---
title: "Algorithm"
bibliography: references.bib
csl: vancouver.csl
vignette: >
  %\VignetteIndexEntry{Algorithm}
  %\VignetteEngine{quarto::html}
  %\VignetteEncoding{UTF-8}
execute: 
  echo: false
---

```{r setup}
#| include: false
library(osdc)
library(dplyr)

algorithm_df <- tibble(name = names(algorithm()), value = algorithm()) |>
  tidyr::unnest_wider(value) |>
  dplyr::select(register, name, title, logic, comments)
```

## General description

The current implementation of the algorithm is concisely described
below. A more complete description of the original implementation of the
classifier is found in Anders Aasted Isaksen's [PhD
Thesis](https://aastedet.github.io/dissertation/) as well as the
validation paper.[@Isaksen2023]

In brief, the algorithm first identifies a diabetes
population---regardless of diabetes type--- based on four types of
inclusion events. It then classifies individuals within this population
as having type 1 diabetes (T1D) based on patterns in glucose-lowering
drug purchases and recorded type-specific diabetes diagnoses.
Individuals who do not meet the criteria for T1D are subsequently
classified as having type 2 diabetes (T2D).

Inclusion events are:

1.  HbA1c measurements of ≥48 mmol/mol.
2.  Hospital diagnoses of diabetes.
3.  Diabetes-specific services received at podiatrist.
4.  Purchase of glucose-lowering drugs.

Some inclusion events are ignored if any of the following exclusion
events apply:

1.  Elevated HbA1c samples are dropped if:
    -   The sample was taken during pregnancies, as these may be related
        to gestational diabetes mellitus rather than T1D or T2D.
2.  Purchases of glucose-lowering drugs are dropped if:
    -   The purchased drug is a GLP-1 receptor agonist, as this drug
        class is used extensively for weight loss among [individuals
        without diabetes](https://doi.org/10.2147/clep.s456170).
    -   The purchased drug is dapagliflozin or empagliflozin, two types
        of SGLT-2 inhibitors recommended in Danish guidelines for
        treatment of heart
        failure (Section 5.5.1 of Dansk Cardiologisk Selskab)
        or [kidney
        failure](https://nephrology.dk/vejledninger/ckd-mbd/kronisk-nyresygdom/sglt2i_ckd_uden_dm/)
        among individuals without diabetes. This pertains only to
        purchases of drugs containing only this class of
        glucose-lowering drug. Purchases of combination drugs containing
        GLP-1 receptor agonists or SGLT-2 inhibitors in combination with
        other glucose-lowering drugs are not dropped.
    -   The drug purchase is made during pregnancies, as these may be
        related to gestational diabetes mellitus, rather than T1D or
        T2D.
    -   If the purchased drug is metformin and is made by women below
        age 40, as these may be treatment for polycystic ovary syndrome
        (PCOS).
    -   If the purchased drug is metformin with a PCOS-related
        indication code.

The diabetes classification date is defined at the date of the second
occurrence of any of the inclusion events listed above. For example, an
individual with two elevated HbA1c tests followed by a glucose-lowering
drug purchase is included at the latest elevated HbA1c test. Had the
second HbA1c test not been performed (or had it returned a result below
the diagnostic threshold), this person would instead have been included
at the date of the first purchase of glucose-lowering drugs. Wherever
possible, all available data for each event is used, except for the
purchases of glucose-lowering drugs, since the patient register data on
obstetric diagnoses necessary to censor glucose-lowering drug purchases
is only complete from 1997 onwards.

## High-level flowchart

![Flow of inclusions and exclusions used for classifying diabetes status
using the osdc package. Light blue and orange boxes represent inclusions
and exclusions, respectively.](images/overview-flow.svg)

### Classifying type 1 diabetes

Diabetes type is classified as either T1D or T2D based on patterns of
purchases of insulin drugs (including analogues) and hospital primary
diagnoses of T1D and T2D.

For an individual to be classified as having T1D, the available data
needs to fulfill the criteria in either of the following branches of
logic:

a.  Must only have recorded purchases of insulin drugs and never any
    other type of glucose-lowering drugs, and have at least one primary
    diagnosis of T1D from a medical specialty hospital department.
b.  Must have a majority of T1D diagnoses from endocrinology departments
    (or from other medical specialty departments, in the absence of
    contacts to endocrinology departments), and a purchase of insulin
    within 180 days of their initial glucose-lowering drug purchase,
    with insulin contributing at least two thirds of all defined daily
    doses of glucose-lowering drugs purchased.

### Classifying type 2 diabetes

Individuals not classified as T1D cases are classified as T2D cases.

## Detailed and technical description

Below are the technical, exact implementation of the above description.
These are the logical conditions and exact variables and registers used
that results in the classification of diabetes status and types. They
are shown in the form of tables for each register with a table at the
end showing the logic that is applied across registers.

### `lpr_diag`

```{r}
algorithm_df |> 
  dplyr::filter(register=="lpr_diag") |>
  dplyr::select(-register) |> 
  knitr::kable()
```

### `lpr_adm`

```{r}
algorithm_df |> 
  dplyr::filter(register=="lpr_adm") |>
  dplyr::select(-register) |> 
  knitr::kable()
```

### `diagnoser`

```{r}
algorithm_df |> 
  dplyr::filter(register=="diagnoser") |>
  dplyr::select(-register) |> 
  knitr::kable()
```

### `kontakter`

```{r}
algorithm_df |> 
  dplyr::filter(register=="kontakter") |>
  dplyr::select(-register) |> 
  knitr::kable()
```

### `lab_forsker`

```{r}
algorithm_df |> 
  dplyr::filter(register=="lab_forsker") |>
  dplyr::select(-register) |> 
  knitr::kable()
```

### `lmdb`

```{r}
algorithm_df |> 
  dplyr::filter(register=="lmdb") |>
  dplyr::select(-register) |> 
  knitr::kable()
```

### Across register logic

```{r}
algorithm_df |> 
  dplyr::filter(is.na(register)) |>
  dplyr::select(-register) |> 
  knitr::kable()
```

## References