In this multimorbidity cluster of physical and mental health conditions, disease sequencing had an important impact on all-cause mortality. Development of diabetes, psychosis, and CHF as an ordered sequence had a poorer prognosis compared to the development of the same conditions in different orders. Comorbid psychosis and CHF as a combination of two diseases had a poorer prognosis compared to combinations of psychosis and diabetes, or CHF and diabetes. CHF had the worst prognosis as a single condition where development of CHF at age 50 was associated with a reduction in life expectancy of 12.38 years compared to the general population without a diagnosis of psychosis, diabetes, or CHF during follow-up, and with the same characteristics in terms of age at diagnosis, deprivation status, and gender.
Development of multimorbidity associated with diabetes, psychosis and CHF as an ordered sequence at age 50 resulted in 13.23 years loss in life expectancy, compared to the general population. The loss in life expectancy for combinations of psychosis and CHF at age 50 (12.95 and 13.45 years), and CHF alone (12.38 years) were comparable to diabetes, psychosis and CHF (13.23 years). These findings would suggest that the key components resulting in an increased loss in life expectancy were attributable to the development of CHF and/or CHF and psychosis in combination. We further identified that identification and therapeutic targets for CHF and psychosis may be beneficial in the first 5 years following the diagnosis of the former condition.
Individuals developed CHF as the first disease later in life, with a median age of 78 years, resulting in a higher proportion of deaths from health states with CHF as the index condition. However, having adjusted for age at entry into the health state, we observed a lower rate of deaths for combinations of three diseases where CHF was the first condition in the sequence, compared to combinations with either diabetes or psychosis as the index condition. This may be a consequence of a survivor effect where individuals who developed subsequent disease after CHF, and CHF and psychosis in combination, must have survived long enough to accrue further disease, and thus these individuals may have a less severe type of CHF, resulting in improved survival.
Diabetes as a single condition had an improved life expectancy compared to the general population including otherwise healthy and otherwise diseased individuals. For example, a 50 year old individual diagnosed with diabetes could expect to gain an average of 0.19 years compared to the general population which may have other life shortening conditions such as cancer. Furthermore, combinations of psychosis and diabetes had an improved life expectancy compared to psychosis alone. For example, a 50 year old individual diagnosed with a combination of psychosis and diabetes could expect to gain an average of 1.28 to 1.74 years compared to psychosis alone. This may be a result of increased management of diabetes and cardiovascular disease in cardiometabolic multimorbidity.20 However, compared to diabetes as a single condition, the development of psychosis in combination with diabetes at age 50 resulted in a mean loss of life expectancy of 0.74 to 1.2 years. These findings are congruent with a recent study evaluating the impact of multimorbidity in diabetes and severe mental illness.21 In this study, individuals with severe mental illness appeared to have an increased all-cause, and cardiovascular disease-specific mortality rate, compared to individuals with diabetes alone. Han, et al. 21 conclude that increased mortality rates may be attributable to the under diagnosis of cardiovascular disease and resultant delays in treatment in individuals with diabetes and severe mental illness. These findings, amongst others reported in cardiometabolic multimorbidity22 and physical-mental health multimorbidity23,24, further highlight the need to identify screening opportunities and/or therapeutic targets in multimorbid populations.
The use of multi-state models offers a flexible framework in which to assess the impact of disease trajectories in multimorbidity, and therefore give rise to the identification of potential screening opportunities, therapeutic targets and risk factors, together with associated implications for future research prioritisation, study design and policy.24
We demonstrate the application of this modelling framework to physical-mental health multimorbidity using population-scale, individual-level, anonymised, linked, routinely collected EHR data sources. Such an approach could be applied to any situation with individuals accumulating conditions over time, for example, in long COVID-19. Application of these models could be used to help inform patients, clinicians, and healthcare decision-makers on the appropriate identification and management of patient care, and thus have the potential to improve patient outcomes.
Transition modelling approaches have previously been used in multimorbidity research,25–33 with several studies specifically looking at state transition modelling.28,29,31,32 Two studies have evaluated trajectories of multimorbidity patterns in terms of transitions between clusters of multimorbidity over time.34,35 However, to our knowledge, none of these approaches have explored the ordering and sequencing of disease occurrence within multimorbidity clusters. Two studies have evaluated the ordering of diseases, 27,36 however, these studies focussed on summary statistics and visual representation of disease trajectories, and did not evaluate the impact of disease sequencing on patient outcomes. A systematic review and comparison of approaches used to model trajectories of diseases in multimorbidity research has also been reported.16
The use of multi-state modelling approaches can have limitations in multimorbidity research. Increasing the number of clusters, and/or the number of diseases within clusters, will exponentially increase the complexity of the model, which may lead to computational difficulties, when applied to population-scale data. For this reason, the number and type of conditions included in the analysis needs careful consideration to ensure that we do not compromise the interpretability of the results by pooling conditions, or increase the granularity of conditions to the extent that we compromise computational efficiency and/or sample size.
A potential limitation of using routinely collected EHR data is the appropriate coding of diagnoses, the time the condition is first recorded and the resulting classification of multimorbidity. Busija et al 37 discusses this potential issue in terms of clustering of diseases in multimorbidity research and the need for harmonised data and definitions. Health Data Research UK (HDRUK) has taken steps to potentially improve the appropriate classification and harmonisation of human disease via the development of a phenotype library.38 However, even with complete harmonisation across data resources, it is widely known that diagnosis codes in EHR data are not always accurate, and there could be a delay from the development of the condition of interest and the resultant recording of the diagnosis that is unlikely to be captured. External factors affecting healthcare utilisation may exacerbate this potential issue, for example, during the recent COVID-19 pandemic. In our analysis, we excluded 16.3% of the WMC population who were registered with General Practices that did not provide data to the SAIL Databank, and thus we were unable to access their complete diagnostic history in terms of diagnoses captured within primary care settings. Furthermore, we were unable to capture diagnoses prior to 1st January 2000 due to poor quality of reporting. As a result, the otherwise healthy/otherwise diseased population may contain individuals with a diagnosis of psychosis, diabetes, and CHF prior to cohort inception, amongst other conditions, and thus provide a reference group for general all-cause mortality. A further potential limitation of using routinely collected, electronic health record data is the difficulty in capturing severity of disease and transient morbidity, such as capturing periods of psychotic episodes for example. Further linkage of data to prescribing and dispensing data at the individual level may help to validate clinical coding and provide an improved understanding of severity of disease and transient health conditions.39,40 Subsequently, modelling techniques could be easily developed to incorporate the transient nature of these conditions in the analysis.
In this paper, we focus on all-cause mortality as the primary outcome of interest. Using routinely collected EHR data sources, modelling approaches could be extended to also assess healthcare utilisation and resource use. Increasingly, trials are being conducted in multimorbid populations, with a focus on alternative outcomes of importance to patients, clinicians, and healthcare decision-makers, such as quality of life measures.41 As more data becomes available in multimorbid populations, further work could extend this modelling approach to assess the impact of trajectories of disease on patient-centred outcomes such as quality of life, independent living, or patient reported outcome measures. Extensions to the modelling approach using longitudinal biomarker data could further enable identification of sub-populations at particular increased risk of both developing specific multimorbid conditions and/or death.42
Patient and public involvement in health data science is still relatively new. From our discussions with public contributors we identified clear potential for future public input into defining multimorbidities, interpreting trajectories of disease and what they mean for public contributors and communities, reviewing and critiquing the content of datasets and the future curation of what variables are included in datasets. Moving from passive supplier of data to active collaborator in health data science offers much promise for the future and will strengthen our understanding of the community validity of such work.
In conclusion, we found that the sequence and timing of disease onset in a trio of long-term conditions associated with high mortality had an important and complex impact on all-cause mortality. Multi-state models provide a flexible framework to analyse trajectories of disease development and their associated impact on patient outcomes, and thus allow identification of potential risk factors, screening opportunities, and targeted therapies for healthcare policy and decision-making.