Skip to content

Converting irregularly spaced time series, such as eletronic health records, into dataframes for tabular classification.

License

Notifications You must be signed in to change notification settings

Aarhus-Psychiatry-Research/timeseriesflattener

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Timeseriesflattener

github actions pytest python versions

PyPI version status

Time series from e.g. electronic health records often have a large number of variables, are sampled at irregular intervals and tend to have a large number of missing values. Before this type of data can be used for prediction modelling with machine learning methods such as logistic regression or XGBoost, the data needs to be reshaped.

In essence, the time series need to be flattened so that each prediction time is represented by a set of predictor values and an outcome value. These predictor values can be constructed by aggregating the preceding values in the time series within a certain time window.

timeseriesflattener aims to simplify this process by providing an easy-to-use and fully-specified pipeline for flattening complex time series.

🔧 Installation

To get started using timeseriesflattener simply install it using pip by running the following line in your terminal:

pip install timeseriesflattener

⚡ Quick start

import datetime as dt

import numpy as np
import polars as pl

# Load a dataframe with times you wish to make a prediction
prediction_times_df = pl.DataFrame(
    {"id": [1, 1, 2], "date": ["2020-01-01", "2020-02-01", "2020-02-01"]}
)
# Load a dataframe with raw values you wish to aggregate as predictors
predictor_df = pl.DataFrame(
    {
        "id": [1, 1, 1, 2],
        "date": ["2020-01-15", "2019-12-10", "2019-12-15", "2020-01-02"],
        "predictor_value": [1, 2, 3, 4],
    }
)
# Load a dataframe specifying when the outcome occurs
outcome_df = pl.DataFrame({"id": [1], "date": ["2020-03-01"], "outcome_value": [1]})

# Specify how to aggregate the predictors and define the outcome
from timeseriesflattener import (
    MaxAggregator,
    MinAggregator,
    OutcomeSpec,
    PredictionTimeFrame,
    PredictorSpec,
    ValueFrame,
)

predictor_spec = PredictorSpec(
    value_frame=ValueFrame(
        init_df=predictor_df, entity_id_col_name="id", value_timestamp_col_name="date"
    ),
    lookbehind_distances=[dt.timedelta(days=1)],
    aggregators=[MaxAggregator(), MinAggregator()],
    fallback=np.nan,
    column_prefix="pred",
)

outcome_spec = OutcomeSpec(
    value_frame=ValueFrame(
        init_df=outcome_df, entity_id_col_name="id", value_timestamp_col_name="date"
    ),
    lookahead_distances=[dt.timedelta(days=1)],
    aggregators=[MaxAggregator(), MinAggregator()],
    fallback=np.nan,
    column_prefix="outc",
)

# Instantiate TimeseriesFlattener and add the specifications
from timeseriesflattener import Flattener

result = Flattener(
    predictiontime_frame=PredictionTimeFrame(
        init_df=prediction_times_df, entity_id_col_name="id", timestamp_col_name="date"
    )
).aggregate_timeseries(specs=[predictor_spec, outcome_spec])
result.df

Output:

id date prediction_time_uuid pred_test_feature_within_30_days_mean_fallback_nan outc_test_outcome_within_31_days_maximum_fallback_0_dichotomous
0 1 2020-01-01 00:00:00 1-2020-01-01-00-00-00 2.5 0
1 1 2020-02-01 00:00:00 1-2020-02-01-00-00-00 1 1
2 2 2020-02-01 00:00:00 2-2020-02-01-00-00-00 4 0

📖 Tutorial

💬 Where to ask questions

Type
🚨 Bug Reports GitHub Issue Tracker
🎁 Feature Requests & Ideas GitHub Issue Tracker
👩‍💻 Usage Questions GitHub Discussions
🗯 General Discussion GitHub Discussions

🎓 Projects

PSYCOP projects use timeseriesflattener, see more at the monorepo.