Skip to content

custom_time_token

Transformations for adding time-derived measurements (e.g., a subject’s age) to a MEDS dataset.

add_new_events_fntr(fn)

Creates a “meta” functor that computes the input functor on a MEDS shard then combines both dataframes.

Parameters:

Name Type Description Default
fn Callable[[DataFrame], DataFrame]

The function that computes the new events.

 required

Returns:

Type Description
Callable[[DataFrame], DataFrame]

A function that computes the new events and combines them with the original DataFrame, returning a
Callable[[DataFrame], DataFrame]

result in proper MEDS sorted order.

Examples:

>>> from datetime import datetime
>>> df = pl.DataFrame(
...     {
...         "subject_id": [1, 1, 1, 1, 2, 2, 3, 3],
...         "time": [
...             None,
...             datetime(1990, 1, 1),
...             datetime(2021, 1, 1),
...             datetime(2021, 1, 1),
...             datetime(1988, 1, 2),
...             datetime(2023, 1, 3),
...             datetime(2022, 1, 1),
...             datetime(2022, 1, 1),
...         ],
...         "code": ["static", "DOB", "lab//A", "lab//B", "DOB", "lab//A", "lab//B", "dx//1"],
...     },
...     schema={"subject_id": pl.UInt32, "time": pl.Datetime, "code": pl.Utf8},
... )
>>> df
shape: (8, 3)
┌────────────┬─────────────────────┬────────┐
│ subject_id ┆ time                ┆ code   │
│ ---        ┆ ---                 ┆ ---    │
│ u32        ┆ datetime[μs]        ┆ str    │
╞════════════╪═════════════════════╪════════╡
│ 1          ┆ null                ┆ static │
│ 1          ┆ 1990-01-01 00:00:00 ┆ DOB    │
│ 1          ┆ 2021-01-01 00:00:00 ┆ lab//A │
│ 1          ┆ 2021-01-01 00:00:00 ┆ lab//B │
│ 2          ┆ 1988-01-02 00:00:00 ┆ DOB    │
│ 2          ┆ 2023-01-03 00:00:00 ┆ lab//A │
│ 3          ┆ 2022-01-01 00:00:00 ┆ lab//B │
│ 3          ┆ 2022-01-01 00:00:00 ┆ dx//1  │
└────────────┴─────────────────────┴────────┘
>>> # As an example, we'll use the age functor defined elsewhere in this module.
>>> age_cfg = DictConfig({"DOB_code": "DOB", "age_code": "AGE", "age_unit": "years"})
>>> age_fn = age_fntr(age_cfg)
>>> age_fn(df)
shape: (2, 4)
┌────────────┬─────────────────────┬──────┬───────────────┐
│ subject_id ┆ time                ┆ code ┆ numeric_value │
│ ---        ┆ ---                 ┆ ---  ┆ ---           │
│ u32        ┆ datetime[μs]        ┆ str  ┆ f32           │
╞════════════╪═════════════════════╪══════╪═══════════════╡
│ 1          ┆ 2021-01-01 00:00:00 ┆ AGE  ┆ 31.001347     │
│ 2          ┆ 2023-01-03 00:00:00 ┆ AGE  ┆ 35.004169     │
└────────────┴─────────────────────┴──────┴───────────────┘
>>> # Now, we'll use the add_new_events functor to add these age events to the original DataFrame.
>>> add_age_fn = add_new_events_fntr(age_fn)
>>> add_age_fn(df)
shape: (10, 4)
┌────────────┬─────────────────────┬────────┬───────────────┐
│ subject_id ┆ time                ┆ code   ┆ numeric_value │
│ ---        ┆ ---                 ┆ ---    ┆ ---           │
│ u32        ┆ datetime[μs]        ┆ str    ┆ f32           │
╞════════════╪═════════════════════╪════════╪═══════════════╡
│ 1          ┆ null                ┆ static ┆ null          │
│ 1          ┆ 1990-01-01 00:00:00 ┆ DOB    ┆ null          │
│ 1          ┆ 2021-01-01 00:00:00 ┆ AGE    ┆ 31.001347     │
│ 1          ┆ 2021-01-01 00:00:00 ┆ lab//A ┆ null          │
│ 1          ┆ 2021-01-01 00:00:00 ┆ lab//B ┆ null          │
│ 2          ┆ 1988-01-02 00:00:00 ┆ DOB    ┆ null          │
│ 2          ┆ 2023-01-03 00:00:00 ┆ AGE    ┆ 35.004169     │
│ 2          ┆ 2023-01-03 00:00:00 ┆ lab//A ┆ null          │
│ 3          ┆ 2022-01-01 00:00:00 ┆ lab//B ┆ null          │
│ 3          ┆ 2022-01-01 00:00:00 ┆ dx//1  ┆ null          │
└────────────┴─────────────────────┴────────┴───────────────┘
Source code in meds_torch/utils/custom_time_token.py 
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
def add_new_events_fntr(fn: Callable[[pl.DataFrame], pl.DataFrame]) -> Callable[[pl.DataFrame], pl.DataFrame]:
    """Creates a "meta" functor that computes the input functor on a MEDS shard then combines both dataframes.

    Args:
        fn: The function that computes the new events.

    Returns:
        A function that computes the new events and combines them with the original DataFrame, returning a
        result in proper MEDS sorted order.

    Examples:
        >>> from datetime import datetime
        >>> df = pl.DataFrame(
        ...     {
        ...         "subject_id": [1, 1, 1, 1, 2, 2, 3, 3],
        ...         "time": [
        ...             None,
        ...             datetime(1990, 1, 1),
        ...             datetime(2021, 1, 1),
        ...             datetime(2021, 1, 1),
        ...             datetime(1988, 1, 2),
        ...             datetime(2023, 1, 3),
        ...             datetime(2022, 1, 1),
        ...             datetime(2022, 1, 1),
        ...         ],
        ...         "code": ["static", "DOB", "lab//A", "lab//B", "DOB", "lab//A", "lab//B", "dx//1"],
        ...     },
        ...     schema={"subject_id": pl.UInt32, "time": pl.Datetime, "code": pl.Utf8},
        ... )
        >>> df
        shape: (8, 3)
        ┌────────────┬─────────────────────┬────────┐
        │ subject_id ┆ time                ┆ code   │
        │ ---        ┆ ---                 ┆ ---    │
        │ u32        ┆ datetime[μs]        ┆ str    │
        ╞════════════╪═════════════════════╪════════╡
        │ 1          ┆ null                ┆ static │
        │ 1          ┆ 1990-01-01 00:00:00 ┆ DOB    │
        │ 1          ┆ 2021-01-01 00:00:00 ┆ lab//A │
        │ 1          ┆ 2021-01-01 00:00:00 ┆ lab//B │
        │ 2          ┆ 1988-01-02 00:00:00 ┆ DOB    │
        │ 2          ┆ 2023-01-03 00:00:00 ┆ lab//A │
        │ 3          ┆ 2022-01-01 00:00:00 ┆ lab//B │
        │ 3          ┆ 2022-01-01 00:00:00 ┆ dx//1  │
        └────────────┴─────────────────────┴────────┘
        >>> # As an example, we'll use the age functor defined elsewhere in this module.
        >>> age_cfg = DictConfig({"DOB_code": "DOB", "age_code": "AGE", "age_unit": "years"})
        >>> age_fn = age_fntr(age_cfg)
        >>> age_fn(df)
        shape: (2, 4)
        ┌────────────┬─────────────────────┬──────┬───────────────┐
        │ subject_id ┆ time                ┆ code ┆ numeric_value │
        │ ---        ┆ ---                 ┆ ---  ┆ ---           │
        │ u32        ┆ datetime[μs]        ┆ str  ┆ f32           │
        ╞════════════╪═════════════════════╪══════╪═══════════════╡
        │ 1          ┆ 2021-01-01 00:00:00 ┆ AGE  ┆ 31.001347     │
        │ 2          ┆ 2023-01-03 00:00:00 ┆ AGE  ┆ 35.004169     │
        └────────────┴─────────────────────┴──────┴───────────────┘
        >>> # Now, we'll use the add_new_events functor to add these age events to the original DataFrame.
        >>> add_age_fn = add_new_events_fntr(age_fn)
        >>> add_age_fn(df)
        shape: (10, 4)
        ┌────────────┬─────────────────────┬────────┬───────────────┐
        │ subject_id ┆ time                ┆ code   ┆ numeric_value │
        │ ---        ┆ ---                 ┆ ---    ┆ ---           │
        │ u32        ┆ datetime[μs]        ┆ str    ┆ f32           │
        ╞════════════╪═════════════════════╪════════╪═══════════════╡
        │ 1          ┆ null                ┆ static ┆ null          │
        │ 1          ┆ 1990-01-01 00:00:00 ┆ DOB    ┆ null          │
        │ 1          ┆ 2021-01-01 00:00:00 ┆ AGE    ┆ 31.001347     │
        │ 1          ┆ 2021-01-01 00:00:00 ┆ lab//A ┆ null          │
        │ 1          ┆ 2021-01-01 00:00:00 ┆ lab//B ┆ null          │
        │ 2          ┆ 1988-01-02 00:00:00 ┆ DOB    ┆ null          │
        │ 2          ┆ 2023-01-03 00:00:00 ┆ AGE    ┆ 35.004169     │
        │ 2          ┆ 2023-01-03 00:00:00 ┆ lab//A ┆ null          │
        │ 3          ┆ 2022-01-01 00:00:00 ┆ lab//B ┆ null          │
        │ 3          ┆ 2022-01-01 00:00:00 ┆ dx//1  ┆ null          │
        └────────────┴─────────────────────┴────────┴───────────────┘
    """

    def out_fn(df: pl.DataFrame) -> pl.DataFrame:
        new_events = fn(df)
        df = df.with_row_index("__idx")
        new_events = new_events.with_columns(pl.lit(0, dtype=df.schema["__idx"]).alias("__idx"))
        return (
            pl.concat([df, new_events], how="diagonal").sort(by=["subject_id", "time", "__idx"]).drop("__idx")
        )

    return out_fn

age_fntr(cfg)

Create a function that adds a subject’s age to a DataFrame.

Parameters:

Name Type Description Default
cfg DictConfig

The configuration for the age function. This must contain the following mandatory keys: - “DOB_code”: The code for the date of birth event in the raw data. - “age_code”: The code for the age event in the output data. - “age_unit”: The unit for the age event when converted to a numeric value in the output data.

 required

Returns:

Type Description
Callable[[DataFrame], DataFrame]

A function that returns the to-be-added “age” events with the subject’s age for all input events with
Callable[[DataFrame], DataFrame]

unique, non-null times in the data, for all subjects who have an observed date of birth. It does
Callable[[DataFrame], DataFrame]

not add an event for times that are equal to the date of birth.

Raises:

Type Description
ValueError

If the input unit is not recognized.

Examples:

>>> from datetime import datetime
>>> df = pl.DataFrame(
...     {
...         "subject_id": [1, 1, 1, 1, 1, 2, 2, 3, 3],
...         "time": [
...             None,
...             datetime(1990, 1, 1),
...             datetime(2021, 1, 1),
...             datetime(2021, 1, 1),
...             datetime(2021, 1, 2),
...             datetime(1988, 1, 2),
...             datetime(2023, 1, 3),
...             datetime(2022, 1, 1),
...             datetime(2022, 1, 1),
...         ],
...         "code": ["static", "DOB", "lab//A", "lab//B", "rx", "DOB", "lab//A", "lab//B", "dx//1"],
...     },
...     schema={"subject_id": pl.UInt32, "time": pl.Datetime, "code": pl.Utf8},
... )
>>> df
shape: (9, 3)
┌────────────┬─────────────────────┬────────┐
│ subject_id ┆ time                ┆ code   │
│ ---        ┆ ---                 ┆ ---    │
│ u32        ┆ datetime[μs]        ┆ str    │
╞════════════╪═════════════════════╪════════╡
│ 1          ┆ null                ┆ static │
│ 1          ┆ 1990-01-01 00:00:00 ┆ DOB    │
│ 1          ┆ 2021-01-01 00:00:00 ┆ lab//A │
│ 1          ┆ 2021-01-01 00:00:00 ┆ lab//B │
│ 1          ┆ 2021-01-02 00:00:00 ┆ rx     │
│ 2          ┆ 1988-01-02 00:00:00 ┆ DOB    │
│ 2          ┆ 2023-01-03 00:00:00 ┆ lab//A │
│ 3          ┆ 2022-01-01 00:00:00 ┆ lab//B │
│ 3          ┆ 2022-01-01 00:00:00 ┆ dx//1  │
└────────────┴─────────────────────┴────────┘
>>> age_cfg = DictConfig({"DOB_code": "DOB", "age_code": "AGE", "age_unit": "years"})
>>> age_fn = age_fntr(age_cfg)
>>> age_fn(df)
shape: (3, 4)
┌────────────┬─────────────────────┬──────┬───────────────┐
│ subject_id ┆ time                ┆ code ┆ numeric_value │
│ ---        ┆ ---                 ┆ ---  ┆ ---           │
│ u32        ┆ datetime[μs]        ┆ str  ┆ f32           │
╞════════════╪═════════════════════╪══════╪═══════════════╡
│ 1          ┆ 2021-01-01 00:00:00 ┆ AGE  ┆ 31.001347     │
│ 1          ┆ 2021-01-02 00:00:00 ┆ AGE  ┆ 31.004084     │
│ 2          ┆ 2023-01-03 00:00:00 ┆ AGE  ┆ 35.004169     │
└────────────┴─────────────────────┴──────┴───────────────┘
>>> age_cfg = DictConfig({"DOB_code": "DOB", "age_code": "AGE", "age_unit": "scores"})
>>> import pytest
>>> with pytest.raises(ValueError):
...     age_fntr(age_cfg)
Source code in meds_torch/utils/custom_time_token.py 
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
def age_fntr(cfg: DictConfig) -> Callable[[pl.DataFrame], pl.DataFrame]:
    """Create a function that adds a subject's age to a DataFrame.

    Args:
        cfg: The configuration for the age function. This must contain the following mandatory keys:
            - "DOB_code": The code for the date of birth event in the raw data.
            - "age_code": The code for the age event in the output data.
            - "age_unit": The unit for the age event when converted to a numeric value in the output data.

    Returns:
        A function that returns the to-be-added "age" events with the subject's age for all input events with
        unique, non-null times in the data, for all subjects who have an observed date of birth. It does
        not add an event for times that are equal to the date of birth.

    Raises:
        ValueError: If the input unit is not recognized.

    Examples:
        >>> from datetime import datetime
        >>> df = pl.DataFrame(
        ...     {
        ...         "subject_id": [1, 1, 1, 1, 1, 2, 2, 3, 3],
        ...         "time": [
        ...             None,
        ...             datetime(1990, 1, 1),
        ...             datetime(2021, 1, 1),
        ...             datetime(2021, 1, 1),
        ...             datetime(2021, 1, 2),
        ...             datetime(1988, 1, 2),
        ...             datetime(2023, 1, 3),
        ...             datetime(2022, 1, 1),
        ...             datetime(2022, 1, 1),
        ...         ],
        ...         "code": ["static", "DOB", "lab//A", "lab//B", "rx", "DOB", "lab//A", "lab//B", "dx//1"],
        ...     },
        ...     schema={"subject_id": pl.UInt32, "time": pl.Datetime, "code": pl.Utf8},
        ... )
        >>> df
        shape: (9, 3)
        ┌────────────┬─────────────────────┬────────┐
        │ subject_id ┆ time                ┆ code   │
        │ ---        ┆ ---                 ┆ ---    │
        │ u32        ┆ datetime[μs]        ┆ str    │
        ╞════════════╪═════════════════════╪════════╡
        │ 1          ┆ null                ┆ static │
        │ 1          ┆ 1990-01-01 00:00:00 ┆ DOB    │
        │ 1          ┆ 2021-01-01 00:00:00 ┆ lab//A │
        │ 1          ┆ 2021-01-01 00:00:00 ┆ lab//B │
        │ 1          ┆ 2021-01-02 00:00:00 ┆ rx     │
        │ 2          ┆ 1988-01-02 00:00:00 ┆ DOB    │
        │ 2          ┆ 2023-01-03 00:00:00 ┆ lab//A │
        │ 3          ┆ 2022-01-01 00:00:00 ┆ lab//B │
        │ 3          ┆ 2022-01-01 00:00:00 ┆ dx//1  │
        └────────────┴─────────────────────┴────────┘
        >>> age_cfg = DictConfig({"DOB_code": "DOB", "age_code": "AGE", "age_unit": "years"})
        >>> age_fn = age_fntr(age_cfg)
        >>> age_fn(df)
        shape: (3, 4)
        ┌────────────┬─────────────────────┬──────┬───────────────┐
        │ subject_id ┆ time                ┆ code ┆ numeric_value │
        │ ---        ┆ ---                 ┆ ---  ┆ ---           │
        │ u32        ┆ datetime[μs]        ┆ str  ┆ f32           │
        ╞════════════╪═════════════════════╪══════╪═══════════════╡
        │ 1          ┆ 2021-01-01 00:00:00 ┆ AGE  ┆ 31.001347     │
        │ 1          ┆ 2021-01-02 00:00:00 ┆ AGE  ┆ 31.004084     │
        │ 2          ┆ 2023-01-03 00:00:00 ┆ AGE  ┆ 35.004169     │
        └────────────┴─────────────────────┴──────┴───────────────┘
        >>> age_cfg = DictConfig({"DOB_code": "DOB", "age_code": "AGE", "age_unit": "scores"})
        >>> import pytest
        >>> with pytest.raises(ValueError):
        ...     age_fntr(age_cfg)
    """

    canonical_unit, seconds_in_unit = normalize_time_unit(cfg.age_unit)
    microseconds_in_unit = int(1e6) * seconds_in_unit

    def fn(df: pl.LazyFrame) -> pl.LazyFrame:
        dob_expr = pl.when(pl.col("code") == cfg.DOB_code).then(pl.col("time")).min().over("subject_id")
        age_expr = (pl.col("time") - dob_expr).dt.total_microseconds() / microseconds_in_unit
        age_expr = age_expr.cast(pl.Float32, strict=False)

        return (
            df.drop_nulls(subset=["time"])
            .unique(subset=["subject_id", "time"], maintain_order=True)
            .select(
                "subject_id",
                "time",
                pl.lit(cfg.age_code, dtype=df.schema["code"]).alias("code"),
                age_expr.alias("numeric_value"),
            )
            .drop_nulls(subset=["numeric_value"])
            .filter(pl.col("numeric_value") > 0)
        )

    return fn

main(cfg)

Adds time-derived measurements to a MEDS cohort as separate observations at each unique time.

Source code in meds_torch/utils/custom_time_token.py 
423
424
425
426
427
428
429
@hydra.main(
    version_base=None, config_path=str(PREPROCESS_CONFIG_YAML.parent), config_name=PREPROCESS_CONFIG_YAML.stem
)
def main(cfg: DictConfig):
    """Adds time-derived measurements to a MEDS cohort as separate observations at each unique time."""

    map_over(cfg, compute_fn=add_time_derived_measurements_fntr)

normalize_time_unit(unit)

Normalize a time unit string to a canonical form and return the number of seconds in that unit.

Note that this function is designed for computing approximate time durations over long periods, not canonical, local calendar time durations. E.g., a “month” is not a fixed number of seconds, but this function will return the average number of seconds in a month, accounting for leap years.

TODO: consider replacing this function with the use of https://github.com/wroberts/pytimeparse

Parameters:

Name Type Description Default
unit str

The input unit to normalize.

 required

Returns:

Type Description
tuple[str, float]

A tuple containing the canonical unit and the number of seconds in that unit.

Raises:

Type Description
ValueError

If the input unit is not recognized.

Examples:

>>> normalize_time_unit("s")
('seconds', 1)
>>> normalize_time_unit("min")
('minutes', 60)
>>> normalize_time_unit("hours")
('hours', 3600)
>>> normalize_time_unit("day")
('days', 86400)
>>> normalize_time_unit("wks")
('weeks', 604800)
>>> normalize_time_unit("month")
('months', 2629746.0)
>>> normalize_time_unit("years")
('years', 31556926.080000002)
>>> normalize_time_unit("fortnight")
Traceback (most recent call last):
    ...
ValueError: Unknown time unit 'fortnight'. Valid units include:
  * seconds: s, sec, secs, second, seconds
  * minutes: m, min, mins, minute, minutes
  * hours: h, hr, hrs, hour, hours
  * days: d, day, days
  * weeks: w, wk, wks, week, weeks
  * months: mo, mos, month, months
  * years: y, yr, yrs, year, years
Source code in meds_torch/utils/custom_time_token.py 
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
def normalize_time_unit(unit: str) -> tuple[str, float]:
    """Normalize a time unit string to a canonical form and return the number of seconds in that unit.

    Note that this function is designed for computing _approximate_ time durations over long periods, not
    canonical, local calendar time durations. E.g., a "month" is not a fixed number of seconds, but this
    function will return the average number of seconds in a month, accounting for leap years.

    TODO: consider replacing this function with the use of https://github.com/wroberts/pytimeparse

    Args:
        unit: The input unit to normalize.

    Returns:
        A tuple containing the canonical unit and the number of seconds in that unit.

    Raises:
        ValueError: If the input unit is not recognized.

    Examples:
        >>> normalize_time_unit("s")
        ('seconds', 1)
        >>> normalize_time_unit("min")
        ('minutes', 60)
        >>> normalize_time_unit("hours")
        ('hours', 3600)
        >>> normalize_time_unit("day")
        ('days', 86400)
        >>> normalize_time_unit("wks")
        ('weeks', 604800)
        >>> normalize_time_unit("month")
        ('months', 2629746.0)
        >>> normalize_time_unit("years")
        ('years', 31556926.080000002)
        >>> normalize_time_unit("fortnight")
        Traceback (most recent call last):
            ...
        ValueError: Unknown time unit 'fortnight'. Valid units include:
          * seconds: s, sec, secs, second, seconds
          * minutes: m, min, mins, minute, minutes
          * hours: h, hr, hrs, hour, hours
          * days: d, day, days
          * weeks: w, wk, wks, week, weeks
          * months: mo, mos, month, months
          * years: y, yr, yrs, year, years
    """
    for canonical_unit, (aliases, seconds) in TIME_DURATION_UNITS.items():
        if unit in aliases:
            return canonical_unit, seconds

    valid_unit_lines = []
    for canonical, (aliases, _) in TIME_DURATION_UNITS.items():
        valid_unit_lines.append(f"  * {canonical}: {', '.join(aliases)}")
    valid_units_str = "\n".join(valid_unit_lines)
    raise ValueError(f"Unknown time unit '{unit}'. Valid units include:\n{valid_units_str}")

time_delta_fntr(cfg)

Create a function that adds time_delta event rows to a DataFrame.

Parameters:

Name Type Description Default
cfg DictConfig

The configuration for the age function. This must contain the following mandatory keys: - “quantile_fp”: The path to a parquet file with pre-defined time delta quantiles. Should have a quantile column of type Int64 and a value column of type Float64 indicating the numerical value for the quantile. - “max_length”: The maximum number of time_delta qunatile tokens to use to approximate the continuous value. - “time_unit”: The unit for the time deltas when dates are converted to a numeric value in the output data.

 required

Returns:

Type Description
Callable[[DataFrame], DataFrame]

A function that returns the to-be-added “time_delta” events with the quantile of the time delta
Callable[[DataFrame], DataFrame]

for all input events with unique, non-null times in the data. The very first event for a subject
Callable[[DataFrame], DataFrame]

has a null time_delta, so it is imputed with a special time start token.

Raises:

Type Description
ValueError

If the input unit is not recognized.

Examples:

>>> from datetime import datetime
>>> df = pl.DataFrame(
...     {
...         "subject_id": [1, 1, 1, 1, 1, 2, 2, 3, 3],
...         "time": [
...             None,
...             datetime(1990, 1, 1),
...             datetime(2021, 1, 1),
...             datetime(2021, 1, 1),
...             datetime(2021, 1, 2),
...             datetime(1988, 1, 2),
...             datetime(2023, 1, 3),
...             datetime(2022, 1, 1),
...             datetime(2022, 1, 1),
...         ],
...         "code": ["static", "DOB", "lab//A", "lab//B", "rx", "DOB", "lab//A", "lab//B", "dx//1"],
...     },
...     schema={"subject_id": pl.UInt32, "time": pl.Datetime, "code": pl.Utf8},
... )
>>> df
shape: (9, 3)
┌────────────┬─────────────────────┬────────┐
│ subject_id ┆ time                ┆ code   │
│ ---        ┆ ---                 ┆ ---    │
│ u32        ┆ datetime[μs]        ┆ str    │
╞════════════╪═════════════════════╪════════╡
│ 1          ┆ null                ┆ static │
│ 1          ┆ 1990-01-01 00:00:00 ┆ DOB    │
│ 1          ┆ 2021-01-01 00:00:00 ┆ lab//A │
│ 1          ┆ 2021-01-01 00:00:00 ┆ lab//B │
│ 1          ┆ 2021-01-02 00:00:00 ┆ rx     │
│ 2          ┆ 1988-01-02 00:00:00 ┆ DOB    │
│ 2          ┆ 2023-01-03 00:00:00 ┆ lab//A │
│ 3          ┆ 2022-01-01 00:00:00 ┆ lab//B │
│ 3          ┆ 2022-01-01 00:00:00 ┆ dx//1  │
└────────────┴─────────────────────┴────────┘
>>> import tempfile
>>> with tempfile.TemporaryDirectory() as tmpdir:
...     fp = f"{tmpdir}/quantile.parquet"
...     pl.DataFrame({"quantile": [25, 50, 75], "value": [1,2,3]}).write_parquet(fp)
...     time_delta_cfg = DictConfig({"quantile_fp": fp, "time_unit": "days", "max_length": 3})
...     time_delta_fn = time_delta_fntr(time_delta_cfg)
...     time_delta_fn(df)
shape: (6, 4)
┌────────────┬─────────────────────┬───────────────┬────────────────────┐
│ subject_id ┆ time                ┆ numeric_value ┆ code               │
│ ---        ┆ ---                 ┆ ---           ┆ ---                │
│ u32        ┆ datetime[μs]        ┆ f32           ┆ str                │
╞════════════╪═════════════════════╪═══════════════╪════════════════════╡
│ 1          ┆ 1990-01-01 00:00:00 ┆ null          ┆ TIME//START//TOKEN │
│ 1          ┆ 2021-01-01 00:00:00 ┆ 11323.0       ┆ TIME//DELTA//TOKEN │
│ 1          ┆ 2021-01-02 00:00:00 ┆ 1.0           ┆ TIME//DELTA//TOKEN │
│ 2          ┆ 1988-01-02 00:00:00 ┆ null          ┆ TIME//START//TOKEN │
│ 2          ┆ 2023-01-03 00:00:00 ┆ 12785.0       ┆ TIME//DELTA//TOKEN │
│ 3          ┆ 2022-01-01 00:00:00 ┆ null          ┆ TIME//START//TOKEN │
└────────────┴─────────────────────┴───────────────┴────────────────────┘
Source code in meds_torch/utils/custom_time_token.py 
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
def time_delta_fntr(cfg: DictConfig) -> Callable[[pl.DataFrame], pl.DataFrame]:
    """Create a function that adds time_delta event rows to a DataFrame.

    Args:
        cfg: The configuration for the age function. This must contain the following mandatory keys:
            - "quantile_fp": The path to a parquet file with pre-defined time delta quantiles.
                Should have a quantile column of type Int64 and a value column of type Float64
                indicating the numerical value for the quantile.
            - "max_length": The maximum number of time_delta qunatile tokens to use to approximate
                the continuous value.
            - "time_unit": The unit for the time deltas when dates are converted to a numeric value
                in the output data.

    Returns:
        A function that returns the to-be-added "time_delta" events with the quantile of the time delta
        for all input events with unique, non-null times in the data. The very first event for a subject
        has a null time_delta, so it is imputed with a special time start token.

    Raises:
        ValueError: If the input unit is not recognized.

    Examples:
        >>> from datetime import datetime
        >>> df = pl.DataFrame(
        ...     {
        ...         "subject_id": [1, 1, 1, 1, 1, 2, 2, 3, 3],
        ...         "time": [
        ...             None,
        ...             datetime(1990, 1, 1),
        ...             datetime(2021, 1, 1),
        ...             datetime(2021, 1, 1),
        ...             datetime(2021, 1, 2),
        ...             datetime(1988, 1, 2),
        ...             datetime(2023, 1, 3),
        ...             datetime(2022, 1, 1),
        ...             datetime(2022, 1, 1),
        ...         ],
        ...         "code": ["static", "DOB", "lab//A", "lab//B", "rx", "DOB", "lab//A", "lab//B", "dx//1"],
        ...     },
        ...     schema={"subject_id": pl.UInt32, "time": pl.Datetime, "code": pl.Utf8},
        ... )
        >>> df
        shape: (9, 3)
        ┌────────────┬─────────────────────┬────────┐
        │ subject_id ┆ time                ┆ code   │
        │ ---        ┆ ---                 ┆ ---    │
        │ u32        ┆ datetime[μs]        ┆ str    │
        ╞════════════╪═════════════════════╪════════╡
        │ 1          ┆ null                ┆ static │
        │ 1          ┆ 1990-01-01 00:00:00 ┆ DOB    │
        │ 1          ┆ 2021-01-01 00:00:00 ┆ lab//A │
        │ 1          ┆ 2021-01-01 00:00:00 ┆ lab//B │
        │ 1          ┆ 2021-01-02 00:00:00 ┆ rx     │
        │ 2          ┆ 1988-01-02 00:00:00 ┆ DOB    │
        │ 2          ┆ 2023-01-03 00:00:00 ┆ lab//A │
        │ 3          ┆ 2022-01-01 00:00:00 ┆ lab//B │
        │ 3          ┆ 2022-01-01 00:00:00 ┆ dx//1  │
        └────────────┴─────────────────────┴────────┘
        >>> import tempfile
        >>> with tempfile.TemporaryDirectory() as tmpdir:
        ...     fp = f"{tmpdir}/quantile.parquet"
        ...     pl.DataFrame({"quantile": [25, 50, 75], "value": [1,2,3]}).write_parquet(fp)
        ...     time_delta_cfg = DictConfig({"quantile_fp": fp, "time_unit": "days", "max_length": 3})
        ...     time_delta_fn = time_delta_fntr(time_delta_cfg)
        ...     time_delta_fn(df)
        shape: (6, 4)
        ┌────────────┬─────────────────────┬───────────────┬────────────────────┐
        │ subject_id ┆ time                ┆ numeric_value ┆ code               │
        │ ---        ┆ ---                 ┆ ---           ┆ ---                │
        │ u32        ┆ datetime[μs]        ┆ f32           ┆ str                │
        ╞════════════╪═════════════════════╪═══════════════╪════════════════════╡
        │ 1          ┆ 1990-01-01 00:00:00 ┆ null          ┆ TIME//START//TOKEN │
        │ 1          ┆ 2021-01-01 00:00:00 ┆ 11323.0       ┆ TIME//DELTA//TOKEN │
        │ 1          ┆ 2021-01-02 00:00:00 ┆ 1.0           ┆ TIME//DELTA//TOKEN │
        │ 2          ┆ 1988-01-02 00:00:00 ┆ null          ┆ TIME//START//TOKEN │
        │ 2          ┆ 2023-01-03 00:00:00 ┆ 12785.0       ┆ TIME//DELTA//TOKEN │
        │ 3          ┆ 2022-01-01 00:00:00 ┆ null          ┆ TIME//START//TOKEN │
        └────────────┴─────────────────────┴───────────────┴────────────────────┘
    """

    _, seconds_in_unit = normalize_time_unit(cfg.time_unit)
    microseconds_in_unit = int(1e6) * seconds_in_unit

    def fn(df: pl.LazyFrame) -> pl.LazyFrame:
        # Apply the function to create the quantile sequence
        assert "subject_id" in df.columns
        df = df.select("subject_id", "time").unique().sort(["subject_id", "time"]).drop_nulls()
        df = df.with_columns(
            (pl.col("time").diff().over("subject_id").dt.total_microseconds() / microseconds_in_unit)
            .alias("numeric_value")
            .cast(pl.Float32)
        )
        df = df.with_columns(
            pl.when(pl.col("numeric_value").is_null())
            .then(pl.lit(TIME_START_TOKEN))
            .otherwise(pl.lit(TIME_DELTA_TOKEN))
            .alias("code")
        )
        return df

    return fn