Unsupervised phenotypic clustering for determining clinical status in children with cystic fibrosis

Discussion

Phenotypic clusters using a range of clinical outcomes collected during routine clinical care allow for a comprehensive overview of CF health, and our results show that they meaningfully represent both mild and severe classes of CF disease. Within clusters, lung function was concordant with disease outcomes, whereas the range of individual lung function values observed within each cluster was wide. In the current era of CF, where up to 85% of affected children are reported to have mild to normal lung function [4–6], a multifactorial tool may provide further insight into disease progression.

The unsupervised algorithm benefits from the exclusion of FEV₁, since young children who cannot perform spirometry are still captured. Where FEV₁ is normal, the clusters may aid to explain who is at a greater risk of hospitalisation or pulmonary exacerbation and who could benefit from targeted management.

It was surprising that FEV₁ in cluster D increased over 1 year in younger children, although it suggests that clinicians may already recognise that these individuals have more severe disease and require more intense interventions. The greater variability in the slopes over 12 months further suggests that these trends may reflect treatment effects.

The validation carried out in the GOSH population was very similar to the results of the TCF data re-analysed with a subset of the original population (n_people=338). Although the GOSH population was at a higher risk of hospitalisations overall, a further investigation into clinical practices reveals differences between these two centres. For instance, routine regular hospitalisations for i.v. antibiotics are common for children with severe disease at GOSH, whereas hospitalisations in Toronto are typically only for acute exacerbations. Despite this significant difference in management approach, FEV₁ values were similar across clusters between the two populations. Clustering on the same variables yielded similar outcomes, highlighting the robustness of the cluster method. To ensure applicability in the modern CF era, the model should be updated in a predominantly newborn screening cohort, as well as in genetically diverse cohorts to ensure generalisability.

Multivariate scores have rarely been used in routine clinical care, but as EHR become more common and data are stored centrally, the implementation of this type of phenotypic clustering into clinical practice will become more feasible. The KNN algorithm can be used to calculate a cluster for each CF encounter based on the input of the nine cluster variables (table 2). As such, following appropriate governance and evaluation, the algorithm could be incorporated into EHR systems to provide clinicians with an overall picture of patient status and inform clinical decision making. Investigation into treatment effects across clusters with chronic therapies (e.g. dornase alfa and hypertonic saline) were limited in this study by missing and inconsistent data, but need to be explored further to better understand whether treatments can modify disease severity status.

The algorithm could also be implemented in patient portals to their EHR or apps to provide patients with a more comprehensive picture of their health status where single clinical measures, such as BMI or recent infection may not be as meaningful or interpretable indicators of health. Future work involving patients/families will highlight the appropriate presentation of cluster labels to better provide insight into overall health status.

Additional potential application use cases of clusters are as an end-point in research studies where changes in FEV₁ are not detectable; improvements in health status may instead be indicated by movement from a severe cluster (C/D) to a mild cluster (A/B). This may be particularly attractive in trials involving young children, and/or in the rapidly evolving era of highly effective cystic fibrosis transmembrane conductance regulator (CFTR) modulator treatments being integrated into routine care. Clusters may also be used in national registry reports to standardise clinics or regions by disease severity for matched comparisons of populations, or to highlight those who may benefit from more resources.

A current barrier to implementation is the issue of missing clinical data, in which future analyses are necessary to triangulate evidence from existing data to ensure robust estimation of clusters. In addition, future work will explore the integration of filters to prevent the miscalculation of clusters from unrealistic clinical values. A limitation of the approach is that we standardised variables, which means the relative importance of each variable in defining the transition between cluster is unknown, and future work is needed to explore this.

Cluster analysis involves largely subjective decisions at each step of the cluster pipeline, and further refinement of the model may include the selection of different parameters to better understand the stability of the model. At present, the optimal model has a good performance and is an important first step, but it may not be the best model as highly effective CFTR modulators begin to change clinical outcomes and prognosis. Future works should determine how to routinely update models as patient characteristics and available data changes. For instance, cough was repeatedly highlighted as important in defining outcomes during iterative clustering, but did not appear to impact the GOSH validation. While cough is limited by subjectivity and a high proportion of missing values, its inclusion allowed the model to encompass patient-reported symptoms. More objective symptoms collected routinely through EHR or patient apps may be required to better capture the lived experience of the patient. CFTR modulator therapy was excluded from the model due to minimal data, but should be considered once these treatments are available to a wider proportion of CF patients. Similarly, lung clearance index should be considered for inclusion as it becomes a more routinely collected clinical variable, and extrapulmonary manifestations not captured in the registry should be included when using EHR data. An advantage of using an integrated cluster algorithm means that the variables can be updated, the time period adjusted, and subgroups refined as new information becomes available.

This analysis focused on time-dependent, continuous variables to assess cluster change over time, but also because clustering with partitioning around medoids is more favourable using continuous variables. For example, the inclusion of genetic sex resulted in clusters completely defined by sex alone, with less association with clinical outcomes. This demonstrates that the unsupervised nature of clustering, in which the algorithm aims to find similar groups without a purpose, requires clinical interpretation to ensure real-world value. In selecting meaningful, continuous variables, the look-back window to capture an individual's clinical history means the frequency of microbiology sampling could introduce a potential bias, and sicker patients with more hospitalisations may be overrepresented.

The phenotypic clusters were broadly comparable to a previous study of an adult CF population, which found that P. aeruginosa, i.v. antibiotics and pancreatic insufficiency were consistent in severe clusters with a high risk of death, and vice versa for milder clusters with a low risk of death [13]. The study identified seven clusters within 25 variables. We restricted the number of possible clusters to between three and five to simplify the practical interpretation of the results, and we were able to reduce the number of variables to nine for a more feasible clinical application. Despite these differences, both approaches identified robust clusters with similar variables associated with severe disease. Other cluster analyses in CF have anchored clusters to physician-determined levels of disease severity, which may be biased by indication [24], or have included FEV₁ in the model [14]. Our analysis provides a more objective categorisation of disease severity, by statistically testing the clusters against known indicators of health that are mainly excluded from the model itself.

Cluster analyses have been applied to other diseases and further emphasise that knowledge-based inclusion of variables is necessary to ensure meaningful translation of the results [11, 12, 25, 26]. Another common approach has been to transform variables into linear combinations [10, 11, 25], which we decided would complicate the interpretation of our results. Two studies have gone further to prospectively apply the clusters by generating an external scoring system using decision tree analyses [13, 25]. The advantage of the KNN algorithm is that it directly allocates clusters to an encounter to predict disease severity. Our results show it has an extremely low error rate (<5%).

Multivariate clinical scoring systems have been derived in CF, but they typically assume an additive/multiplicative independent association between the variables included [27–30]. Cluster analysis instead makes no assumption about the nature of the relationship, only how individuals are similar to one another through shared patterns across the variables. Additionally, multivariate models must be linked to an event of interest or a specific outcome, and the advantage of this approach is that the clusters are correlated with important clinical outcomes, but not developed with an explicit prediction model of a single outcome.