Data Model ¶

OneEHR uses a three-table input specification. Users prepare plain CSV files -- there is no dataset registry or special format.

Overview ¶

Table	Required	Description
`dynamic.csv`	Yes	Longitudinal event table (long format)
`static.csv`	No	Patient-level static covariates
`label.csv`	No	Label events (task-agnostic long format)

[dataset]
dynamic = "data/dynamic.csv"
static = "data/static.csv"     # optional
label = "data/label.csv"       # optional

`dynamic.csv` (required)¶

The core input: a unified longitudinal event table in long format (one row per event).

Required columns ¶

Column	Type	Description
`patient_id`	string	Patient identifier
`event_time`	datetime	Timestamp (parseable by `pandas.to_datetime`)
`code`	string	Measurement or concept name (e.g. `"heart_rate"`, `"diagnosis_A01"`)
`value`	numeric or string	Measured value

Example ¶

patient_id,event_time,code,value
P001,2023-01-01 08:00,heart_rate,72
P001,2023-01-01 08:00,blood_pressure_sys,120
P001,2023-01-01 08:00,diagnosis,A01
P001,2023-01-02 10:00,heart_rate,80
P002,2023-01-01 09:30,heart_rate,68
P002,2023-01-01 09:30,lab_glucose,5.4

Key points ¶

Timestamps are irregular -- events for the same patient can occur at any time
OneEHR bins events into fixed time windows during preprocessing (e.g. 1-hour, 1-day)
value can be numeric or categorical; type inference happens automatically during binning (a code is considered numeric if 90%+ of its values parse as numbers)
Numeric values: aggregated per bin via preprocess.numeric_strategy (mean or last)
Categorical values: encoded per bin via preprocess.categorical_strategy (one-hot or count)

`static.csv` (optional)¶

One row per patient with static covariates (demographics, baseline characteristics).

Required columns ¶

Column	Type	Description
`patient_id`	string	Patient identifier (must match `dynamic.csv`)

All other columns are treated as static features. Numeric columns become num__* features; categorical columns become cat__*__* one-hot features.

Example ¶

patient_id,age,sex,insurance
P001,65,M,Medicare
P002,42,F,Private
P003,78,M,Medicaid

Tip

Static features are concatenated with dynamic features for tabular models, and joined to test predictions for all models.

`label.csv` (optional)¶

A task-agnostic label event table in long format. This allows multiple label types to coexist in one file.

Required columns ¶

Column	Type	Description
`patient_id`	string	Patient identifier
`label_time`	datetime	When the label was observed
`label_code`	string	Label type identifier (e.g. `"outcome"`, `"los"`)
`label_value`	numeric	Label value

Example ¶

patient_id,label_time,label_code,label_value
P001,2023-01-05,outcome,1
P001,2023-01-05,los,4.5
P002,2023-01-03,outcome,0
P002,2023-01-03,los,2.0

How tables flow through OneEHR ¶

dynamic.csv ──┐
              ├──→ preprocess ──→ binned.parquet ──→ train
static.csv  ──┤                                       │
              │                                        ▼
label.csv ────┘                                    models + metrics

Preprocess bins dynamic.csv into fixed time windows and builds feature columns
Static features from static.csv are encoded into patient-level features
Labels come from label.csv
All downstream commands read from the materialized run artifacts instead of reinterpreting raw files

← Quickstart Tutorials →