Healthcare utilization after a first hospitalization for COPD: a new approach of State Sequence Analysis based on the '6W' multidimensional model of care trajectories

The studied population included all patients living in the province of Quebec, Canada, aged 40–84 years, with a first and urgent hospitalization for COPD (main diagnosis ICD-10: J40-J44, J47) between January 1st and December 31st, 2013 (Fig. 1). To increase the likelihood of correct diagnosis, we included only patients aged at least 40 years old and diagnosed with COPD in the 2-year period before the index date (defined as at least two physician ambulatory visits or at least one hospitalization with a secondary diagnosis; ICD-9: 490–494, 496, ICD-10: J40-J44, J47). A hospitalization was considered urgent if the patient was admitted following an ED visit and if the type of hospital admission was considered urgent. The index date refers to the date of the ED visit leading to such hospitalization. To select relatively stable patients with infrequent exacerbations, patients who were hospitalized with a primary diagnosis of COPD in the 3 years before the index date were excluded. We also excluded patient with a diagnosed cancer (as reported in the secondary diagnoses at index hospitalization), and institutionalized patients, identified through the location of the medical services in the 2 years preceding index date. Finally, to describe healthcare use during the one-year follow-up, we excluded patients that died during this period.

CTs were defined as sequences in time (dimension “when”) of healthcare utilization associated with the dimensions of the “6W” model [7] and measured in the year after the index date (date of ED arrival to index hospitalization). The main variable consisted of a classification (typology) of CTs (see Statistical analysis section). Information used to define healthcare sequences were: the category of consultation’s care setting (hospital, ED, outpatient clinic, primary care clinic, CLSC) (dimension “where”); the reason of consultation (respiratory disease, CVD, other) (dimension “why”); and the encountered care provider’s category (pulmonologist, cardiologist, internist, other MD specialist, general practitioner (GP), respiratory therapist in CLSC, nurse in CLSC) (dimension “which”).

Demographics and clinical characteristics of the patient (dimension “who”) included: sex; age; physical and mental health conditions, as summarized by a comorbidity index, the severity of index hospitalization; GP affiliation (yes/no); and public prescription drug insurance plan (PPDIP) status. We identified physical and mental conditions using the diagnoses reported in MED-ECHO and in the medical services register in the 2-year period before index date (one diagnosis during a hospitalization or at least two in the medical services register). The comorbidity index selected is proposed by Simard et al., [34] which uses a combination of 31 conditions from the 17 Charlson’s and the 30 Elixhauser medical conditions [35, 36]. The severity of the index hospitalization was measured using variables length of stay, intensity of resources index (NIRRU), and clinical severity index (Weak, Moderate, High, and Extreme). To determine if a patient was affiliated to a GP, we considered all ambulatory visits to GPs (excluding EDs) during the 2-year period before index date. A patient was considered affiliated to a GP if at least 75% of these visits were made to the same GP [37]. If a patient had only one visit to a GP during that period, they were considered affiliated to a GP. The PPDIP status includes four categories: not admissible to PPDIP (individuals with a private drug insurance plan), admissible to PPDIP and age ≥ 65 years with guaranteed income supplement (GIS), admissible to PPDIP and being a recipient of last-resort financial assistance (LRFA), or regular recipient of PPDIP.

Patients’ residential characteristics were also considered as covariables and included the rural-urban characteristic of the residential neighbourhood (dissemination area) (metropolitan: ≥100,000 inhabitants, small town: 10,000–100,000 inhabitants, rural: < 10,000 inhabitants with high to low metropolitan influence) as well as its material and social deprivation quartiles [38].

To characterize the typology of CTs (homogeneous groups of CTs), we used a state sequence analysis (SSA) [39]. This method was specifically developed to analyse sequential data [24‐26]. Because of the multidimensional nature of CTs (the where, why, and which dimensions), we used a modified version of SSA (Fig. 2). The main steps of this multidimensional modified version of SSA were to: 1) define the cohort of patients, the observation period and the time unit (e.g. days, weeks, months); 2) for each of the three dimensions, select categorical states, specify their priorities as many are not mutually exclusive (e.g. primary care consultation, emergency visit and urgent hospitalization occurring in the same time unit) and measure the state sequences to generate patient-sequences; 3) for each of the three dimensions, calculate the distance between each pair of patient-sequences using an appropriate distance or dissimilarity measure method, resulting in three distance matrices; 4) calculate a pooled distance matrix by summing the three dimension-specific matrices; 5) based on the pooled distance matrix calculated in step 4, choose and apply a classification method resulting in groups of distinct patient-sequences - the CT typology; and finally 6) display results by visual representations offered by SSA to interpret the CT typology [32].

For this study, we measured CTs in the year following the index hospital admission (index date) and chose “weeks” as the time unit (step 1). We defined the following dimension-specific states (step 2): A) the “where” dimension, i.e., the type of care setting or unit patients consulted. Seven possible states, in priority order: hospital (urgent) (state 1), ED (state 2), hospital (non-urgent) (state 3), outpatient clinic (state 3), primary care or private clinic/other (state 4), consultation in CLSC (state 5), and a state for “no healthcare utilization” (state 7); B) the “why” dimension, i.e., the reason for the consultation. Four possible states, in priority order: respiratory disease (state 1), CVD (state 2), other reason (state 3), and a state for “no healthcare utilization” (state 4); C) the “which” dimension, i.e., the type of physician consulted. Height possible states, in priority order: pulmonologist (state 1), cardiologist (state 2), internist (state 3), other MD specialist (state 4), GP (state 5), respiratory therapists in CLSC (state 6), nurse in CLSC (state 7), and a state for “no healthcare utilization” (state 8). Then, for each patient, we defined three CT sequences (one for each dimension) consisting of 52 states (from 1 to 7 for the where dimension, from 1 to 4 for the why dimension, and from 1 to 8 for the which dimension), one for each time unit of follow-up after index date (52 weeks). In case a patient has several states during the same unit of time (e.g. urgent hospitalization and consultation in CLSC during the same week), each state was given a priority (in the same order as listed above, e.g. state 1 hospital (urgent) has the priority over all other states).

In step 3, we chose optimal matching, a method largely used in social sciences [32, 40‐42] to measure the distance (or dissimilarity) between patients’ CT sequences for each of the three dimensions. For each pair of CT sequences, this method measures the minimal cost of transforming one sequence into the other. This minimal cost constitutes the “distance” between two CT sequences (or patients). In optimal matching, only three kinds of modifications are allowed: substitution, deletion, or insertion. For the primary analysis, we chose a deletion/insertion cost of 1 and a substitution-cost matrix based on the estimated transition rates, the rational being to set a high cost when changes between two states are seldom observed and lower cost when they are frequent [32]. Three distance matrices, one for each dimension, were created at the end of this step. We explored another substitution-cost matrix in sensitivity analyses.

In step 4, we calculated a pooled distance matrix between CT sequences as the sum of the three dimension-specific distances to propose a unique typology of CTs that accounts for all three dimensions. Hence, two CT sequences that are similar on all three dimensions will have a small sum of dimension-specific distances. Conversely, two CT sequences dissimilar on all three dimensions will have a large sum of dimension-specific distances.

Then, based on this pooled distance matrix, we performed a hierarchical cluster analysis (HCA) in step 5 to classify similar CTs (or patients with similar CTs) [40], i.e., patients with similar sum of dimension-specific distances were classified in the same group. In HCA, each patient starts in his own cluster, and then pairs of clusters are merged as one moves up the hierarchy, until all patients are combined in a unique group. The Ward’s linkage criterion, also largely used in social sciences with dissimilarity measures [25, 40], was chosen to find the pair of clusters that lead to the minimum increase in total within-cluster variance after merging. The choice of the optimal number of groups or clusters was guided on statistical criteria (sum of squares or inertia).

Finally, to interpret and visualize the CT types (step 6), we benefited from the various visual representations offered by SSA. Among them, State Distribution Plots show the distribution of states for each time unit point, and Sequence Index Plots use line segments to show how individuals move from one state to another over time, each line representing an individual’s CT sequence. Once each patient was classified in a specific cluster (with similar CTs), we compared covariables between groups using the usual descriptive statistics (Chi-2 test, t-test, Kruskal Wallis test).

Sensitivity analyses: Since optimal matching offers the opportunity to assign different “costs” to different modifications, we also explored the impact of using different costs matrices. This was done using weeks as the time unit. We also explored other different time units (days and months).

SSA was performed using the TraMineR package in R [43]. All other analyses were performed using SAS 9.4.

To see if the results were sensitive to the matrix of substitution cost used in the SSA, we repeated the analysis with a matrix of constant costs. As in the main analysis with transition rates costs, three high healthcare users’ groups emerged, one constituted of patients with respiratory consultations, one with CVD consultations, and one with consultations for other reasons (Supplementary Figure 1). To see if the results were sensitive to the time unit used in the SSA, we reran the analyses by months and days instead of weeks. Results by months were similar to those obtained in the SSA analysis by weeks, with three high healthcare users groups: respiratory, CVD, and other reasons (Supplementary Figure 2). However, results by days produced a different CT typology (Supplementary Figures 3 and 4): two “low” healthcare utilization patterns in day-based CT Types 1 and 2 shared by almost 90% of patients, and two tiny clusters of very high healthcare utilization patterns in day-based CT Types 4 and 5, shared by 1.4 and 1.3% of patients respectively. The day-based CT Type 4 displays mainly urgent hospitalizations for respiratory causes in the first half of the follow-up year, while the CT Type 5 displays mainly non-urgent hospitalizations and use of CLSC services, related mostly to conditions other than respiratory. Unlike week and month-based typology of CTs, care use for CVD diagnosis does not emerge clearly in day-based typology.

Patients with a first hospitalization for COPD had been grouped into five structurally distinct types of one-year ‘post-acute’ Care Trajectories. At first glance, the emerging CT typology is somewhat reassuring. First, the intensity of care use seems to be related more to the initial patients’ demographic factors and the severity of their condition, rather than prominent socioeconomic inequalities. Second, results revealed that the most common CT type is the Type 1 “low healthcare utilization” shared by more than 50% of patients, followed by the Type 2 “moderate healthcare utilization” shared by nearly 30% of patients. Third, as expected, the most discriminative factors associated with “high healthcare utilization” patterns, revealed by CT Types 3 to 5, were related to increased age, comorbidities, and the illness condition of patients at the index hospitalization, measured by length of stay, intensity of resources used and clinical severity index. Fourth, no manifest or considerable socioeconomic inequalities in patterns of care use were observed, except for patients of the CT type 4, which came more frequently from rural areas, and were more frequently admissible to the public prescription drug insurance plan with last-resort financial assistance. Finally, the emerging multidimensional typology of CTs shows a connection between the “why”, “which” and “where” dimensions, which seems consistent with the real-world practice. For example, state distribution plots of CTs show that among the “high healthcare utilization” groups, the CT Type 3, the “respiratory group”, had more frequent urgent or “unplanned” rehospitalizations for respiratory diseases, more noticeable in the first half of the follow-up year, consulted mostly pulmonologists, at a hospital or an outpatient clinic, and received health services from respiratory therapists at ambulatory care settings. On the other hand, the CT Type 4, the “CVD group”, had more frequent emergency visits, consulted more cardiologists at ED, hospital or outpatient clinics, and received more health services from nurses at ambulatory care setting. Non-urgent hospitalizations were mostly observed in patients of the CT Type 5, the “other conditions group”, which had more consultations with other categories of specialists at outpatient clinics. Results also show that outpatient and primary care were largely used by patients with high healthcare utilization, corresponding to patients with a more severe initial illness condition, for which post-acute rehabilitation care services delivered at CLSC settings are the most crucial [6, 44].

In this study, almost 60% of patients had at least one readmission (urgent or non-urgent) in the year following the index hospitalization, 36.1% were readmitted to a hospital for a respiratory cause, and 11.6% for a CVD cause. Although one-year post-acute outcomes after a first hospitalization for COPD are rarely reported, these results are generally in agreement with well-known outcomes for COPD patients: readmissions for respiratory conditions are frequent and comorbidities are common, especially CVD [6, 44].

SSA delivered relatively consistent results when imposing constant costs of transitions in distance measures instead of measures based on the inherent data transition rates. Nonetheless, results should be taken with caution, since the choice of time granularity could affect state sequences results at different degrees. The typology of CTs is broadly similar when using weeks or months as the time unit, although the size of clusters differs. However, the emerging typology of CTs differs notably when using the day as the time unit (Supplementary Figures 3 and 4), where almost 90% of patients shared the two “lowest” healthcare utilization patterns day-based CT types 1 and 2, compared to 81% in week-based CTs. The day granularity also reveals two tiny atypical or extreme clusters of very high healthcare utilization patterns: a first one with recurrent urgent hospitalizations in the first half of the follow-up year for respiratory diagnoses (day-based CT Type 4), and a second one with non-urgent hospitalizations and nurse services use in ambulatory care, associated mostly to CVD and other conditions (day-based CT Type 5). Another interesting point about the time granularity effect is that primary care encounters emerge more intensely in month and week-based sequences, compared to sequences with day time granularities. Since many categorical states of healthcare utilization are not mutually exclusive, the priorities of each state needed to be established a priori. As a result, hospital and emergency events arose over ambulatory care visits, and specialists’ consultations arose over primary care providers, regardless of the time granularity. However, week-based sequences allow the emergence of lower-priority states, since the probability of primary care and non-specialist encounters increases when using larger time units. Taking this into consideration, rather than a limit, a change in time granularity offers different complementary perspectives of care use patterns. For instance, some “alarming” day-based CT types arose, which may require additional investigation. In particular, the day-based CT type 4 reveals extreme lengths of stay both in urgent hospitalizations and urgent hospital readmissions for respiratory cause, despite an initial severity index broadly similar to most of the other groups (data not shown).

State distribution plots offer a useful visualization of the CT typology, but as expected, such large sets of complex sequences could be problematic to display by sequence index plots due to overplotting, although certain graph simplification and smoothing techniques could be applied [28, 30]. Nonetheless, displaying sequences at the individual level, the sequence index plot of day-based CTs (Supplementary Figure 3), while “noisy”, offers a view of point events, interval events and transitions, such as physician consultations and healthcare services at primary care settings, hospital length of stay, as well as transitions and gaps, which are undetectable in state distribution sequences plots. Also, graphical representations of some CT attributes (total number of days in each care settings and hospital readmission) and characteristics of the study cohort for each CT type provide valuable complementary information (Table 2, Figs. 6 and 7).

Although evidence-based guidelines regarding interventions in primary care such as early pulmonary rehabilitation and counseling could improve patients’ quality of life, exercise tolerance and dyspnea, the effectiveness of such interventions in preventing rehospitalizations for respiratory causes remains unclear [45‐48]. However, for each CT type, the relatively large number of consultations at outpatient, primary care and community-based clinics suggests that access to primary care is adequate. The patterns of healthcare utilization described by our approach could contribute to a better evaluation of the impact of new organizational models of healthcare services according to the patient’s condition and concomitant diseases.

This study has several notable strengths. First, it uses an exhaustive longitudinal dataset of patients hospitalized for COPD in Quebec. This dataset used linked medico-administrative data from multiple sources which provides a comprehensive picture of healthcare services utilization at both inpatient and outpatient settings, including community-based healthcare services provided in ambulatory care (CLSCs). Moreover, the multidimensional approach of care trajectories allowed the possibility to include as much as 19 states, since these states are partitioned into three distinct sequence-dimensions, thus reducing the complexity of each sequence and avoiding the “overplotting” issue [28, 30, 31, 33]. To our knowledge, this is the first study which proposes a comprehensive perspective of care trajectories, allowing a more intuitive and straightforward examination of the most common shared patterns of care use as a whole.

Results need to be taken with caution nonetheless, since the CT typology emerged from the analysis of a specific cohort of patients with a first hospitalization for COPD (infrequent exacerbations). Also, administrative data has inherent limitations: some important variables related to patients’ individual and clinical attributes, such as severity of COPD, body mass index, smoking status, as well as social or caregiver support, which may considerably affect patients’ health condition and care use, are not routinely collected. There are also limitations related to the choice of the clustering method used in the SSA approach. For example, one possible problem in clustering analysis is that different algorithms may lead to different results. However, the graphical inspection of CT sequences (sequence index plots) in clusters (types) can help to evaluate the quality of a partition. Although not presented in this paper, other techniques exist to help the visualization of complex sequence data, for example by analyzing dissimilarities using multi-dimensional scaling (MDS) and smoothing techniques [28, 49].