. Parents/carers were approached via UK parent support organisations. This process led to 22 different study groups from 15 countries joining the PICNICC collaboration. Ethical approval for the study was granted in the UK and individual members were advised to contact their local ethics board to determine the need for approval (Phillips ).

Data acquisition

Each collaborative group shared, episode-by-episode, de-identified individual level data from their original studies of new-onset episodes of FN as defined by the original studies. Candidate predictors for inclusion in the risk-prediction model were requested for each study. These were patient-level variables measured on admission, and included: patient age at, and date of, each episode of FN; underlying malignancy type; remission status; chemotherapy type and date of last cycle; type of central venous line; in/out-patient status; maximum temperature; clinical observations including global assessment of illness severity and mucositis; blood count parameters; inflammatory biomarkers; and empirical antibiotics used (Phillips ). A minimum requirement of patient age, malignancy type, one marker of clinical status, antibiotics given, white cell and neutrophil counts and one specified outcome was set. Data were mapped into similar reporting units, for example, mmol into g dl−1, and using predefined structures for cardiovascular, respiratory and mucositis outcomes. The outcome used in this analysis was of MDI. We permitted any definition described in the original study and did not insist on the HIS consensus definition (Immunocompromised Host Society, 1990). The majority of studies used microbiological identification of a pathogen from a sterile site. This outcome was chosen after data collection as it was the most frequently and consistently reported clinically relevant outcome, and was highly correlated with the less frequently reported ‘serious medical complication' as suggested by the recently published core outcome set (Haeusler ).

Statistical methods

IPD from the 22 data sets were cleaned, and the availability of variables and the outcome of interest (MDI) were summarised. Studies were excluded if they did not provide the outcome of interest. Univariable analysis was undertaken to examine the unadjusted effect of each candidate predictor, using random effects logistic regression as described below. The multivariable risk-prediction model was developed using logistic regression, hierarchical of episodes within studies. For multivariable analysis, a complete case analysis was undertaken. Studies were removed if they had not collected information on potentially informative variables that had been deemed important a priori and focused on those predictors with the most complete information (⩽15% missing). These were selected from variables that prior systematic reviews suggested would provide the largest complete data set (see Figure 1). Variables were selected for inclusion in the model by a stepwise selection process—examining improvement in fit with each additional variable added to the model using Akaike's Information Criterion (AIC), with a P-value of <0.15 and net reclassification improvement (Pencina ) used in determining inclusion. Different functional forms for continuous variables were assessed by the use of fractional polynomials before inclusion (Royston and Sauerbrei, 2008). Clustering of patients within studies was accounted for by fitting a random-effect on the intercept. The impact of other potential sources of heterogeneity (e.g., in the effects of particular variables across studies) and other degrees of clustering (e.g., through repeated episodes of the same individual) were evaluated with further random effects. Heuristic shrinkage (Steyerberg, 2009) and the bootstrap procedure (Harrell ) using 5000 episodes with replacement were used to adjust for potential overfitting. Performance was internally checked using all-bar-one cross-validation predictive performance (Royston ), where the model is repeatedly applied to the derivation data after a single data set has been removed, to assess for systematic miscalibration driven by outlying data sources. Type of antibiotic therapy was incorporated into the model in a sensitivity analysis, as initial discussions considered this may reflect otherwise occult clinical suspicion of serious infection. The calibration of the model was checked by plotting observed against expected values for population deciles, with discrimination calculated using the area under the receiver–operator curve (AUROC), also known as the C-statistic.

Figure 1

Per study proportion of missing predictors. Grey items are predictors and studies that were included in the multivariable analyses.

Results

Data

IPD information from 5127 episodes of FN in 3504 patients from 22 data sets was provided for analysis (Supplementary Table 1, online).

A wide variety of malignancies were represented in the IPD data sets. Overall there were 52% of patients with acute leukaemias, 27% with solid tumours of childhood, 11% with lymphoma, 7% with brain tumours, 3% with non-malignant conditions treated with chemotherapy (e.g., Langerhans cell histiocytosis) and 2% of children with adult-type solid tumours (e.g., carcinomas of bowel). Only two of the data sets (Genoa, PINE) included patients who had undergone haemopoetic stem cell transplantation (HSCT). The median age of the patients at first episode was 6.8 years, with a range of 50 days to 25 years old, with 14% of patients 16 years or older; 56% of the patients were male.

The IPD collected included a wide range of outcomes and potential predictor variables with marked differences in completeness, interpretation and consistency between data sets. No data set completely reported every item (Figure 1). The most common reason for absent data was that the studies did not collect that variable. The unrecorded data effectively reduced the data set available for multivariable analysis to around 1101 episodes in 742 patients over seven studies. The sparseness of information on biomarkers in comparison with the clinical data meant that these were assessed only in their univariable relationship with MDI and not as part of a multivariable model.

We were unable to obtain data from 30 studies (including 5387 episodes) that had been identified in our previous systematic reviews. Authors of two studies explained that they were unable to participate because the studies were conducted over 20 years ago and data were no longer retrievable; other study groups did not respond or give reasons for not taking part. These unavailable studies did not significantly differ from those included in terms of number of participants (P=0.66), number of episodes (P=0.93), number of events (P=0.67) or geographical region (P=0.25).

Univariable associations with microbiologically defined infection

The results of univariable (unadjusted) analyses showed associations between MDI and many variables (Table 1) including the presence of an untunnelled central line, the clinical appearance of being significantly unwell (as defined by the original studies), of documented cardiovascular compromise (shock), a high temperature, raised serum biomarkers, low white cell counts, neutrophil counts or platelets and a diagnosis of acute myeloid leukaemia or relapsed acute leukaemia. There were very few data on patients who had recently received HSCT (3% of all episodes). A surprising univariable association found that patients with more severe mucositis had a smaller chance of MDI than those with less severe mucositis (Table 1). The estimated associations were generally consistent betweyen studies (e.g., Figure 2). Notably, there was no relationship demonstrated between age and risk of MDI.

Table 1

Univariable associations with MDI

Predictor name	N of episodes	N of studies	N of MDI	OR	95% CI	P-value
Patient demographics
Age (per 10-year increase)	5152	22	1412	1.0	0.98–1.1	0.38
Sex—female	1892	20	549	1	(Reference)
Sex—male	2400		656	0.96	0.84–1.1	0.56
Central venous line type—None	294	6	105	1	(Reference)
Central venous line type—Port	416		102	1.1	0.66–1.7	0.80
Central venous line type—Hickman	357		98	1.4	0.87–2.3	0.17
Central venous line type—Untunnelled	15		6	3.1	0.95–9.5	0.054
Type of malignancy—acute lymphoblastic leukaemia	1805	22	558	1	(Reference)
Type of malignancy—acute myeloid leukaemia	671		251	1.4	1.2–1.8	0.0002
Type of malignancy—relapsed acute leukaemia	192		71	1.4	1.0–2.0	0.06
Type of malignancy—osteosarcoma	204		26	0.38	0.25–0.59	<0.0001
Type of malignancy—Ewing's sarcoma	173		26	0.54	0.32–0.78	0.002
Intensity of chemotherapy—Non-HSCT	2292	12	668	1	(Reference)
Intensity of chemotherapy—HSCT	193		52	1.5	0.96–2.46	0.08
Clinical features
Temperature (per degree Celsius)	2321	18	638	1.9	1.6–2.2	<0.0001
Physician assessment—‘not severely unwell'	2348	17	603	1	(Reference)
Physician assessment—‘severely unwell'	545		219	2.2	1.8–2.7	<0.0001
Clinical assessment—Not shocked	2243	12	699	1	(Reference)
Clinical assessment—Shock	135		73	2.4	1.7–3.4	<0.0001
Mucositis (per grades II–IV, compared with I)	442	11	133	0.89	0.80–1.0	0.050
In-patient at onset	838	18	272	1	(Reference)
Out-patient at onset	1973		482	0.69	0.53–0.91	0.0078
Investigations
Haemoglobin (g dl−1)	2658	16	765	1	0.99–1.1	0.11
Loge (white cell count, 106 l−1)	2845	16	679	0.72	0.66–0.78	<0.0001
Loge (platelets, 109 l−1)	2380	16	621	0.80	0.74–0.87	<0.0001
Loge (absolute monocyte count, cells per mm3)	1469	11	431	0.80	0.75–0.86	<0.0001
Loge (absolute neutrophil count, cells per mm3)	4382	21	1186	0.92	0.90–0.95	<0.0001
Loge (CRP, mg dl−1)	1267	11	366	1.1	0.96–1.2	0.25
Loge (IL-8, pg ml−1)	477	5	94	1.8	1.5–2.3	<0.0001

Abbreviations: CI=confidence interval; HSCT=haemopoetic stem cell transplant; MDI=microbiologically defined infection; OR=odds ratio.

Figure 2

Relation between probability of MDI and natural log of absolute monocyte count by study. Data points indicated by rug plot.

The biomarkers with sufficient data to examine (<90% missing) were C-reactive protein (CRP) and interleukin 8. Each biomarker was assessed as the natural logarithm of its reported value to reduce skewness. CRP was not significantly associated with MDI (see Table 1) but IL-8 had a strongly significant association with MDI (OR=1.81, 95% CI 1.48–2.28, τ2=0.00, I2=0%) see Figure 3.

Figure 3

Unadjusted association between MDI and natural log of IL8 values by study.

Multivariable model to predict microbiologically defined infection

Because of the pattern of missing data, the final analysis data set used to build the multivariable model contained 1101 episodes of FN from 742 patients in seven studies (Figure 1). Overall 24% (236) of those episodes exhibited a MDI. The analysis examined 15 candidate variables (Figure 1 and Supplementary Table 2, online), giving an event-to-predictor ratio of 15.7. There was little between-study heterogeneity in the parameter estimates, and no significant difference in the univariable parameter estimates between-studies included and excluded from the multivariable analysis.

Six simple variables were included in the model to predict MDI: type of malignancy; maximum temperature; clinical evaluation of being severely unwell; haemoglobin; white cell count; and absolute monocyte count (AMC).

Logit(p(MDI))=α+β0+β1 tumour type+β2 temperature+β3 unwell+β4 Hb+β5 loge (white cell count)+β6 loge(absolute monocyte count),

where the subscript i refers to the ith episode, j the jth tumour type, where β1=0 for acute lymphoblastic laekaemia, and k the kth data set and β0k∼N(0, τ2).

The predictions calibrated well against those expected even when calculated using an average intercept (Figure 4, AUROC 0.723, 95% CI 0.711–0.759).

Figure 4

Calibration of PICNICC predictive model. The calibration plot shows the relationship between ideal (dashed line) and observed (dotted line) predicted values. The rug plot along the bottom demonstrates the distribution of the data points, and the triangles show predictions grouped by decile of the population.

There appeared to be little optimism in the predictive estimates as assessed by bootstrap; the mean difference in AUC values between the bootstrapped estimates applied in their bootstrapped data sets, and the estimates applied in the original data set, being −0.0031 (0.7244 vs 0.7213). The heuristic shrinkage estimate for the model (Steyerberg, 2009) was calculated to be 0.97, in keeping with the very small differences produced by bootstrapping. Forcing the antibiotic used into the model (sensitivity analysis) did not affect its predictive performance. Using nonlinear forms did not improve the model fit. Assessment of how the model performed using all-bar-one cross-validation, also showed no significant systematic differences in calibration (see Supplementary Figure 5). Final parameter estimates can be found in Table 2.

Table 2

Final model beta-estimates and standard errors

Item	Beta Estimate	Std. Error	Odds Ratio	95% CI
Acute myeloid leukaemia	0.62	0.26	1.86	1.13–3.07
Brain	−0.43	0.39	0.65	0.3–1.38
Carcinoma	15.84	1455.4	7500	0–10 000
Ewings	−0.61	0.66	0.54	0.15–1.99
Germ cell tumour (extracranial)	−0.07	0.88	0.94	0.17–5.21
Hepatoblastoma	0.45	0.57	1.57	0.51–4.8
High-grade brain tumour	−0.33	0.46	0.72	0.29–1.78
Hodgkins	−0.39	0.7	0.68	0.17–2.68
High-risk neuroblastoma	0.88	0.66	2.4	0.66–8.76
LCH	−13.39	1025.44	0	0–10 000
Low-grade brain tumour	−13.45	677.94	0	0–10 000
Neuroblastoma	0.45	0.5	1.57	0.59–4.13
Non-Hodgkins Lymphoma	−0.45	0.32	0.64	0.34–1.19
Osteosarcoma	−1.13	0.57	0.32	0.11–0.98
Other tumour tyoe	0.76	0.77	2.13	0.47–9.6
Retinoblastoma	0.52	0.86	1.68	0.31–9
Relapsed acute leukaemia	0.35	0.18	1.42	1–2.03
Rhabdomyosarcoma	−0.23	0.32	0.79	0.42–1.48
Other sarcoma	0.18	0.82	1.2	0.24–5.98
Wilms	−0.47	0.66	0.63	0.17–2.3
Centralised temperature (actual temperature—37 °C)	0.54	0.14	1.71	1.29–2.27
Physician impression of clinically severe unwell	0.75	0.19	2.11	1.45–3.08
Haemoglobin (g dl−1)	0.17	0.05	1.19	1.08–1.31
Loge (white cell count × 106)	−0.28	0.1	0.75	0.62–0.92
Loge (absolute monocyte count per mm3)	−0.2	0.06	0.82	0.73–0.92

Abbreviations: CI=confidence interval; LCH=Langerhans cell histiocytosis.

Note: acute lymphoblastic leukaemia is the reference case.

Classes in italics have too little information to be clinically meaningful.

Assessing if a complex model gives meaningful benefits over a simpler one can be seen by comparing the AUROC values apparent in the derivation data. The model using clinical variables only (tumour type, temperature, severely unwell) gave an AUROC of 0.697 (95% CI 0.660–0.734). The addition of the simple full blood count variables (haemoglobin, white count and monocyte count) improved the prediction further to an AUROC of 0.736 (95% CI 0.698–0.774).

Discussion

The final multivariable predictive model had six simple components: tumour type, temperature, clinical description of being ‘severely unwell' and the results of measurements of haemoglobin concentration, total white cell count and AMC. It showed moderate discrimination (AUROC 0.723, 95% CI 0.711–0.759) and good calibration between predicted and actual estimates of the risk of MDI when assessed across the range of predictive values. The model was robust to bootstrap and cross-validation sensitivity analyses, and built by adding specialist investigations only after considering the simpler pieces of information, ensuring that if extra tests with additional costs are required, they are shown to add predictive power to existing variables.(Altman ).

The prediction model developed contained five items which showed consistent relationships across the different study groups (type of malignancy, temperature, unwellness, AMC and total white cell count) and one item (haemoglobin) that was less consistent. The heterogeneity in this item makes it difficult to tell if the inclusion of haemoglobin would be applicable in alternative data sets and subsequently in clinical practice.

The inclusion of clinical impression of the child being significantly unwell as a predictive variable is controversial. The clinical impression of a trained physician is held by some to be of great importance, and by others as unhelpfully subjective and poorly reproducible. Previous investigators have demonstrated objective replicability in paediatric oncology practice (Klaassen ), and shown its value in predicting infectious risk in children without an oncological diagnosis(Bruel ; Scott ) and it is a key element of the widely used (adult) low-risk scoring system of the Multinational Supportive Care in Cancer organisation. (Klastersky ) We believe that our analysis supports the use of this variable as providing additional useful information above the objective elements of ‘vital signs' and full blood count in routine practice.

The value of using both total white count and AMC, rather than one of these items or absolute neutrophil count (ANC), was assessed during the model building. The use of AMC improved AIC (model fit), discrimination and calibration compared with ANC, as did the additional use of total white count. While this appears to contradict the common understanding of how neutrophil counts most importantly indicate the risk of infection—the clinical situation is defined as ‘neutropenic sepsis', not ‘monocytopenic sepsis'—this analysis confirms data from multiple prior investigations (Rackoff ; Klaassen ; Baorto ; Tezcan ). The finding may be explained physiologically by both the neutrophil response to infection and the observed pattern of marrow recovery after chemotherapy, with monocytes preceding neutrophils (Mullen and Buchanan, 1990).

The study was limited in its reliance on study-defined MDI as the key outcome to be predicted, rather than a more comprehensive, patient-important assessment of an adverse outcome of the episode. The lack of a core, agreed definition of ‘significant adverse outcome' of FN remains a problem to researchers in this area (Haeusler , 2013b). If one was to be universally agreed, the PICNICC data may be suitable to recalibrate using an alternative outcome.

Relying on the investigators' original definition of MDI allows the different quality control processes (e.g., numbers of sets of blood cultures taken and volume of blood, or the use of prophylactic antibiotics before presentation) to vary between-studies. This may introduce different degrees of bias between-studies, compared with the ‘true' MDI status of patients. If this was an important issue, the between-study heterogeneity in predictors would have been marked, and this was not the case in this analysis.

The issue of missing data, almost entirely at the level of study where predictor or outcome variables were ‘not recorded' rather than ‘missing', led to a smaller data set being available for the development of the multivariable analyses. Largely, this data set was similar to the total data set when the univariable analyses were compared, except for removing patients who had received HSCT. Although developments in the handling of missing data within studies using simulation have produced guidelines using imputation techniques to maximise the value of the IPD data collected (Burgess ; Resche-Rigon ; Ahmed ), the application of such methods in the situation of unreported data has yet to be explored. Future studies will benefit from adhering to the core items recommended in the recent international position paper on research in this area (Haeusler , 2013b).

The clinical implementation of the prediction model will require it to be clearly understood by the clinical teams, based on sound data, and easily integrated into practice. The full PICNICC model has complexity (with the series of different predictors for type of malignancy, and the use of log-transformed data) which makes it likely to be unwieldy unless made easily applicable, for example on a spreadsheet (for a basic version see: http://goo.gl/3AOa9R) or smart phone ‘app', into which the clinicians would enter the data and return a calculated probability of MDI. The model, despite its derivation from data taken from different geographies and eras, requires evaluation in clinical practice from alternative data sets. This could be accomplished by a further global collaboration, building on the existing work, to formally assess the validity in further new data sets, or by individual institutions against locally collected data as a service improvement project. It also needs work undertaking to determine how a prediction of risk can be turned into a decision to employ a particular management strategy, which will require qualitative work with health-care practitioners, families and young people with febrile neutropenia.

We need to collect greater quantities of information on the additional benefit of particular biomarkers and good quality data on their comparative efficacy in initial risk stratification. Moving beyond the initial treatment of FN, and focusing on how we should treat patients with either a defined MDI or those without a clear cause, the patterns of how biomarkers change over time which reflect response to treatment will require evaluation, and also how these patterns may vary both between individuals and within individuals after different elements of their cancer treatment.

This individual patient data meta-analysis represents the largest of its kind; the output of a global collaboration which has shared thousands of items of data and developed a predictive model for MDI which is robust to internal validation techniques. It has demonstrated that such a project is feasible across many different jurisdictions and eras of study, and can provide important verifications and negations of commonly held beliefs.

This prediction model provides a robust method of determining an individual's chance of MDI during an episode of FN. It should be validated, and used in the context of a clinical trial to enable shared decisions to be made with parents or young people about intensity and location of care for each episode. Such a study should be achievable to prove or disprove the utility of this approach to treatment (Manji ) and solidly improve the management of this frequent complication of childhood cancer therapy.