Challenges in interpreting wheeze phenotypes: the clinical implications of statistical learning techniques

Understanding the heterogeneity of childhood allergic sensitization and its relationship with asthma

PURPOSE OF REVIEW

To review the current state of knowledge on the relationship between allergic sensitization and asthma; to lay out a roadmap for the development of IgE biomarkers that differentiate, in individual sensitized patients, whether their sensitization is important for current or future asthma symptoms, or has little or no relevance to the disease.

RECENT FINDINGS

The evidence on the relationship between sensitization and asthma suggests that some subtypes of allergic sensitization are not associated with asthma symptoms, whilst others are pathologic. Interaction patterns between IgE antibodies to individual allergenic molecules on component-resolved diagnostics (CRD) multiplex arrays might be hallmarks by which different sensitization subtypes relevant to asthma can be distinguished. These different subtypes of sensitization are associated amongst sensitized individuals at all ages, with different clinical presentations (no disease, asthma as a single disease, and allergic multimorbidity); amongst sensitized preschool children with and without lower airway symptoms, with different risk of subsequent asthma development; and amongst sensitized patients with asthma, with differing levels of asthma severity.

SUMMARY

The use of machine learning-based methodologies on complex CRD data can help us to design better diagnostic tools to help practising physicians differentiate between benign and clinically important sensitization.

Understanding progression from pre-school wheezing to school-age asthma: Can modern data approaches help?

Preschool wheezing and childhood asthma create a heavy disease burden which is only exacerbated by the complexity of the conditions. Preschool wheezing exhibits both "curricular" and "aetiological" heterogeneity: that is, heterogeneity across patients both in the time-course of its development and in its underpinning pathological mechanisms. Since these are not fully understood, but clinical presentations across patients may nonetheless be similar, current diagnostic labels are imprecise-not mapping cleanly onto underlying disease mechanisms-and prognoses uncertain. These uncertainties also make a identifying new targets for therapeutic intervention difficult. In the past few decades, carefully designed birth cohort studies have collected "big data" on a large scale, incorporating not only a wealth of longitudinal clinical data, but also detailed information from modalities as varied as imaging, multiomics, and blood biomarkers. The profusion of big data has seen the proliferation of what we term "modern data approaches" (MDAs)-grouping together machine learning, artificial intelligence, and data science-to make sense and make use of this data. In this review, we survey applications of MDAs (with an emphasis on machine learning) in childhood wheeze and asthma, highlighting the extent of their successes in providing tools for prognosis, unpicking the curricular heterogeneity of these conditions, clarifying the limitations of current diagnostic criteria, and indicating directions of research for uncovering the etiology of the diseases underlying these conditions. Specifically, we focus on the trajectories of childhood wheeze phenotypes. Further, we provide an explainer of the nature and potential use of MDAs and emphasize the scope of what we can hope to achieve with them.

Current state and prospects of artificial intelligence in allergy

The field of medicine is witnessing an exponential growth of interest in artificial intelligence (AI), which enables new research questions and the analysis of larger and new types of data. Nevertheless, applications that go beyond proof of concepts and deliver clinical value remain rare, especially in the field of allergy. This narrative review provides a fundamental understanding of the core concepts of AI and critically discusses its limitations and open challenges, such as data availability and bias, along with potential directions to surmount them. We provide a conceptual framework to structure AI applications within this field and discuss forefront case examples. Most of these applications of AI and machine learning in allergy concern supervised learning and unsupervised clustering, with a strong emphasis on diagnosis and subtyping. A perspective is shared on guidelines for good AI practice to guide readers in applying it effectively and safely, along with prospects of field advancement and initiatives to increase clinical impact. We anticipate that AI can further deepen our knowledge of disease mechanisms and contribute to precision medicine in allergy.

Phenotype and endotype based treatment of preschool wheeze

INTRODUCTION

Preschool wheeze (PSW) is a significant public health issue, with a high presentation rate to emergency departments, recurrent symptoms, and severe exacerbations. A heterogenous condition, PSW comprises several phenotypes that may relate to a range of pathobiological mechanisms. However, treating PSW remains largely generalized to inhaled corticosteroids and a short acting beta agonist, guided by symptom-based labels that often do not reflect underlying pathways of disease.

AREAS COVERED

We review the observable features and characteristics used to ascribe phenotypes in children with PSW and available pathobiological evidence to identify possible endotypes. These are considered in the context of treatment options and future research directions. The role of machine learning (ML) and modern analytical techniques to identify patterns of disease that distinguish phenotypes is also explored.

EXPERT OPINION

Distinct clusters (phenotypes) of severe PSW are characterized by different underlying mechanisms, some shared and some unique. ML-based methodologies applied to clinical, biomarker, and environmental data can help design tools to differentiate children with PSW that continues into adulthood, from those in whom wheezing resolves, identifying mechanisms underpinning persistence and resolution. This may help identify novel therapeutic targets, inform mechanistic studies, and serve as a foundation for stratification in future interventional therapeutic trials.

Modeling Wheezing Spells Identifies Phenotypes with Different Outcomes and Genetic Associates

Rationale: Longitudinal modeling of current wheezing identified similar phenotypes, but their characteristics often differ between studies. Objectives: We propose that a more comprehensive description of wheeze may better describe trajectories than binary information on the presence/absence of wheezing. Methods: We derived six multidimensional variables of wheezing spells from birth to adolescence (including duration, temporal sequencing, and the extent of persistence/recurrence). We applied partition-around-medoids clustering on these variables to derive phenotypes in five birth cohorts. We investigated within- and between-phenotype differences compared with binary latent class analysis models and ascertained associations of these phenotypes with asthma and lung function and with polymorphisms in asthma loci 17q12-21 and CDHR3 (cadherin-related family member 3). Measurements and Main Results: Analysis among 7,719 participants with complete data identified five spell-based wheeze phenotypes with a high degree of certainty: never (54.1%), early-transient (ETW) (23.7%), late-onset (LOW) (6.9%), persistent (PEW) (8.3%), and a novel phenotype, intermittent wheeze (INT) (6.9%). FEV₁/FVC was lower in PEW and INT compared with ETW and LOW and declined from age 8 years to adulthood in INT. 17q12-21 and CDHR3 polymorphisms were associated with higher odds of PEW and INT, but not ETW or LOW. Latent class analysis- and spell-based phenotypes appeared similar, but within-phenotype individual trajectories and phenotype allocation differed substantially. The spell-based approach was much more robust in dealing with missing data, and the derived clusters were more stable and internally homogeneous. Conclusions: Modeling of spell variables identified a novel intermittent wheeze phenotype associated with lung function decline to early adulthood. Using multidimensional spell variables may better capture wheeze development and provide a more robust input for phenotype derivation.

Trajectories of IgE sensitization to allergen molecules from childhood to adulthood and respiratory health in the EGEA cohort

BACKGROUND

Longitudinal studies assessing the association of profiles of allergen-specific IgE (sIgE) sensitization to a large range of allergen molecules and respiratory health are rare. We aimed to assess trajectories of molecular sIgE sensitization profiles from childhood to adulthood and their associations with respiratory health.

METHODS

IgE reactivity to microarrayed allergen molecules were measured in childhood (EGEA1) and 12 years later in adult life (EGEA2) among 291 EGEA participants (152 with asthma). At each time point, sIgE sensitization profiles were identified by latent class analysis (LCA) by considering IgE-reactivity to the 38 most prevalent respiratory allergens. The LCA-defined profiles were then studied in association with respiratory health.

RESULTS

At baseline, the mean (min-max) age of the population was 11 (4.5-16) years. The LCA identified four sIgE sensitization profiles which were very similar at both time points (% at EGEA1 and EGEA2); A: "no/few allergen(s)" (48%, 39%), B: "pollen/animal allergens" (18%, 21%), C: "most prevalent house dust mite allergens" (22%, 27%) and D: "many allergens" (12%, 13%). Overall, 73% of the participants remained in the same profile from childhood to adulthood. The profiles were associated with asthma and rhinitis phenotypes. Participants of profiles C and D had lower FEV₁ % and FEF_25-75 % as compared to profile A. Similar patterns of associations were observed for participants with asthma. There was no association with change in lung function.

CONCLUSION

Using high-resolution sIgE longitudinal data, the LCA identified four molecular sensitization profiles, mainly stable from childhood to adulthood, that were associated with respiratory health.

Recurrent Severe Preschool Wheeze: From Pre-Specified Diagnostic Labels to Underlying Endotypes

Rationale: Preschool wheezing is heterogeneous, but the underlying mechanisms are poorly understood.

Objectives: To investigate lower airway inflammation and infection in preschool children with different clinical diagnoses undergoing elective bronchoscopy and BAL.

Methods: We recruited 136 children aged 1–5 years (105 with recurrent severe wheeze [RSW]; 31 with nonwheezing respiratory disease [NWRD]). Children with RSW were assigned as having episodic viral wheeze (EVW) or multiple-trigger wheeze (MTW). We compared lower airway inflammation and infection in different clinical diagnoses and undertook data-driven analyses to determine clusters of pathophysiological features, and we investigated their relationships with prespecified diagnostic labels.

Measurements and Main Results: Blood eosinophil counts and percentages and allergic sensitization were significantly higher in children with RSW than in children with a NWRD. Blood neutrophil counts and percentages, BAL eosinophil and neutrophil percentages, and positive bacterial culture and virus detection rates were similar between groups. However, pathogen distribution differed significantly, with higher detection of rhinovirus in children with RSW and higher detection of Moraxella in sensitized children with RSW. Children with EVW and children with MTW did not differ in terms of blood or BAL-sample inflammation, or bacteria or virus detection. The Partition around Medoids algorithm revealed four clusters of pathophysiological features: 1) atopic (17.9%), 2) nonatopic with a low infection rate and high use of inhaled corticosteroids (31.3%), 3) nonatopic with a high infection rate (23.1%), and 4) nonatopic with a low infection rate and no use of inhaled corticosteroids (27.6%). Cluster allocation differed significantly between the RSW and NWRD groups (RSW was evenly distributed across clusters, and 60% of the NWRD group was assigned to cluster 4; P < 0.001). There was no difference in cluster membership between the EVW and MTW groups. Cluster 1 was dominated by Moraxella detection (P = 0.04), and cluster 3 was dominated by Haemophilus or Staphylococcus or Streptococcus detection (P = 0.02).

Conclusions: We identified four clusters of severe preschool wheeze, which were distinguished by using sensitization, peripheral eosinophilia, lower airway neutrophilia, and bacteriology.

Differential associations of allergic disease genetic variants with developmental profiles of eczema, wheeze and rhinitis

BACKGROUND

Allergic diseases (eczema, wheeze and rhinitis) in children often present as heterogeneous phenotypes. Understanding genetic associations of specific patterns of symptoms might facilitate understanding of the underlying biological mechanisms.

OBJECTIVE

To examine associations between allergic disease-related variants identified in a recent genome-wide association study and latent classes of allergic diseases (LCADs) in two population-based birth cohorts.

METHODS

Eight previously defined LCADs between birth and 11 years: "No disease," "Atopic march," "Persistent eczema and wheeze," "Persistent eczema with later-onset rhinitis," "Persistent wheeze with later-onset rhinitis," "Transient wheeze," "Eczema only" and "Rhinitis only" were used as the study outcome. Weighted multinomial logistic regression was used to estimate associations between 135 SNPs (and a polygenic risk score, PRS) and LCADs among 6345 individuals from The Avon Longitudinal Study of Parents and Children (ALSPAC). Heterogeneity across LCADs was assessed before and after Bonferroni correction. Results were replicated in Manchester Asthma and Allergy Study (MAAS) (n = 896) and pooled in a meta-analysis.

RESULTS

We found strong evidence for differential genetic associations across the LCADs; pooled PRS heterogeneity P-value = 3.3 × 10-14 , excluding "no disease" class. The associations between the PRS and LCADs in MAAS were remarkably similar to ALSPAC. Two SNPs (a protein-truncating variant in FLG and a SNP within an intron of GSDMB) had evidence for differential association (pooled P-values ≤ 0.006). The FLG locus was differentially associated across LCADs that included eczema, with stronger associations for LCADs with comorbid wheeze and rhinitis. The GSDMB locus in contrast was equally associated across LCADs that included wheeze.

CONCLUSIONS AND CLINICAL RELEVANCE

We have shown complex, but distinct patterns of genetic associations with LCADs, suggesting that heterogeneous mechanisms underlie individual disease trajectories. Establishing the combination of allergic diseases with which each genetic variant is associated may inform therapeutic development and/or predictive modelling.

Causes of variability in latent phenotypes of childhood wheeze

BACKGROUND

Latent class analysis (LCA) has been used extensively to identify (latent) phenotypes of childhood wheezing. However, the number and trajectory of discovered phenotypes differed substantially between studies.

OBJECTIVE

We sought to investigate sources of variability affecting the classification of phenotypes, identify key time points for data collection to understand wheeze heterogeneity, and ascertain the association of childhood wheeze phenotypes with asthma and lung function in adulthood.

METHODS

We used LCA to derive wheeze phenotypes among 3167 participants in the ALSPAC cohort who had complete information on current wheeze recorded at 14 time points from birth to age 16½ years. We examined the effects of sample size and data collection age and intervals on the results and identified time points. We examined the associations of derived phenotypes with asthma and lung function at age 23 to 24 years.

RESULTS

A relatively large sample size (>2000) underestimated the number of phenotypes under some conditions (eg, number of time points <11). Increasing the number of data points resulted in an increase in the optimal number of phenotypes, but an identical number of randomly selected follow-up points led to different solutions. A variable selection algorithm identified 8 informative time points (months 18, 42, 57, 81, 91, 140, 157, and 166). The proportion of asthmatic patients at age 23 to 24 years differed between phenotypes, whereas lung function was lower among persistent wheezers.

CONCLUSIONS

Sample size, frequency, and timing of data collection have a major influence on the number and type of wheeze phenotypes identified by using LCA in longitudinal data.

Development of allergic sensitization and its relevance to paediatric asthma

PURPOSE OF REVIEW

The purpose of this review is to summarize the recent evidence on the distinct atopic phenotypes and their relationship with childhood asthma. We start by considering definitions and phenotypic classification of atopy and then review evidence on its association with asthma in children.

RECENT FINDINGS

It is now well recognized that both asthma and atopy are complex entities encompassing various different sub-groups that also differ in the way they interconnect. The lack of gold standards for diagnostic markers of atopy and asthma further adds to the existing complexity over diagnostic accuracy and definitions. Although recent statistical phenotyping studies contributed significantly to our understanding of these heterogeneous disorders, translating these findings into meaningful information and effective therapies requires further work on understanding underpinning biological mechanisms.

SUMMARY

The disaggregation of allergic sensitization may help predict how the allergic disease is likely to progress. One of the important questions is how best to incorporate tests for the assessment of allergic sensitization into diagnostic algorithms for asthma, both in terms of confirming asthma diagnosis, and the assessment of future risk.

Validation of childhood asthma predictive tools: A systematic review

BACKGROUND

There is uncertainty about the clinical usefulness of currently available asthma predictive tools. Validation of predictive tools in different populations and clinical settings is an essential requirement for the assessment of their predictive performance, reproducibility and generalizability. We aimed to critically appraise asthma predictive tools which have been validated in external studies.

METHODS

We searched MEDLINE and EMBASE (1946-2017) for all available childhood asthma prediction models and focused on externally validated predictive tools alongside the studies in which they were originally developed. We excluded non-English and non-original studies. PROSPERO registration number is CRD42016035727.

RESULTS

From 946 screened papers, eight were included in the review. Statistical approaches for creation of prediction tools included chi-square tests, logistic regression models and the least absolute shrinkage and selection operator. Predictive models were developed and validated in general and high-risk populations. Only three prediction tools were externally validated: the Asthma Predictive Index, the PIAMA and the Leicester asthma prediction tool. A variety of predictors has been tested, but no studies examined the same combination. There was heterogeneity in definition of the primary outcome among development and validation studies, and no objective measurements were used for asthma diagnosis. The performance of tools varied at different ages of outcome assessment. We observed a discrepancy between the development and validation studies in the tools' predictive performance in terms of sensitivity and positive predictive values.

CONCLUSIONS

Validated asthma predictive tools, reviewed in this paper, provided poor predictive accuracy with performance variation in sensitivity and positive predictive value.

Features of asthma which provide meaningful insights for understanding the disease heterogeneity

BACKGROUND

Data-driven methods such as hierarchical clustering (HC) and principal component analysis (PCA) have been used to identify asthma subtypes, with inconsistent results.

OBJECTIVE

To develop a framework for the discovery of stable and clinically meaningful asthma subtypes.

METHODS

We performed HC in a rich data set from 613 asthmatic children, using 45 clinical variables (Model 1), and after PCA dimensionality reduction (Model 2). Clinical experts then identified a set of asthma features/domains which informed clusters in the two analyses. In Model 3, we reclustered the data using these features to ascertain whether this improved the discovery process.

RESULTS

Cluster stability was poor in Models 1 and 2. Clinical experts highlighted four asthma features/domains which differentiated the clusters in two models: age of onset, allergic sensitization, severity, and recent exacerbations. In Model 3 (HC using these four features), cluster stability improved substantially. The cluster assignment changed, providing more clinically interpretable results. In a 5-cluster model, we labelled the clusters as: "Difficult asthma" (n = 132); "Early-onset mild atopic" (n = 210); "Early-onset mild non-atopic: (n = 153); "Late-onset" (n = 105); and "Exacerbation-prone asthma" (n = 13). Multinomial regression demonstrated that lung function was significantly diminished among children with "Difficult asthma"; blood eosinophilia was a significant feature of "Difficult," "Early-onset mild atopic," and "Late-onset asthma." Children with moderate-to-severe asthma were present in each cluster.

CONCLUSIONS AND CLINICAL RELEVANCE

An integrative approach of blending the data with clinical expert domain knowledge identified four features, which may be informative for ascertaining asthma endotypes. These findings suggest that variables which are key determinants of asthma presence, severity, or control may not be the most informative for determining asthma subtypes. Our results indicate that exacerbation-prone asthma may be a separate asthma endotype and that severe asthma is not a single entity, but an extreme end of the spectrum of several different asthma endotypes.

Disaggregating asthma: Big investigation versus big data

We are facing a major challenge in bridging the gap between identifying subtypes of asthma to understand causal mechanisms and translating this knowledge into personalized prevention and management strategies. In recent years, "big data" has been sold as a panacea for generating hypotheses and driving new frontiers of health care; the idea that the data must and will speak for themselves is fast becoming a new dogma. One of the dangers of ready accessibility of health care data and computational tools for data analysis is that the process of data mining can become uncoupled from the scientific process of clinical interpretation, understanding the provenance of the data, and external validation. Although advances in computational methods can be valuable for using unexpected structure in data to generate hypotheses, there remains a need for testing hypotheses and interpreting results with scientific rigor. We argue for combining data- and hypothesis-driven methods in a careful synergy, and the importance of carefully characterized birth and patient cohorts with genetic, phenotypic, biological, and molecular data in this process cannot be overemphasized. The main challenge on the road ahead is to harness bigger health care data in ways that produce meaningful clinical interpretation and to translate this into better diagnoses and properly personalized prevention and treatment plans. There is a pressing need for cross-disciplinary research with an integrative approach to data science, whereby basic scientists, clinicians, data analysts, and epidemiologists work together to understand the heterogeneity of asthma.

Relationship between cytokine expression patterns and clinical outcomes: two population-based birth cohorts

BACKGROUND

Models that incorporate patterns of multiple cytokine responses to allergens, rather than individual cytokine production, may better predict sensitization and asthma.

OBJECTIVE

To characterize the patterns of peripheral blood mononuclear cells' (PBMCs) cytokine responses to house dust mite (HDM) allergens among children from two population-based birth cohorts using machine learning techniques.

METHODS

PBMCs collected at 8 years of age from the UK Manchester Asthma and Allergy Study (n = 268) and at 14 years of age from the Australian Raine Study (n = 1374) were cultured with HDM extract (10 μg/ml). Cytokine expression (IL-13, IL-5, IFN-γ, and IL10) was measured in the supernatant. Cytokine patterns were identified using a Gaussian mixture model clustering, and classification stability was assessed by bootstrapping.

RESULTS

A six-class model indicated complex latent structure of cytokine expression. Based on the characteristics of each class, we designated them as follows: 'Nonresponders' (n = 905, 55%); 'IL-10 responders' (n = 49, 3%); 'IFN-γ and IL-13 medium responders' (n = 56, 3.4%); 'IL-13 medium responders' (n = 351, 21.4%); 'IL-5 and IL-13 medium responders' (n = 77, 4.7%); and 'IL-13 and IL-5 high responders' (n = 204, 12.4%). 'IL-13 and IL-5 high responders' were at much higher risk of HDM sensitization and asthma compared to all other classes, with 88% of children assigned to this class being sensitized and 28.5% having asthma.

CONCLUSION

Using model-based clustering, we identified several distinct patterns of cytokine response to HDM and observed interplay between cytokine expression level, cytokine patterns (especially IL-13 and IL-5), and clinical outcomes. 'IL-13 and IL-5 high responders' class was strongly associated with HDM sensitization. However, among HDM-sensitized children, one-third showed no PBMC response to HDM, and the majority of HDM-sensitized children did not have asthma or wheeze. Our findings suggest that positive HDM 'allergy tests' and asthma are associated with a broad range of immunophenotypes, which may have important implications for the use of cytokine-targeted treatment approaches.

A clustering approach to identify severe bronchiolitis profiles in children

OBJECTIVE

Although bronchiolitis is generally considered a single disease, recent studies suggest heterogeneity. We aimed to identify severe bronchiolitis profiles using a clustering approach.

METHODS

We analysed data from two prospective, multicentre cohorts of children younger than 2 years hospitalised with bronchiolitis, one in the USA (2007-2010 winter seasons, n=2207) and one in Finland (2008-2010 winter seasons, n=408). Severe bronchiolitis profiles were determined by latent class analysis, classifying children based on clinical factors and viral aetiology.

RESULTS

In the US study, four profiles were identified. Profile A (12%) was characterised by history of wheezing and eczema, wheezing at the emergency department (ED) presentation and rhinovirus infection. Profile B (36%) included children with wheezing at the ED presentation, but, in contrast to profile A, most did not have history of wheezing or eczema; this profile had the largest probability of respiratory syncytial virus infection. Profile C (34%) was the most severely ill group, with longer hospital stay and moderate-to-severe retractions. Profile D (17%) had the least severe illness, including non-wheezing children with shorter length of stay. Two of these profiles (A and D) were replicated in the Finnish cohort; a third group ('BC') included Finnish children with characteristics of profiles B and/or C in the US population.

CONCLUSIONS

Several distinct clinical profiles (phenotypes) were identified by a clustering approach in two multicentre studies of children hospitalised for bronchiolitis. The observed heterogeneity has important implications for future research on the aetiology, management and long-term outcomes of bronchiolitis, such as future risk of childhood asthma.

Distinguishing Asthma Phenotypes Using Machine Learning Approaches

Asthma is not a single disease, but an umbrella term for a number of distinct diseases, each of which are caused by a distinct underlying pathophysiological mechanism. These discrete disease entities are often labelled as 'asthma endotypes'. The discovery of different asthma subtypes has moved from subjective approaches in which putative phenotypes are assigned by experts to data-driven ones which incorporate machine learning. This review focuses on the methodological developments of one such machine learning technique-latent class analysis-and how it has contributed to distinguishing asthma and wheezing subtypes in childhood. It also gives a clinical perspective, presenting the findings of studies from the past 5 years that used this approach. The identification of true asthma endotypes may be a crucial step towards understanding their distinct pathophysiological mechanisms, which could ultimately lead to more precise prevention strategies, identification of novel therapeutic targets and the development of effective personalized therapies.

The Study Team for Early Life Asthma Research (STELAR) consortium 'Asthma e-lab': team science bringing data, methods and investigators together

We created Asthma e-Lab, a secure web-based research environment to support consistent recording, description and sharing of data, computational/statistical methods and emerging findings across the five UK birth cohorts. The e-Lab serves as a data repository for our unified dataset and provides the computational resources and a scientific social network to support collaborative research. All activities are transparent, and emerging findings are shared via the e-Lab, linked to explanations of analytical methods, thus enabling knowledge transfer. eLab facilitates the iterative interdisciplinary dialogue between clinicians, statisticians, computer scientists, mathematicians, geneticists and basic scientists, capturing collective thought behind the interpretations of findings.