timeseriesflattener: A Python package for summarizing features from (medical) time series

Martin Bernstoff, Kenneth Enevoldsen, Jakob Damgaard, Andreas Danielsen, Lasse Hansen

Abstract:

Time series from e.g. electronic health records often have a large number of variables that
are sampled at irregular and differing intervals. Before this type of data can be used for
prediction modelling with machine learning methods such as logistic regression or XGBoost
(Chen & Guestrin, 2016), the data needs to be reshaped. In essence, the time series need to
be flattened so that each prediction time is represented by a vector of predefined length. This
vector should hold the 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. This process of flattening the data lays the foundation for further analyses and
requires handling a number of tasks such as 1) how to deal with missing values, 2) which value
to use if none fall within the prediction window, 3) how to handle variables measured multiple
times within the chosen time window, and 4) how to handle predictors that attempt to look
further back than the start of the dataset.
timeseriesflattener aims to simplify this process by providing an easy-to-use and fullyspecified pipeline for flattening complex time series. timeseriesflattener implements all the
functionality required for aggregating features in specific time windows, grouped by e.g. patient
IDs, in a computationally efficient manner. The package is currently used for feature extraction
from electronic health records in studies based on the Psychiatric Clinical Outcome Prediction
Cohort (PSYCOP) projects (Hansen et al., 2021).