We utilized the UK Biobank (UKB) dataset, a large-scale biomedical database and research resource containing in-depth genetic and health information from over half a million UK participants. Following a formal data application process, we obtained permission to access and utilize the UKB dataset for our research. UK Biobank Dataset The dataset includes comprehensive health data collected through physical measurements, biological samples, and self-reported questionnaires, complemented by linked electronic health records. The data was initially collected between 2006 and 2010, with ongoing updates to include new information on health outcomes and laboratory tests as well as long-term follow-ups.
After the data cleaning steps, the cleaned dataset includes 413,146 observations and 113 variables. Although we will not use all the variables in this dataset, we will select multiple key variables to investigate our research questions:
expo_npar_cat: Neutrophil-to-albumin ratio categorized
into quartiles (Q1 = lowest, Q4 = highest).
age: Participants’ age, categorized into 10-year age
groups (e.g., 30–39, 40–49, etc.).
sex: Male or female.
income: Seven income brackets, ranging from 1 (lowest)
to 7 (highest).
townsend: Townsend deprivation index, a measure of
socioeconomic status, where higher values indicate greater deprivation.
total_met: Total metabolic equivalent of task score,
derived from physical activity data and grouped into categories.
diet_quality: Diet quality score, ranging from 0
(poorest) to 7 (highest), based on adherence to DASHguideline.
sleep_hour: Nightly sleep duration, measured in hours.
smoke_cat: Smoking status, categorized as current
smoker, past smoker, or non-smoker.
total_alcohol: Total alcohol consumption (ml/week),
grouped into categories (non-drinker, low, moderate, high, very high).
non_hdl: Non-HDL cholesterol level, a continuous
variable indicating cardiovascular risk. (mg/dl)
tg: Triglyceride levels, a continuous measure of blood
lipid levels. (mg/dl)
bp_cat: Blood pressure, categorized into intervals
(e.g., BP=20, 40, 60, etc.).
Lymphocyte-to-Monocyte Ratio (LMR): \[\text{LMR} = \frac{\text{Absolute Lymphocyte Count}}{\text{Absolute Monocyte Count}}\]
Systemic Immune-Inflammation Index (SII): \[\text{SII} = \frac{\text{Neutrophil Count} \times \text{Platelet Count}}{\text{Lymphocyte Count}}\]
Naples Prognostic Score (NPS): This score is based on a combination of factors, typically including:
The final NPS is the sum of these points (range: 0–4).
Neutrophil-Percentage-to-Albumin Ratio (NPAR): \[\text{NPAR} = \frac{\text{Neutrophil Percentage}}{\text{Albumin Level (g/dL)}}\]
Note: The variables were categorized into tertiles based on the 25% and 75% percentiles after being ranked in ascending order, except for NPS, which was dichotomized based on whether it was ≥ 2.
View Full Part One Data Cleaning Steps on GitHub
The pre-cleaning process involved several key steps to transform the
raw UKB dataset into a clean, analyzable format. Initially, variables
were renamed to improve clarity and usability. For instance,
n_21022_0_0 was renamed to age,
n_31_0_0 to sex, and n_21000_0_0
to ethnicity. New variables were derived from existing data
to enhance analysis potential. For example, physical activity metrics
such as MET_walk (n_22037_0_0),
MET_moderate (n_22038_0_0), and
MET_vigorous (n_22039_0_0) were combined to
create total_MET, which was further categorized into
activity levels using pa_cat. Invalid or missing data
values, such as -3 and -1, were replaced with
. (missing), while specific imputations were applied where
necessary, such as assigning 0.5 to dietary variables like
cooked_vegetable_0 when valid data was unavailable.
Variables were categorized for analytical ease, such as grouping
physical activity, sleep, and dietary patterns. Dietary components like
vegetables, fruits, fish, red meat, and grains were processed to create
binary indicators of “healthy” consumption (e.g.,
healthy_vegetable for ≥3 servings/day of vegetables and
healthy_red_meat for ≤1.5 servings/week of red meat). An
aggregate dietary quality score (diet_quality) was computed
by summing these indicators and further categorized into quality
levels.
For sleep patterns, multiple variables such as chronotype
(Chronotype_preference_0), insomnia
(Insomnia_0), snoring (snoring_0), and daytime
sleepiness (daytime_sleepiness_0) were processed to
calculate a composite sleep quality score (sleep_quality).
This score was then combined with categorized sleep duration
(sleep_cat) to refine the assessment of overall sleep
health.
Lifestyle factors like smoking and alcohol consumption were also
carefully processed. Smoking history variables
(smoke_current and smoke_past) were combined
to categorize individuals’ smoking behavior into smoke_cat.
Alcohol consumption was calculated using variables like
redwine_week, beer_week, and
spirits_week, and total intake was categorized into risk
levels.
Biomarkers such as lipid and glucose levels were processed to
generate new variables like non_hdl (non-HDL cholesterol
derived from tc and hdl) and
glu_cat (glucose levels categorized based on clinical
thresholds). Blood pressure variables (sbp and
dbp) were similarly used to classify individuals into
health risk categories.
The cleaned and processed dataset, containing all these refined and derived variables, was saved as a new csv to be used for further analysis. This comprehensive pre-cleaning ensured that the dataset was both reliable and tailored to support the research objectives.
View Full Part Two Data Cleaning Steps on GitHub
The subsequent data cleaning process was designed to ensure the
dataset’s quality and readiness for complex analyses, including
multi-state modeling. The process began with importing the csv file
exported in the previous steps and standardizing variable names using
janitor::clean_names() to improve readability and
usability. Key variables were then transformed and categorized. For
instance, sex was converted into a factor variable with 0
representing female and 1 representing male, while
ethnicity_cat was categorized as 0 for white and 1 for
others. Age was grouped into meaningful categories
(age_cat): <50, 50–59,
60–69, and ≥70. Socioeconomic variables, like
income, and lifestyle variables, such as smoking status
(smoke_cat), were also transformed into categorical
factors.
Participants with pre-existing conditions at baseline were excluded
to ensure accurate survival and progression analyses. This included
those with non-alcoholic fatty liver disease (NAFLD) or cirrhosis,
identified through the variables nafld_date and
cirrho_date. Outcome variables were created to classify
participants’ health status. For example, nafld_outcome and
cirrho_outcome assigned 0 for healthy participants and 1
for those who developed the disease post-baseline. Missing disease dates
were carefully imputed using other known dates, such as the death date
(death_date) or lost-to-follow-up date
(lost_date). For participants with no recorded outcomes, a
default date of 2022-10-31 was used to mark the end of
follow-up.
To calculate disease progression metrics, survival durations were
computed from the baseline to either disease diagnosis or the end of
follow-up. These metrics were stored as nafld_surv_duration
and cirrho_surv_duration. Blood test variables were also
renamed and cleaned, including lymphocytes (b_lympho),
monocytes (b_mono), platelets (b_plate), and
neutrophils (b_neutro). Observations with missing values in
these variables were dropped to ensure reliable downstream analyses.
New biomarkers were derived to assess participants’ health status and
inflammation levels. For example, the lymphocyte-to-monocyte ratio
(expo_lmr), systemic immune-inflammation index
(expo_sii), and neutrophil-to-albumin ratio
(expo_npar) were calculated. Critical thresholds were
applied to categorize these biomarkers into clinical risk groups, such
as low albumin levels (cri_album <4 g/dL) or high
neutrophil-to-lymphocyte ratios (cri_nlr >3). Biomarkers
were further categorized into quartiles using a custom function
(cut_1_23_4) to allow for stratified analyses.
The final dataset included essential variables, such as demographics
(sex, age), lifestyle factors
(smoke_cat, total_alcohol), biomarkers (e.g.,
tc for total cholesterol, hdl for high-density
lipoprotein, hba1c for blood glucose), and disease
outcomes. These variables were selected and exported as
data_prepared.dta. This ensured the dataset was robust,
consistent, and ready for both descriptive and inferential analyses,
particularly for modeling disease progression pathways.