CPRD-Pedro-MODY

R scripts for MODY prediction in CPRD

MODY T1D model

Bayesian hierarchical model combining case-control data with UNITED (population representative data). It uses a mixture approach splitting patients based on a latent variable T. If the patient is $C^- \cup A^+$ then $T = 1$. If the patient is $C^+ \cap A^-$ then $T = 0$. Patients with $T = 1$ have an informative prior Beta distribution, modelling the probability of MODY, with $mean = 0.029%$. Patients with $T = 0$ are modelled using shrinkage recalibration logistic model which scales the odds ratios and adjusts the intercept based on the general population data. In case T is missing, T is modelled using a logistic regression with the variables BMI, age of diagnosis, age of recruitment and parent history of diabetes (continuous variables are modelled with a restricted cubic spline, 3 knots calculated using UNITED dataset).

How to make predictions

The function used for predictions is available at 00.prediction_functions.R. In order to use it, you must load the function into the environment.

source("00.prediction_functions.R")

You also need to load the Bayesian model posteriors (model parameters) and format them the right way.

rcs_parms <- readRDS("model_posteriors/rcs_parms.rds")
posterior_samples_T1D <- readRDS("model_posteriors/type_1_model_posteriors.rds")
posterior_samples_T1D_obj <- list(post = posterior_samples_T1D$samples)
class(posterior_samples_T1D_obj) <- "T1D"

In order to make predictions, you need the following variables:

• pardm: Parent history of Diabetes, with history represented by 1 and no history represented by 0
• agerec: Age at recruitment
• hba1c: HbA_{1c}
• agedx: Age at diagnosis
• sex: Sex, with Male represented by a 0 and Female represented by a 1
• bmi: BMI
• T: C-peptide and autoantibody testing, with C+ and A- represented by 0, C- or A+ represented by 1 and any other combination represented by NA.

The dataset has to be formatted as a tibble.

predictions_x <- as_tibble(as.matrix(select(patients_dataset, pardm, agerec, hba1c, agedx, sex, bmi, T)))

Last thing to do is the predictions themselves. For that, use

predictions_T1D <- predict(posterior_samples_T1D_obj, predictions_x, rcs_parms) %>%
  apply(., 2, function(x) {
    data.frame(prob = mean(x), LCI = quantile(x, probs = 0.025), UCI = quantile(x, probs = 0.975))
  }) %>%
  bind_rows()

this will make the predictions and calculate the mean (lower and upper credible intervals at 2.5% and 97.5%) probability of having a MODY gene. If you are not looking at the uncertainty, only use the mean prediction.

MODY T2D model

Bayesian shrinkage recalibration logistic model which scales the odds ratios and adjusts the intercept based on the general population data.

How to make predictions

The function used for predictions is available at 00.prediction_functions.R. In order to use it, you must load the function into the environment.

source("00.prediction_functions.R")

You also need to load the Bayesian model posteriors (model parameters).

posterior_samples_T2D <- readRDS("model_posteriors/type_2_model_posteriors.rds")
posterior_samples_T2D_obj <- list(post = posterior_samples_T2D$samples)
class(posterior_samples_T2D_obj) <- "T2D"

In order to make predictions, you need the following variables:

• pardm: Parent history of Diabetes, with history represented by 1 and no history represented by 0
• agerec: Age at recruitment
• hba1c: HbA_{1c}
• agedx: Age at diagnosis
• sex: Sex, with Male represented by a 0 and Female represented by a 1
• bmi: BMI
• insoroha: Patient currently on insulin or tables, with TRUE represented by 1 and FALSE represented by 0

The dataset has to be formatted as a tibble.

predictions_x <- as_tibble(as.matrix(select(patients_dataset, pardm, agerec, hba1c, agedx, sex, bmi, insoroha)))

Last thing to do is the predictions themselves. For that, use

predictions_T2D <- predict(posterior_samples_T2D_obj, predictions_x) %>%
  apply(., 2, function(x) {
    data.frame(prob = mean(x), LCI = quantile(x, probs = 0.025), UCI = quantile(x, probs = 0.975))
  }) %>%
  bind_rows()

this will make the predictions and calculate the mean (lower and upper credible intervals at 2.5% and 97.5%) probability of having a MODY gene. If you are not looking at the uncertainty, only use the mean prediction.