Installation

For development

pip install . -e

The -e flag marks the install as editable, "overwriting" the package as you edit the source files.

Recommended to also add black as a pre-commit hook: pre-commit install

For use

pip install git+https://github.com/Aarhus-Psychiatry-Research/psycop-ml-utils.git

Usage

Loading data from SQL

Currently only contains one function to load a view from SQL, sql_load

from loaders import sql_load

view = "[FOR_SFI_fritekst_resultat_udfoert_i_psykiatrien_aendret_2011]"
sql = "SELECT * FROM [fct]." + view
df = sql_load(sql, chunksize = None)

Flattening time series

To train baseline models (logistic regression, elastic net, SVM, XGBoost/random forest etc.), we need to represent the longitudinal data in a tabular, flattened way.

In essence, we need to generate a training example for each prediction time, where that example contains "latest_blood_pressure" (float), "X_diagnosis_within_n_hours" (boolean) etc.

To generate this, I propose the time-series flattener class (TimeSeriesFlattener). It builds a dataset like described above.

TimeSeriesFlattener

class FlattenedDataset:
    def __init__():
        """Class containing a time-series flattened.

        Args:
            prediction_times_df (DataFrame): Dataframe with prediction times.
            prediction_timestamp_colname (str, optional): Colname for timestamps. Defaults to "timestamp".
            id_colname (str, optional): Colname for patients ids. Defaults to "dw_ek_borger".
        """

    def add_outcome():
        """Adds an outcome-column to the dataset

        Args:
            outcome_df (DataFrame): Cols: dw_ek_borger, datotid, value if relevant.
            lookahead_days (float): How far ahead to look for an outcome in days. If none found, use fallback.
            resolve_multiple (str): What to do with more than one value within the lookahead.
                Suggestions: earliest, latest, mean, max, min.
            fallback (List[str]): What to do if no value within the lookahead.
                Suggestions: latest, mean_of_patient, mean_of_population, hardcode (qualified guess)
            timestamp_colname (str): Column name for timestamps
            values_colname (str): Colname for outcome values in outcome_df
            id_colname (str): Column name for citizen id
            new_col_name (str): Name to use for new col. Automatically generated as '{new_col_name}_within_{lookahead_days}_days'.
                Defaults to using values_colname.
        """

    def add_predictor():
        """Adds a predictor-column to the dataset

        Args:
            predictor_df (DataFrame): Cols: dw_ek_borger, datotid, value if relevant.
            lookahead_days (float): How far ahead to look for an outcome in days. If none found, use fallback.
            resolve_multiple (str): What to do with more than one value within the lookahead.
                Suggestions: earliest, latest, mean, max, min.
            fallback (List[str]): What to do if no value within the lookahead.
                Suggestions: latest, mean_of_patient, mean_of_population, hardcode (qualified guess)
            outcome_colname (str): What to name the column
            id_colname (str): Column name for citizen id
            timestamp_colname (str): Column name for timestamps
        """

Inspiration-code can be found in previous commits.

Example

• Update examples as API matures

import FlattenedDataset

dataset = FlattenedDataset(prediction_times_df = prediction_times, prediction_timestamp_colname = "timestamp", id_colname = "dw_ek_borger")

dataset.add_outcome(
    outcome_df=type_2_diabetes_df,
    lookahead_days=730,
    resolve_multiple="max",
    fallback=[0],
    name="t2d",
)

dataset.add_predictor(
    predictor=hba1c,
    lookback_window=365,
    resolve_multiple="max",
    fallback=["latest", 40],
    name="hba1c",
)

Dataset now looks like this:

For binary outcomes, add_predictor with fallback = [0] would take a df with only the times where the event occurred, and then generate 0's for the rest.

I propose we create the above functionality on a just-in-time basis, building the features as we need them.