UK Biobank Analyses

Overview

This repository contains code used in producing the UKB-DRS risk score by Anatürk, Patel, and colleagues. This project uses data from the UK Biobank to develop a novel risk score for dementia prediction. The score is mainly developed through a series of R scripts which derive diagnoses for dementia and other relevant conditions before deriving risk score formulae for the UKB-DRS and other previously developed risk scores.

Contents

• scripts/ contains R scripts used for data manipulation, deriving risk scores, and assessing performance
• results/ contains output plots, lasso fit outputs, as well as the UKB-DRS calculator in excel format
• models/ contains final lasso fitting in .rda format All files in results/ and models/ are created via scripts in scripts/

Requirements

For optimal performance, we recommend a system with >=16GB of RAM and >=4 CPU cores. Development was carried out on a Ubuntu 18.04.6 LTS (Bionic Beaver) system. R version 3.6.3 was used. See scripts/session_info_R for specific package versions. For dementia ascertain python scripts, python version 3.6.5 was used.

1. Dementia ascertain

1.1 Mining GP prescription and clinical events data

Dementia cases were identified through primary care and secondary care records, death reports and self-report (at baseline only, to exclude pre-existing cases). Participants were classified as a dementia case if they had a record of a primary or secondary diagnosis of dementia in their health/death records or had been prescribed a common 'dementia drugs' (e.g., memantine).

The script extract_diagnoses_from_primary_care.py searches through the gp clinical events records (i.e., gp_clinical.csv) and identifies participants with read codes (e.g., bnf, dmd, read_v2) corresponding to several diseases of interest (e.g., dementia, stroke). The output is a csv file of all participants who have atleast one of the listed read codes of interest ("participants_with_dementia.csv"). These participants are then assigned a value of "1" and merged back to the original dataframe in the series of r scripts described below. Similarly, extract_prescriptions_from_primary_care.py mines the database containing prescription records (i.e. gp_scripts.csv). The list of read codes for each disease/medication class of interest was created using all_lkps_maps_v2.

1.2 Self-report prescription data

To identify individuals taking hormone replacement therapies (HRTs), anti-hypertensive medications and other treatments of interest, we have grouped all of the drugs recorded in data-field 20003 according to their ATC codes [Wu et al. 2019, Supplementary Figure 1].

2. Main analysis

The main analysis is performed using a series of R scripts. A brief description of each script is provided below:

1. 1_derive_diagnoses.R: Generates variables corresponding to an individual's status for several diseases of interest (e.g,., "1" for dementia, "0" for no dementia). It also derives corresponding dates_of_diagnosis and time_to_diagnosis variables. Next, we exclude people based on prespecified criteria, such as anyone reporting a diagnosis of dementia at baseline or anyone who developed dementia within a year of enrolling into the UKB study. Finally, we also generate variables to reflect medication use (e.g., antidepressants).

2. 2_anu_adri.R: Recodes all demographic and genetic variables of interest to compute the ANU-ADRI.

3. 3_caide.R: Recodes all relevant variables to compute the CAIDE, with the predicted probability of developing dementia (according to the CAIDE) also calculated.

4. 4_bdsi_retired.R: Originally used to calculate the BDSI but this script has now been retired as one of the core components cannot (e.g. money problems) be calculated in the biobank due to a lack of information.

5. 5_framingham.R: Calculates several versions of the framingham risk score as well as predicted probability of developing CVD.

6. 6_recode_variables.R: Here, we're recoding demographic and other variables of interest prior to our main analyses.

7. 7_DRS.R: Computing the DRS risk score.

8. 9_derive_hypertension_cholesterol_Dx.R: Derive some extra diagnoses

9. 10_cox_regression_LASSO.R: Run LASSO for feature selection, identify most predictive factors

10. 11_fit_cox.R: Fit a cox model, derive coefficients

11. 12_fit_crr.R: Fit a competing risk model, derive coefficients

12. 13_discrimination_and_calibration.R : This r script computes performance, via discriminative ability and calibration, of each model.

13. 14_compute_agespecific_auc.R: Test performance in age specific subsets

14. 15_sensitivity_time_to_diagnosis.R: Test performance in shorter time windows

15. 16_include_vs_exclude.R: Compare included vs excluded subjects