Community prevalence of multimorbidity in Aotearoa New Zealand based on linked general practice and hospitalisation records

Multimorbidity, the presence of two or more co-existent long-term conditions, is associated with an increased risk of hospitalisation, greater use of unplanned care and higher mortality.^1–3 Life expectancy in Aotearoa New Zealand has increased by over 10 years in the last 50 years, but 20–30% of these years are lived in poor health due to multimorbidity.¹ Common long-term conditions (LTCs) contributing to this burden include diabetes, cancers, cardiovascular diseases, respiratory diseases, mental illness, chronic kidney disease and dementia. Many co-occur and comprise a major part of the primary care workload—almost 70% of general practice (GP) visits.⁴ Multimorbidity is a stronger predictor of healthcare utilisation and costs than age or sex⁵ and is the strongest predictor of adverse drug events.^6,7

In 2017, Sarfati and Stanley developed a New Zealand general population multimorbidity index “M3” utilising 61 LTC categories identified via ICD-10-AM (International Statistical Classification of Diseases and Related Health Problems, 10th Revision, Australian Modification) codes from hospital records.⁸ They also developed the P3 index containing 30 condition categories from pharmaceutical dispensing records.² Both are weighted indices representing the association between LTCs and 1-year mortality^8,9 and have been used to describe the prevalence of multimorbidity.² However, both the M3 and P3 indices have limitations for estimating the prevalence of multimorbidity. M3 was derived solely from hospital administrative data and cannot capture LTCs that rarely lead to admission or are not recorded if they are unrelated to the primary cause of admission. The P3 index, based on community pharmacy dispensing, is more closely aligned to what might be expected in primary care diagnosis data, but many LTCs cannot be determined from these records because medications may have multiple indications (and were not included in the P3 index) or treatment may not include prescribed medication.²

There remains uncertainty about the extent to which hospital data alone can capture the full burden of multimorbidity in the community. Measurement of multimorbidity is critical for health service planning, care delivery, hospital and primary care funding and understanding inequities in healthcare utilisation and outcomes.^4,10 With an ageing population, the complexity and need for care in the community is increasing, but funding and planning decisions are often based on hospitalisation data alone. Research suggests that this primary care information gap will under-estimate the prevalence of multimorbidity and clinical priority conditions driving health inequity and life expectancy gaps, such as those experienced by Māori and Pacific people.¹¹ Multimorbidity has been reported to be more common among Māori and Pacific people,² and so it is important to assess whether different data sources affect observed (in)equity patterns. In New Zealand, 24% of patients with ischaemic cardiovascular disease (CVD) diagnosed by their general practitioner were not recorded as having a prior CVD-related hospital admission.¹² In the United Kingdom’s QRISK3 equation, 23% of angina and 55% of transient ischaemic attack (TIA) outcomes came only from GP records.¹³

Using both GP and hospital records, we aimed to quantify the community prevalence of multimorbidity for the 61 LTCs included in the M3 index and determine the extent to which adding GP data affects estimation of its “true” burden.

Methods

This cross-sectional study of linked data from GP and hospital records is described according to the STrengthening the Reporting of OBservational studies in Epidemiology (STROBE) checklist for the reporting of observational studies.¹⁴

Study participants, data sources and linkage

ProCare is a primary health organisation (PHO) in Auckland, and in 2014 it included over 170 member practices, serving approximately 820,000 enrolled patients, covering 51% of the Auckland population (17% of New Zealand population). Adults ≥18 years enrolled in a ProCare practice at the study index date of 1 January 2014 and who had GP electronic health records (EHRs) were eligible for inclusion in this study (Appendix Figure 1). With practice approval, ProCare PHO extracts and centrally stores EHR data for population health improvement and approved research. GP-recorded LTC data from practices using patient management systems that employed Read coding were obtained from this repository.

ICD-10-AM–coded hospitalisation data were obtained from the National Minimum Data Set (NMDS), which holds records of all publicly and most privately funded hospital discharges in New Zealand since 1988.¹⁵ These datasets (from inception to the end of 2013) were linked via an encrypted unique health identifier (the National Health Index [NHI]) assigned to everyone upon their first interaction with the New Zealand health system. Age, gender and ethnicity as of 1 January 2014 were obtained from the NHI collection.

Ethnicity was categorised into six broad groups of Māori, Pacific, Indian, Chinese, European and Other ethnicities (representing Middle Eastern/Latin American/African/Other/Other Asian) using a modified total ethnicity approach, with individuals identifying with multiple ethnicities counted in each group they identified with. The European group was defined as mutually exclusive for comparison.

Area-based socio-economic deprivation (New Zealand Index of Deprivation 2013 [NZDep2013]) was assigned to each patient using meshblock codes from the national PHO enrolment database.¹⁶ NZDep2013 was analysed in quintiles from 1 (least deprived) to 5 (most deprived).

Multimorbidity definition

Multimorbidity was defined as having two or more of 61 ICD-10-AM–coded condition categories in the M3 index.² From the GP EHRs, any recorded diagnosis matching one of the 61 M3 index codes was extracted, whether it had been entered using Read codes, Read terms or free text. Free-text entries were classified using methods previously described.¹⁷ In brief, a quarter of primary care classifications contained either unrecognised Read codes or were free-text only. To improve accuracy, clinical raters independently identified condition-related text, including synonyms, negation terms and common spelling errors in randomised samples. Clinician inter-rater reliability was high (kappa≥0.9). Over 10 iterations, text classification performance improved, achieving average sensitivity, specificity and positive predictive value of 88%, 99% and 95%, respectively.

To avoid double-counting, primary cancers (e.g., breast cancer, malignant melanoma) were excluded if a metastatic cancer diagnosis was also present. This rule was applied across both GP and hospital data sources.

Statistical analysis

The cohort was categorised based on socio-demographic factors, with age and ethnic groups selected to align with those used by Stanley et al.² for comparability. Prevalence estimates were reported with 95% confidence intervals (95% CIs) calculated using the Wilson method for binomial distributions.

Agreement between GP and hospital discharge diagnoses was assessed with Cohen’s kappa statistic for each of the 61 conditions, categorising the extent of agreement as: slight (0.01–0.20), fair (0.21–0.40), moderate (0.41–0.60), substantial (0.61–0.80) and almost perfect (0.81–1.00). Missing data were minimal and supplemented using linked national datasets; no imputation was performed. All analyses were conducted using R software version 4.3.1 (R Core Team, 2023, R Foundation for Statistical Computing, Vienna, Austria).

Ethics approval

Ethical approval for this study was granted by the Health and Disability Ethics Committees (HDEC) under the VIEW programme (MEC/07/19/AM25) and subsequently under the VAREANZ programme (2022 EXP 13442 and 2023 EXP 18564), with annual reapproval since (2025 PR 13442). The need for informed participant consent was formally waived by the Northern B HDEC (2022 EXP 13442). The study was also approved by the ProCare Clinical Governance committee for the secondary use of de-identified patient data.

Results

A total of 623,475 individuals aged ≥18 years were identified. GP data for 374,071 (60.0%) were available from practices using “Read coding” patient management systems. Of these, 251,058 patients also had hospital records. Among the included patients, over 7 million entries in GP medical history/classification fields and over 1 million unique hospital discharge events were identified (Appendix Figure 1).

Median age was 46 years (interquartile interval [IQI] 33–58) and 55% were women (Table 1). There were 6,802 (1.8%) people recorded as having more than one non-European ethnicity resulting in a total ethnicity count of 379,405. The total ethnicity composition was 8% Māori, 11% Pacific people, 6% Indian, 7% Chinese, 7% Other ethnicities and 62% sole European. Of the cohort, 32% were living in communities with the highest levels of deprivation (NZDep2013 Q4 and Q5).

The prevalence of multimorbidity using both data sources was 30.8% (95% CI: 30.7–31.0) with 14.1% (14.0–14.2) identified using only hospital records and 16.7% (16.6–16.8) exclusively from GP records (Table 1).

View Table 1–2, Figure 1–4.

Figure 1 shows the prevalence of multimorbidity and denotes individuals with three, four, five and six or more long-term conditions (MM3+, MM4+, MM5+, MM6+). Within each of these categories, the prevalence of more complex multimorbidity is at least doubled after taking GP records into account.

Prevalence by age group and gender

Multimorbidity increased markedly with age, affecting over 60% of men and women by age 65–74 years from combined hospital and GP records (Figure 2, Appendix Tables 1 and 2). Across all age groups, the under-estimate of multimorbidity using only hospitalisation data was greater for women—from 7% under-estimated at age 18–24 years to 36% at age 65–74 years. The range for men was 6% at age 18–24 years to 29% at age 65–74 years.

Prevalence by age group, total ethnicity and deprivation quintile

By age 35–44 years, 28% Māori, 25% Pacific people, 20% Indian and 18% Europeans had multimorbidity (Figure 3, Appendix Table 2). By 65–74 years, the prevalence of multimorbidity for the same groups was 76%, 74%, 74% and 61%, respectively. Overall, Māori and Pacific people are more likely to be identified with multimorbidity from hospitalisation data, with European and Chinese more likely identified using solely GP data. The relative under-estimate of prevalence using hospital data alone was lowest for Chinese people until the age of 55 years.

Multimorbidity was highly patterned by socio-economic deprivation (Figure 4, Appendix Table 2). By age 35–44 years, the prevalence of multimorbidity was 15% in the least deprived (NZDep2013 Q1) compared with 26% in the most deprived communities (NZDep2013 Q5). By 65–74 years, this difference had widened somewhat: 57% in NZDep2013 Q1 vs 72% in NZDep2013 Q5. Overall, the prevalence under-estimate when using hospital data alone was lower for NZDep2013 Q5 compared to NZDep2013 Q1. By all socio-demographic strata, the largest under-estimate in multimorbidity prevalence between hospital and combined hospital and GP records occurred from age 55 to 84 (Figures 2, 3, 4, Appendix Tables 1 and 2).

Table 2 lists the five most common conditions contributing to multimorbidity when identified only in GP EHRs (i.e., not identifiable within hospital records) among patients aged 55–64 years, 65–74 years and 75–84 years. Across all strata the most common M3 long-term condition categories found in GP records were uncomplicated hypertension, joint or spinal disorder, chronic pulmonary disease, diabetes (uncomplicated or complicated) and metabolic disorder. In addition, osteoporosis was in the top five additional conditions for women and inner-ear disorder for men.

The agreement between primary care and hospital discharge diagnoses using Cohen’s kappa statistic was slight for 23 conditions (<0.20 for drug abuse, inflammatory bowel disease, metastatic cancer, etc.), fair for 25 conditions (0.21–0.40 for cardiac valvular disease, lung cancer, pancreatitis, dementia, etc.), moderate for 12 conditions (0.41–0.60 for prostate, breast and colorectal cancer, cerebrovascular disease, myocardial infarction and cardiac disease, lymphomas and leukaemia, etc.) and substantial for one condition (0.61–0.80: acquired immunodeficiency syndrome [AIDS]) (Appendix Table 3).

Discussion

In this study of 374,071 primary care patients aged 18 years and over, 31% were identified as having multimorbidity: 14% from hospital records and 17% exclusively from GP EHRs. If only hospital administrative records are used, 54% of multimorbidity in the community would not be recognised. This under-estimate was greater for women than men and increased with age, being highest at 65–74 years (36% for women and 29% for men). The under-estimate was higher for European and Chinese people and for those living in the least-deprived areas. Overall, over 60% of men and women aged 65–74 had multimorbidity, and across all ages, one in 10 people (11.8%) had four or more LTCs, and one in 20 (5%) had six or more LTCs. The most common LTC categories found exclusively in GP records were uncomplicated hypertension, joint or spinal disorder, chronic pulmonary disease, diabetes, metabolic disorder, osteoporosis (women) and inner-ear disorder (men). For the majority of LTCs in the M3 index, the agreement between hospital and GP records was only slight or fair. This misclassification is likely occurring in both directions and would substantially affect prevalence estimates if relying on either data source alone. These findings highlight the potential for misclassification when relying solely on hospital data, particularly for those conditions indicated as having low agreement.

The closest comparison for an estimate of multimorbidity prevalence comes from the New Zealand population-based study of multimorbidity by Stanley et al. who used a 5-year look-back of ICD-10-AM–coded hospitalisations from the M3 index (7.9%) and 1-year look-back of national community pharmaceutical dispensing records (27.9%; 30 conditions from the P3 index).² Our estimate of 14% from ICD-10-AM–coded hospitalisations is likely to be due to our much longer look-back period to 1988.

Recent systematic reviews have reported a pooled prevalence of multimorbidity in community settings ranging from 33 to 42%.^18–21 Although prevalence varied by geographic region and country income level, the estimates in our study sit at the lower end of this range. However, individual studies included in these reviews were highly heterogeneous, varying by mean age of the study sample (e.g., only including adults over 60 years), data source (e.g., self-reported population-based surveys, primary care EHRs, aged care facilities, hospitalised patients, pharmaceutical dispensing), whether data sources were single or combined (e.g., inpatient and/or emergency department and/or GP EHRs), the definition of multimorbidity (e.g., two or more, three or more LTCs) and multimorbidity measurement tools (e.g., the number of and specific LTCs included). This substantial variability in estimates of multimorbidity prevalence is illustrated in a recent review of 16 studies of primary care EHRs which reported estimates ranging from 14% to 93.9%.²² Within this review, there were two English studies including all people aged over 18 years (the same age range as this study) with reported prevalence from GP EHRs of 21%²³ and 29%.²⁴ A further study that combined primary care EHRs with data from secondary care²⁵ (inpatient and outpatient) found an overall multimorbidity prevalence of 21.6%. This was considerably lower than our findings; however, the study included all ages and multimorbidity is rare in people less than 18 years.²⁵

We stratified our prevalence estimates by socio-demographic status. Our findings were consistent with published literature showing more than half the adult population worldwide over 60 years of age have multimorbidity,^19,26 and that multimorbidity is more prevalent in women than men,^19,26 by lower socio-economic status^2,26 and for Māori and Pacific people compared with European ethnic groups.² Of note, the under-estimate of multimorbidity prevalence using hospital records alone was more strongly apparent for women. Similarly, Stanley et al. reported that women had lower age- and sex- standardised prevalence when measured by M3 (61 hospitalised conditions) than men, but higher when measured by P3 (30-condition community pharmaceutical dispensing).²

The slight or fair agreement between hospital and GP records for 48 out of 61 conditions listed in the M3 index is likely multifactorial. For example, hospital diagnoses may not reach the intended practice or be fully updated in GP electronic health records. If hospitalised, many LTCs are not coded if not the major or contributing reason for hospitalisation. For people aged 55–84 years with multimorbidity and without previous hospital admissions, the most common codes were for cardiometabolic disease, chronic pulmonary disease and joint or spinal disease. While it was outside the scope of this study to characterise multimorbidity clusters, common patterns reported in systematic reviews include at least two or more clusters relating to cardiovascular disease, musculoskeletal, respiratory and mental health conditions (primarily characterised by depression and anxiety disorders).^19,22,26

Strengths of this study were the large population sample and the ability to individually link patient data from GP EHRs and hospital admissions. A further strength was the ability to utilise unstructured text as well as structured data within GP EHRs to accurately identify LTCs.¹⁷ These multimorbidity estimates are likely to be generalisable to New Zealand adults as the age and sex distribution of our cohort is similar to that of the national adult population.²

The main study limitation is that we report multimorbidity prevalence in a 2014 cohort; thus, it is over 10 years old. However, we purposefully chose the 2014 cohort to enable us to investigate the impact of multimorbidity on 5-year CVD risk (research in progress) and implications for risk management. To assist primary care decision making we will need to understand whether measures of multimorbidity are associated with fatal and non-fatal CVD events and if so, which measure or measures are the most practical to be used in a primary care setting.

A further limitation includes the inability to accurately map, at the time of data extraction, to SNOMED codes for the M3 ICD-10-AM conditions, resulting in only 60% of the ProCare-enrolled adult population being included. However, while prevalence estimates may have differed somewhat with the addition of practices using patient management systems utilising SNOMED classification, our aim was to investigate the impact of including both primary and secondary care datasets.

Implications and future research

This study has highlighted the previously under-recognised extent and complexity of multimorbidity in the community and exposes the “unsung” GP workload, which will have increased as this cohort is now over 10 years older. Funding and planning decisions that are based on routine hospital administrative data, while capturing the more severe end of the disease spectrum, will not track emerging trends in LTC prevalence in the community, nor the needs for primary healthcare services. The next steps are to build on our methodology (available on request) to include SNOMED-coded primary care LTC information to more accurately quantify multimorbidity prevalence.

Conclusion

Multimorbidity prevalence is substantially under-estimated if relying solely on hospital discharge data, capturing under half of the total multimorbidity burden found when GP-level data are included. Without the combination of primary care and hospital data, we will continue to under-rate the complexity of care and health system resources required to support the needs of this important and growing group of patients.

View Appendix.