NMI - Nordic Multimorbidity Index
Weighted mortality-risk score from Nordic registers
NMI (Nordic Multimorbidity Index) is a weighted score built from Nordic hospital diagnoses and filled prescriptions. It combines 50 mortality predictors - 29 diagnosis-based and 21 prescription-based - into a single continuous score.
Article: Kristensen et al. Clinical Epidemiology 2022 - doi:10.2147/CLEP.S353398
Important: NMI is a mortality index, not a measure of comorbidity burden. NMI was developed and validated to predict 5-year mortality (Kristensen et al. 2022, ~425,000 Danish residents aged ≥40), and its predictors were selected for how much they added to the mortality model - not to cover disease burden. As a result:
- Several common chronic diseases are not included in the model (e.g. hypertension and type 1 diabetes).
- Some items are mortality proxies, not diseases in their own right (e.g. laxatives, iron).
- Some weights are negative (e.g. statins -3, ACE/ARB inhibitors -2), which only makes sense in a mortality model, not in a measure of disease burden.
If your goal is comorbidity burden, NMI is the wrong tool - use a hospital-ICD-based multimorbidity index instead. A dedicated one (ICD-10 + ATC codes) is in development; see Comorbidity for the general approach. Use NMI only when your goal is specifically mortality risk.
What NMI gives you
A continuous mortality-risk score per person, validated to predict 5-year mortality in Nordic populations (and stronger than the Charlson and Elixhauser indices for that outcome). Use it as:
- A covariate summarising mortality-associated illness in a Cox or logistic model.
- A summary variable for mortality risk in Table 1.
It is not a clean count of chronic diseases - for comorbidity burden, see the warning above.
When to use it (and when not)
Use NMI when your question is mortality risk: adjusting for mortality-associated illness burden in a survival or logistic model, or wanting a validated 5-year-mortality predictor.
Do not use NMI to measure comorbidity burden (a count or index of chronic disease) - it omits common conditions by design (see above). Use a hospital-ICD multimorbidity index for that.
Watch for over-adjustment: your own outcome must not be one of the score’s predictors. NMI handles dementia studies, for example, by dropping the dementia predictor (dx_F00_G30 and rx_N06D), since dementia would otherwise be part of the score you adjust for.
R implementation
There is no official R package (SDU publishes a Stata implementation). The R version below runs in two steps - extract, then compute. These code examples are based on Pernille Errebo Tonnesen’s translation of the SDU Stata code to R.
1. Extract diagnoses and prescriptions
What NMI needs, and over which period:
- Diagnoses, a 5-year lookback before index, from all of: LPR2 somatic and psychiatric (up to March 2019) and LPR3 (from March 2019), across all diagnosis types. A somatic-only pull (e.g.
lpr_admalone) misses the psychiatric predictors (F00-03, F10, F17). - Prescriptions (ATC), a 6-month lookback before index, from LMDB - only that window, not the whole register.
Diagnoses - You can use the two approaches from Extract from LPR: either extract all diagnoses and ATC codes for the whole cohort (not recommended - time-, RAM- and space-heavy, especially for a large cohort), or filter to only the codes NMI needs before collect(), so the saved .rds files stay small.
You can also use the get_lpr_diagnoses() pattern from Extract from LPR as in the example here:
See the code: extract NMI diagnoses (LPR2 + LPR3)
#=============================================================
# 1. EXTRACT diagnoses and prescriptions
#=============================================================
library(dplyr)
cohort <- readRDS("datasets/full_cohort.rds") # your study population from Phase 10 (pnr + index_date)
#-------------------------------------------------------------
# NMI diagnosis codes (3-char ICD, no D-prefix).
# paste0("C", 76:80) makes "C76".."C80" - glues the letter onto each number
# in the range, so we don't type every code by hand.
#-------------------------------------------------------------
nmi_icd <- c(
"B18", # chronic viral hepatitis
"C34", # lung cancer
"C50", # breast cancer
"C61", # prostate cancer
"C67", # bladder cancer
# NMI predictor "tumour of brain or meninges" (only the subcodes below score, see the computation):
"C70", # cancer of the meninges
"C71", # brain cancer
"C75", # cancer of other endocrine glands (only C751-753 score)
"D32", # benign meningeal tumour
"D33", # benign brain/nervous-system tumour (only D330-332)
"D35", # benign endocrine-gland tumour (only D352-354)
"D42", # meningeal tumour, uncertain behaviour
"D43", # brain/nervous-system tumour, uncertain behaviour (only D430-432)
"D44", # endocrine-gland tumour, uncertain behaviour (only D443-445)
paste0("C", 76:80), # secondary/metastatic and ill-defined cancer (C76-C80)
paste0("C", 91:95), # leukaemia (C91-C95)
paste0("D", 50:64), # anaemias (D50-D64)
"E11", # type 2 diabetes
"E86", # volume depletion (dehydration)
paste0("F0", 0:3), # dementia (F00-F03, scores together with G30)
"G30", # Alzheimer's disease (dementia)
"F10", # alcohol use disorder
"F17", # tobacco/nicotine use disorder
paste0("G", 20:22), # Parkinson's disease and parkinsonism (G20-G22)
"G35", # multiple sclerosis
"G40", # epilepsy (G40-G41)
"G41", # status epilepticus (epilepsy predictor)
# NMI predictor "aortic and mitral valve disease" (I05/I06/I34/I35):
"I05", # rheumatic mitral valve disease
"I06", # rheumatic aortic valve disease
"I34", # non-rheumatic mitral valve disorder
"I35", # non-rheumatic aortic valve disorder
# NMI predictor "heart failure" (only subcodes I110/I13[02]/I42[06789]/I50 score):
"I11", # hypertensive heart disease (only I110)
"I13", # hypertensive heart and kidney disease (only I130/I132)
"I42", # cardiomyopathy (only I420/I426-429)
"I50", # heart failure
paste0("I", 60:69), # cerebrovascular disease / stroke (I60-I69)
# NMI predictor "atherosclerosis / peripheral arterial disease" (I70/I73/I74/I77):
"I70", # atherosclerosis
"I73", # peripheral vascular disease
"I74", # arterial embolism and thrombosis
"I77", # other arterial disorders
"I71", # aortic aneurysm and dissection (I71-I72)
"I72", # other aneurysm (aneurysm predictor)
paste0("J1", 2:8), # pneumonia (J12-J18)
# NMI predictor "chronic lower respiratory disease" (J41-44, J47, J961/J969):
"J41", # simple chronic bronchitis
"J42", # unspecified chronic bronchitis
"J43", # emphysema
"J44", # COPD (chronic obstructive pulmonary disease)
"J47", # bronchiectasis
"J96", # respiratory failure (only J961/J969)
"J84", # other interstitial lung disease
paste0("K0", c(2, 3, 4, 5, 6, 8)), # dental and periodontal disease (K02-K06, K08)
# NMI predictor "alcoholic liver disease" (K70/K72/K74/K766/K767):
"K70", # alcoholic liver disease
"K72", # hepatic failure
"K74", # fibrosis and cirrhosis of the liver
"K76", # other liver disease (only K766/K767)
"L89", # pressure ulcer
"N18", # chronic kidney disease (N18-N19)
"N19" # unspecified kidney failure
)
# You can now extract the ICD and ATC codes using approach 1 or 2, or the get_lpr_diagnoses() helper from Phase 9 (Extract from LPR). With the latter, the code looks like this:
nmi_diagnoses <- get_lpr_diagnoses(
pnr_vector = unique(cohort$pnr), # all the cohort's pnr
icd_codes = nmi_icd # only NMI codes are collected
)
saveRDS(nmi_diagnoses, "datasets/nmi_diagnoses.rds") # columns: pnr, date_contact, c_diag, c_diagtype, icd3Prescriptions - extract LMDB for the cohort, keeping only NMI’s ATC classes. read_register() requires fastreg set up with the path to your registers - see Phase 4 if you did not convert them from SAS yourself:
See the code: extract NMI prescriptions (LMDB)
library(fastreg) # read_register()
library(arrow) # open_dataset() - fallback without fastreg
library(dplyr)
pnr_vec <- unique(cohort$pnr)
#-------------------------------------------------------------
# NMI drug classes (ATC) as one regex. Each class:
# A06A laxatives
# A07DA antipropulsives (antidiarrhoeals)
# A10A insulins and analogues
# B01AC antiplatelet drugs
# B03A iron preparations
# C01AA digitalis glycosides (digoxin)
# C03C loop (high-ceiling) diuretics
# C03EB loop + potassium-sparing diuretics
# C03DA aldosterone antagonists (e.g. spironolactone)
# C09C/C09D angiotensin II receptor blockers (ARBs), plain and combinations
# C10AA statins
# H02AB systemic glucocorticoids
# J01C penicillins (beta-lactam antibacterials)
# N02A opioids
# N02BE anilides (paracetamol)
# N05BA benzodiazepine anxiolytics
# N05CD/CF benzodiazepine and related hypnotics (z-drugs)
# N05AA-AL antipsychotics
# N05AX other antipsychotics
# N06A antidepressants
# N06D anti-dementia drugs
# N07BC drugs used in opioid dependence
# R03AC02-05 short-acting inhaled beta-2 agonists
# R03BB04-07 long-acting inhaled anti-muscarinics (anticholinergics)
#-------------------------------------------------------------
nmi_atc <- "^(A06A|A07DA|A10A|B01AC|B03A|C01AA|C03C|C03EB|C03DA|C09[CD]|C10AA|H02AB|J01C|N02A|N02BE|N05BA|N05C[DF]|N05A[A-L]|N05AX|N06A|N06D|N07BC|R03AC0[2-5]|R03BB0[4-7])"
nmi_lmdb <- read_register("lmdb") %>% # without fastreg: open_dataset("path/to/lmdb/")
rename_with(tolower) %>%
semi_join(tibble(pnr = pnr_vec), by = "pnr") %>% # only the cohort
filter(grepl(nmi_atc, atc)) %>% # keep rows where the atc column matches the nmi_atc regex (NMI drug classes) - BEFORE collect()
collect()
saveRDS(nmi_lmdb, "datasets/nmi_lmdb.rds") # pnr, eksd, atcYou started from full_cohort.rds (Phase 10: pnr + index_date) and now have nmi_diagnoses.rds and nmi_lmdb.rds. Save them. Because the extraction is filtered to only NMI’s codes, the files are small, so you rarely need to clean up. If memory is tight anyway, free it with rm() + gc() (see Phase 5) rather than restarting the whole R session.
2. Compute the NMI score
See the full NMI computation (corrected against the Stata source)
#=============================================================
# 2. COMPUTE the NMI score
#=============================================================
library(dplyr)
library(stringr)
library(lubridate)
library(tidyr)
#-------------------------------------------------------------
# 2.0 Load data
#-------------------------------------------------------------
cohort <- readRDS("datasets/full_cohort.rds") # pnr, index_date
dx_raw <- readRDS("datasets/nmi_diagnoses.rds") # from get_lpr_diagnoses(): pnr, date_contact, c_diag
rx_raw <- readRDS("datasets/nmi_lmdb.rds") # pnr, eksd, atc
#-------------------------------------------------------------
# 2.1 Diagnoses: 5-year lookback before index (Stata: 1825 days)
#-------------------------------------------------------------
# Four steps, one at a time - we keep reassigning dx all the way through.
# 2.1.1 Attach the index date to each diagnosis (inner join, so only cohort members are kept)
dx <- dx_raw %>%
inner_join(cohort %>% select(pnr, index_date), by = "pnr")
# 2.1.2 Keep only diagnoses inside the 5-year window before index
dx <- dx %>%
filter(date_contact >= index_date - days(1825), # 1825 days = 5 years before index (date_contact = d_inddto in LPR2 / kont_starttidspunkt in LPR3, renamed in Phase 9)
date_contact <= index_date - days(1)) # up to the day before index (not index day itself)
# 2.1.3 Set one TRUE/FALSE flag per NMI predictor (str_detect() matches c_diag with D-prefix against the predictor's ICD codes)
dx <- dx %>%
mutate(
dx_B18 = str_detect(c_diag, "^DB18"),
dx_C34 = str_detect(c_diag, "^DC34"),
dx_C50 = str_detect(c_diag, "^DC50"),
dx_C61 = str_detect(c_diag, "^DC61"),
dx_C67 = str_detect(c_diag, "^DC67"),
dx_C70_D432 = str_detect(c_diag, "^DC7[01]|^DC75[1-3]|^DD32|^DD33[0-2]|^DD35[2-4]|^DD42|^DD43[0-2]|^DD44[3-5]"),
dx_C76_C80 = str_detect(c_diag, "^DC7[6-9]|^DC80"),
dx_C91_C95 = str_detect(c_diag, "^DC9[1-5]"),
dx_D50_D64 = str_detect(c_diag, "^DD5[0-9]|^DD6[0-4]"),
dx_E11 = str_detect(c_diag, "^DE11"),
dx_E86 = str_detect(c_diag, "^DE86"),
dx_F00_G319 = str_detect(c_diag, "^DF0[0-3]|^DG30"),
dx_F10 = str_detect(c_diag, "^DF10"),
dx_F17 = str_detect(c_diag, "^DF17"),
dx_G20_22 = str_detect(c_diag, "^DG2[0-2]"),
dx_G35 = str_detect(c_diag, "^DG35"),
dx_G40_G41 = str_detect(c_diag, "^DG4[0-1]"),
dx_I05_35 = str_detect(c_diag, "^DI0[56]|^DI3[45]"),
dx_I110_I50 = str_detect(c_diag, "^DI110|^DI13[02]|^DI42[06789]|^DI50"),
dx_I60_I69 = str_detect(c_diag, "^DI6[0-9]"),
dx_I70_I77 = str_detect(c_diag, "^DI7[0347]"), # I70, I73, I74, I77 (Stata: ^I7[0347])
dx_I71_I72 = str_detect(c_diag, "^DI7[12]"),
dx_J12_J18 = str_detect(c_diag, "^DJ1[2-8]"),
dx_J41_J47 = str_detect(c_diag, "^DJ4[12347]|^DJ96[19]"),
dx_J84 = str_detect(c_diag, "^DJ84"),
dx_K02_K08 = str_detect(c_diag, "^DK0[234568]"),
dx_K70_K767 = str_detect(c_diag, "^DK7[024]|^DK76[67]"),
dx_L89 = str_detect(c_diag, "^DL89"),
dx_N18_N19 = str_detect(c_diag, "^DN1[89]")
)
# 2.1.4 Collapse to one row per person (one column per dx_ predictor valued 0/1; any() = 1 if at least one matching diagnosis in the 5-year window, else 0)
dx <- dx %>%
group_by(pnr) %>%
summarise(across(starts_with("dx_"), ~ as.integer(any(.))), .groups = "drop")
#-------------------------------------------------------------
# 2.2 Prescriptions: 6-month lookback before index
#-------------------------------------------------------------
# Same pattern as the diagnoses (attach, filter, flag, collapse), here in a single chain.
rx <- rx_raw %>%
inner_join(cohort %>% select(pnr, index_date), by = "pnr") %>%
filter(eksd >= index_date - days(180), # 180 days = 6 months before index (eksd = redemption date)
eksd <= index_date - days(1)) %>% # up to the day before index
# One TRUE/FALSE flag per drug predictor: matches atc against the ATC prefix
mutate(
rx_A06A = str_detect(atc, "^A06A"),
rx_A07DA = str_detect(atc, "^A07DA"),
rx_A10A = str_detect(atc, "^A10A"),
rx_B01AC = str_detect(atc, "^B01AC"),
rx_B03A = str_detect(atc, "^B03A"),
rx_C01AA = str_detect(atc, "^C01AA"),
rx_C03C_C03EB = str_detect(atc, "^C03C|^C03EB"),
rx_C03DA = str_detect(atc, "^C03DA"),
rx_C09C_C09D = str_detect(atc, "^C09[CD]"),
rx_C10AA = str_detect(atc, "^C10AA"),
rx_H02AB = str_detect(atc, "^H02AB"),
rx_J01C = str_detect(atc, "^J01C"),
rx_N02A = str_detect(atc, "^N02A"),
rx_N02BE = str_detect(atc, "^N02BE"),
rx_N05BA_N05CF = str_detect(atc, "^N05BA|^N05C[DF]"),
rx_N05AA_N05AX = str_detect(atc, "^N05A[A-L]|^N05AX"),
rx_N06A = str_detect(atc, "^N06A"),
rx_N06D = str_detect(atc, "^N06D"),
rx_N07BC = str_detect(atc, "^N07BC"),
rx_R03AC02_R03AC05 = str_detect(atc, "^R03AC0[2-5]"),
rx_R03BB04_R03BB07 = str_detect(atc, "^R03BB0[4-7]")
) %>%
# Collapse to one row per person: 1 = at least one fill in the 6-month window, else 0
group_by(pnr) %>%
summarise(across(starts_with("rx_"), ~ as.integer(any(.))), .groups = "drop")
#-------------------------------------------------------------
# 2.3 Combine and score (weights from Table 2)
#-------------------------------------------------------------
nmi_data <- cohort %>% # reload your cohort here if needed (the name may be something other than "cohort")
left_join(dx, by = "pnr") %>% # attach the diagnosis flags to every cohort member
left_join(rx, by = "pnr") %>% # and the drug flags
# people with no diagnosis/prescription aren't in dx/rx -> left_join gives NA -> set to 0 (they have none)
mutate(across(c(starts_with("dx_"), starts_with("rx_")), ~ replace_na(., 0))) %>%
mutate(
nmi =
10 * dx_B18 + 19 * dx_C34 + 4 * dx_C50 + 5 * dx_C61 + 8 * dx_C67 +
8 * dx_C70_D432 + 22 * dx_C76_C80 + 8 * dx_C91_C95 + 5 * dx_D50_D64 +
2 * dx_E11 +
6 * dx_E86 + 9 * dx_F00_G319 + 12 * dx_F10 + 4 * dx_F17 +
7 * dx_G20_22 + 7 * dx_G35 + 5 * dx_G40_G41 +
2 * dx_I05_35 + 4 * dx_I110_I50 + 4 * dx_I60_I69 +
5 * dx_I70_I77 + 4 * dx_I71_I72 +
4 * dx_J12_J18 + 4 * dx_J41_J47 + 7 * dx_J84 +
5 * dx_K02_K08 + # dental, weight 5
13 * dx_K70_K767 + # fixed: liver disease, weight 13 (was 13 * dx_K02_K08)
11 * dx_L89 + 7 * dx_N18_N19 +
8 * rx_A06A + 5 * rx_A07DA +
4 * rx_A10A +
2 * rx_B01AC + 5 * rx_B03A + 4 * rx_C01AA + 5 * rx_C03C_C03EB + 3 * rx_C03DA +
(-2) * rx_C09C_C09D + (-3) * rx_C10AA + # negative weights: protective in the mortality model (lower the score)
2 * rx_H02AB + 1 * rx_J01C + 2 * rx_N02A + 2 * rx_N02BE +
1 * rx_N05BA_N05CF + 7 * rx_N05AA_N05AX + 3 * rx_N06A + 11 * rx_N06D +
7 * rx_N07BC + 3 * rx_R03AC02_R03AC05 + 5 * rx_R03BB04_R03BB07
)
saveRDS(nmi_data %>% select(pnr, nmi), "datasets/extract_nmi.rds")3. Look at and describe the score
The score is one number per person in extract_nmi.rds. Join it to your analysis dataset (Phase 12) and inspect it - NMI is right-skewed (many low, a few high), so median [IQR] often describes it better than the mean, though mean + CI is fine for comparing groups.
See the code: describe the NMI score
library(dplyr)
nmi <- readRDS("datasets/extract_nmi.rds") # pnr, nmi
analysis <- readRDS("datasets/analysis.rds") %>% # your assembled dataset (Phase 12)
left_join(nmi, by = "pnr")
# Distribution overall
summary(analysis$nmi)
hist(analysis$nmi)
# By group (e.g. exposed vs comparison cohort): mean + 95% CI and median [IQR]
analysis %>%
group_by(group) %>%
summarise(
n = n(),
mean_nmi = mean(nmi, na.rm = TRUE), # na.rm = TRUE: remove NA (missing values) before computing
sd_nmi = sd(nmi, na.rm = TRUE),
median_nmi = median(nmi, na.rm = TRUE),
q1 = quantile(nmi, 0.25, na.rm = TRUE),
q3 = quantile(nmi, 0.75, na.rm = TRUE)
) %>%
mutate(
# 95% CI for the mean: mean ± 1.96 * standard error (SE = sd / sqrt(n))
ci_low = mean_nmi - 1.96 * sd_nmi / sqrt(n),
ci_high = mean_nmi + 1.96 * sd_nmi / sqrt(n)
)For a formatted Table 1 across groups, gtsummary::tbl_summary(by = group) does this automatically - see Descriptive tables.
Consider whether the 5-year lookback should be an inclusion criterion. NMI’s diagnosis window needs at least 5 years of register data before the index date. If you make full lookback an inclusion criterion (only people with at least 5 years of prior residency/coverage in Denmark), apply it already when you build your cohort, not here - and apply the same requirement to any comparison cohort, so both groups have the same depth of data before index. Decide it in Phase 1 and apply it in Phase 10.
Example: make the 5-year lookback an inclusion criterion (conceptual)
The idea is to require that each person has resided in Denmark (and is therefore covered by the registers) for at least 5 years before their index date. You check this against residence/migration data (VNDS/BEF) and apply it as a filter before you match a comparison cohort:
# Conceptual - adapt to your residence data
kohort <- kohort %>%
mutate(lookback_ok = as.numeric(index_date - residence_start) >= 365.25 * 5) %>% # at least 5 years before index
filter(lookback_ok) # keep only full lookbackApply the same requirement to the comparison cohort’s assigned index dates, so the two groups are comparable.
More information
See the article for the full predictor list and weights. See Overview of registers for confirmed column names in LPR and LMDB.