Skip to content

MedSamples

he MedSamples object is designed to store key fields such as "label", "patient id", and "requested prediction time", along with additional information. Data is stored in a tab-separated file format. The "pred_0" field, representing a prediction result for performance analysis, is optional.

  • EVENT_FIELDS: Static field; always set to "SAMPLE"
  • id: Numeric patient identifier
  • time : Requested prediction time; only data prior to this point is used for prediction
  • outcome: The label or outcome; can be binary (0/1) or numeric (for regression)
  • outcomeTime: Event time if the patient is labeled "1", or "end of followup" for "0". This field is optional; if unused, you can specify a placeholder date like "19000101". Useful for filtering by time windows.
  • split: Optional field for specifying patient split. Typically, splits are assigned based on patient id in a separate file, so please specify "-1"
  • pred_0 - optional prediction result

Example file 

1
2
3
4
EVENT_FIELDS    id  time    outcome outcomeTime split
SAMPLE  1   20250101    0   20250820    -1
SAMPLE  1   20250201    0   20250820    -1
SAMPLE  2   20250101    1   20250820    -1

Explain:

  • Patient 1 has two prediction points (20250101 and 20250201), both labeled "0". Follow-up is documented until 20250820.
  • Patient 2 has one prediction point (20250101) with label "1". The event date is 20250820.

Overview

MedSamples is a data structure and set of helper functions for managing samples—combinations of individuals, time-points (dates), and associated information such as outcomes, outcome times, splits, and predictions.

The data is organized in a three-tier hierarchy:

  1. MedSamples: Represents a collection of MedIdSamples for multiple patients
  2. MedIdSamples: Represents a set of samples for a specific patient (id).
  3. MedSample: Represents a single sample.

Notes:

  • It is not strictly enforced that all samples within a MedIdSample share the same id, but many functions may not work correctly otherwise.
  • Different samples for the same id may have different splits, and these may not match the split parameter in MedIdSample. However, it is strongly recommended to maintain a single split per id.

Samples can be read from and written to CSV or binary files.