Could the clinical interpretability of subgroups detected using clustering methods be improved by using a novel two-stage approach?

A Patient Stratification Approach to Identifying the Likelihood of Continued Chronic Depression and Relapse Following Treatment for Depression

Background: Subgrouping methods have the potential to support treatment decision making for patients with depression. Such approaches have not been used to study the continued course of depression or likelihood of relapse following treatment. Method: Data from individual participants of seven randomised controlled trials were analysed. Latent profile analysis was used to identify subgroups based on baseline characteristics. Associations between profiles and odds of both continued chronic depression and relapse up to one year post-treatment were explored. Differences in outcomes were investigated within profiles for those treated with antidepressants, psychological therapy, and usual care. Results: Seven profiles were identified; profiles with higher symptom severity and long durations of both anxiety and depression at baseline were at higher risk of relapse and of chronic depression. Members of profile five (likely long durations of depression and anxiety, moderately-severe symptoms, and past antidepressant use) appeared to have better outcomes with psychological therapies: antidepressants vs. psychological therapies (OR (95% CI) for relapse = 2.92 (1.24–6.87), chronic course = 2.27 (1.27–4.06)) and usual care vs. psychological therapies (relapse = 2.51 (1.16–5.40), chronic course = 1.98 (1.16–3.37)). Conclusions: Profiles at greater risk of poor outcomes could benefit from more intensive treatment and frequent monitoring. Patients in profile five may benefit more from psychological therapies than other treatments.

Adding Physical Impairment to Risk Stratification Improved Outcome Prediction in Low Back Pain

OBJECTIVE

Identifying subgroups of low back pain (LBP) has the potential to improve prediction of clinical outcomes. Risk stratification is one such strategy that identifies similar characteristics indicative of a common clinical outcome trajectory. The purpose of this study was to determine if an empirically derived subgrouping approach based on physical impairment measures improves information provided from the STarT Back Tool (SBT).

METHODS

At baseline in this secondary analysis of a cohort study, patients (N = 144) receiving physical therapy for LBP completed the SBT and tests (active lumbar flexion, extension, lateral bending, and passive straight-leg raise) from a validated physical impairment index. Clinical outcomes were assessed at 4 weeks and included the Numerical Pain Rating Scale and Oswestry Disability Index. Exploratory hierarchical agglomerative cluster analysis identified empirically derived subgroups based on physical impairment measures. Independent samples t testing and chi-square analysis were used to assess baseline subgroup differences in demographic and clinical measures. Spearman rho correlation coefficient was used to assess baseline SBT risk and impairment subgroup relationships, and a 3-way mixed-model ANOVA was used to assessed SBT risk and impairment subgroup relationships with clinical outcomes at 4 weeks.

RESULTS

Two physical impairment-based subgroups emerged from cluster analysis: (1) low-risk impairment (n = 119, 81.5%), characterized by greater lumbar mobility; and (2) high-risk impairment (n = 25, 17.1%), characterized by less lumbar mobility. A weak, positive relationship was observed between baseline SBT risk and impairment subgroups (rs = .170). An impairment-by-SBT risk-by-time interaction effect was observed for Oswestry Disability Index scores but not for Numerical Pain Rating Scale scores at 4 weeks.

CONCLUSIONS

Physical impairment subgroups were not redundant with SBT risk categories and could improve prediction of 4-week LBP disability outcomes. Physical impairment subgroups did not improve the prediction of 4-week pain intensity scores.

IMPACT

Subgroups based on physical impairment and psychosocial risk could lead to better prediction of LBP disability outcomes and eventually allow for treatment options tailored to physical and psychosocial risk.

Latent variable mixture modelling and individual treatment prediction

Understanding which groups of patients are more or less likely to benefit from specific treatments has important implications for healthcare. Many personalised medicine approaches in mental health employ variable-centred approaches to predicting treatment response, yet person-centred approaches that identify clinical profiles of patients can provide information on the likelihood of a range of important outcomes. In this paper, we discuss the use of latent variable mixture modelling and demonstrate its use in the application of a patient profiling algorithm using routinely collected patient data to predict outcomes from psychological treatments. This validation study analysed data from two services, which included n = 44,905 patients entering treatment. There were different patterns of reliable recovery, improvement and clinical deterioration from therapy, across the eight profiles which were consistent over time. Outcomes varied between different types of therapy within the profiles: there were significantly higher odds of reliable recovery with High Intensity therapies in two profiles (32.5% of patients) and of reliable improvement in three profiles (32.2% of patients) compared with Low Intensity treatments. In three profiles (37.4% of patients) reliable recovery was significantly more likely if patients had CBT vs Counselling. The developments and potential application of latent variable mixture approaches are further discussed.

Toward the Development of Data-Driven Diagnostic Subgroups for People With Patellofemoral Pain Using Modifiable Clinical, Biomechanical, and Imaging Features

BACKGROUND

Unfavorable treatment outcomes for people with patellofemoral pain (PFP) have been attributed to the potential existence of subgroups that respond differently to treatment.

OBJECTIVES

This study aimed to identify subgroups within PFP by combining modifiable clinical, biomechanical, and imaging features and exploring the prognosis of these subgroups.

METHODS

This was a longitudinal cohort study, with baseline cluster analyses. Baseline data were analyzed using a 2-stage cluster analysis; 10 features were analyzed within 4 health domains before being combined at the second stage. Prognosis of the subgroups was assessed at 12 months, with subgroup differences reported as global rating of change and analyzed with an exploratory logistic regression adjusted for known confounders.

RESULTS

Seventy participants were included (mean age, 31 years; 43 [61%] female). Cluster analysis revealed 4 subgroups: "strong," "pronation and malalignment," "weak," and "active and flexible." Descriptively, compared to the strong subgroup (55% favorable), the odds of a favorable outcome were lower in the weak subgroup (31% favorable; adjusted odds ratio [OR] = 0.30; 95% confidence interval [CI]: 0.07, 1.36) and the pronation and malalignment subgroup (50%; OR = 0.64; 95% CI: 0.11, 3.66), and higher in the active and flexible subgroup (63%; OR = 1.24; 95% CI: 0.20, 7.51). After adjustment, compared to the strong subgroup, differences between some subgroups remained substantive, but none were statistically significant.

CONCLUSION

In this relatively small cohort, 4 PFP subgroups were identified that show potentially different outcomes at 12 months. Further research is required to determine whether a stratified treatment approach using these subgroups would improve outcomes for people with PFP.

LEVEL OF EVIDENCE

Diagnosis, level 2b. J Orthop Sports Phys Ther 2019;49(7):536-547. doi:10.2519/jospt.2019.8607.

Identifying heterogeneous health profiles of primary care utilizers and their differential healthcare utilization and mortality - a retrospective cohort study

BACKGROUND

Heterogeneity of population health needs and the resultant difficulty in health care resources planning are challenges faced by primary care systems globally. To address this challenge in population health management, it is critical to have a better understanding of primary care utilizers' heterogeneous health profiles. We aimed to segment a population of primary care utilizers into classes with unique disease patterns, and to report the 1 year follow up healthcare utilizations and all-cause mortality across the classes.

METHODS

Using de-identified administrative data, we included all adult Singapore citizens or permanent residents who utilized Singapore Health Services (SingHealth) primary care services in 2012. Latent class analysis was used to identify patient subgroups having unique disease patterns in the population. The models were assessed by Bayesian Information Criterion and clinical interpretability. We compared healthcare utilizations in 2013 and one-year all-cause mortality across classes and performed regression analysis to assess predictive ability of class membership on healthcare utilizations and mortality.

RESULTS

We included 100,747 patients in total. The best model (k = 6) revealed the following classes of patients: Class 1 "Relatively healthy" (n = 58,213), Class 2 "Stable metabolic disease" (n = 26,309), Class 3 "Metabolic disease with vascular complications" (n = 2964), Class 4 "High respiratory disease burden" (n = 1104), Class 5 "High metabolic disease without complication" (n = 11,122), and Class 6 "Metabolic disease with multi-organ complication" (n = 1035). The six derived classes had different disease patterns in 2012 and 1 year follow up healthcare utilizations and mortality in 2013. "Metabolic disease with multiple organ complications" class had the highest healthcare utilization (e.g. incidence rate ratio = 19.68 for hospital admissions) and highest one-year all-cause mortality (hazard ratio = 27.97).

CONCLUSIONS

Primary care utilizers are heterogeneous and can be segmented by latent class analysis into classes with unique disease patterns, healthcare utilizations and all-cause mortality. This information is critical to population level health resource planning and population health policy formulation.

Latent class analysis derived subgroups of low back pain patients - do they have prognostic capacity?

BACKGROUND

Heterogeneity in patients with low back pain is well recognised and different approaches to subgrouping have been proposed. One statistical technique that is increasingly being used is Latent Class Analysis as it performs subgrouping based on pattern recognition with high accuracy. Previously, we developed two novel suggestions for subgrouping patients with low back pain based on Latent Class Analysis of patient baseline characteristics (patient history and physical examination), which resulted in 7 subgroups when using a single-stage analysis, and 9 subgroups when using a two-stage approach. However, their prognostic capacity was unexplored. This study (i) determined whether the subgrouping approaches were associated with the future outcomes of pain intensity, pain frequency and disability, (ii) assessed whether one of these two approaches was more strongly or more consistently associated with these outcomes, and (iii) assessed the performance of the novel subgroupings as compared to the following variables: two existing subgrouping tools (STarT Back Tool and Quebec Task Force classification), four baseline characteristics and a group of previously identified domain-specific patient categorisations (collectively, the 'comparator variables').

METHODS

This was a longitudinal cohort study of 928 patients consulting for low back pain in primary care. The associations between each subgroup approach and outcomes at 2 weeks, 3 and 12 months, and with weekly SMS responses were tested in linear regression models, and their prognostic capacity (variance explained) was compared to that of the comparator variables listed above.

RESULTS

The two previously identified subgroupings were similarly associated with all outcomes. The prognostic capacity of both subgroupings was better than that of the comparator variables, except for participants' recovery beliefs and the domain-specific categorisations, but was still limited. The explained variance ranged from 4.3%-6.9% for pain intensity and from 6.8%-20.3% for disability, and highest at the 2 weeks follow-up.

CONCLUSIONS

Latent Class-derived subgroups provided additional prognostic information when compared to a range of variables, but the improvements were not substantial enough to warrant further development into a new prognostic tool. Further research could investigate if these novel subgrouping approaches may help to improve existing tools that subgroup low back pain patients.

Identifying subgroups of patients using latent class analysis: should we use a single-stage or a two-stage approach? A methodological study using a cohort of patients with low back pain

Background

Heterogeneity in patients with low back pain (LBP) is well recognised and different approaches to subgrouping have been proposed. Latent Class Analysis (LCA) is a statistical technique that is increasingly being used to identify subgroups based on patient characteristics. However, as LBP is a complex multi-domain condition, the optimal approach when using LCA is unknown. Therefore, this paper describes the exploration of two approaches to LCA that may help improve the identification of clinically relevant and interpretable LBP subgroups.

Methods

From 928 LBP patients consulting a chiropractor, baseline data were used as input to the statistical subgrouping. In a single-stage LCA, all variables were modelled simultaneously to identify patient subgroups. In a two-stage LCA, we used the latent class membership from our previously published LCA within each of six domains of health (activity, contextual factors, pain, participation, physical impairment and psychology) (first stage) as the variables entered into the second stage of the two-stage LCA to identify patient subgroups. The description of the results of the single-stage and two-stage LCA was based on a combination of statistical performance measures, qualitative evaluation of clinical interpretability (face validity) and a subgroup membership comparison.

Results

For the single-stage LCA, a model solution with seven patient subgroups was preferred, and for the two-stage LCA, a nine patient subgroup model. Both approaches identified similar, but not identical, patient subgroups characterised by (i) mild intermittent LBP, (ii) recent severe LBP and activity limitations, (iii) very recent severe LBP with both activity and participation limitations, (iv) work-related LBP, (v) LBP and several negative consequences and (vi) LBP with nerve root involvement.

Conclusions

Both approaches identified clinically interpretable patient subgroups. The potential importance of these subgroups needs to be investigated by exploring whether they can be identified in other cohorts and by examining their possible association with patient outcomes. This may inform the selection of a preferred LCA approach.

Electronic supplementary material

The online version of this article (doi:10.1186/s12891-017-1411-x) contains supplementary material, which is available to authorized users.

Using existing questionnaires in latent class analysis: should we use summary scores or single items as input? A methodological study using a cohort of patients with low back pain

Background

Latent class analysis (LCA) is increasingly being used in health research, but optimal approaches to handling complex clinical data are unclear. One issue is that commonly used questionnaires are multidimensional, but expressed as summary scores. Using the example of low back pain (LBP), the aim of this study was to explore and descriptively compare the application of LCA when using questionnaire summary scores and when using single items to subgrouping of patients based on multidimensional data.

Materials and methods

Baseline data from 928 LBP patients in an observational study were classified into four health domains (psychology, pain, activity, and participation) using the World Health Organization’s International Classification of Functioning, Disability, and Health framework. LCA was performed within each health domain using the strategies of summary-score and single-item analyses. The resulting subgroups were descriptively compared using statistical measures and clinical interpretability.

Results

For each health domain, the preferred model solution ranged from five to seven subgroups for the summary-score strategy and seven to eight subgroups for the single-item strategy. There was considerable overlap between the results of the two strategies, indicating that they were reflecting the same underlying data structure. However, in three of the four health domains, the single-item strategy resulted in a more nuanced description, in terms of more subgroups and more distinct clinical characteristics.

Conclusion

In these data, application of both the summary-score strategy and the single-item strategy in the LCA subgrouping resulted in clinically interpretable subgroups, but the single-item strategy generally revealed more distinguishing characteristics. These results 1) warrant further analyses in other data sets to determine the consistency of this finding, and 2) warrant investigation in longitudinal data to test whether the finer detail provided by the single-item strategy results in improved prediction of outcomes and treatment response.