Skip to content

custom_tokenization

Functions for tokenizing MEDS datasets.

Here, tokenization refers specifically to the process of converting a longitudinal, irregularly sampled, continuous time sequence into a temporal sequence at the level that will be consumed by deep-learning models.

All these functions take in normalized data – meaning data where there are no longer any code modifiers, as those have been normalized alongside codes into integer indices (in the output code column). The only columns of concern here thus are subject_id, time, code, numeric_value.

extract_seq_of_subject_events(df)

This function extracts sequences of subject events, which are sequences of measurements.

The result of this can be naturally tensorized into a JointNestedRaggedTensorDict object.

Parameters:

Name Type Description Default
df LazyFrame

The input data.

 required

Returns:

Type Description
LazyFrame

A pl.LazyFrame object containing the sequences of subject events, with the following columns: - subject_id: The subject ID. - time_delta_days: The time delta in days, as a list of floats (ragged). - code: The code, as a list of lists of ints (ragged in both levels). - numeric_value: The numeric value as a list of lists of floats (ragged in both levels).

Examples:

>>> from datetime import datetime
>>> df = pl.DataFrame({
...     "subject_id": [1, 1, 1, 1, 2, 2, 2],
...     "time": [
...         None, datetime(2021, 1, 1), datetime(2021, 1, 1), datetime(2021, 1, 13),
...         None, datetime(2021, 1, 2), datetime(2021, 1, 2)],
...     "code": [100, 101, 102, 103, 200, 201, 202],
...     "numeric_value": pl.Series([1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0], dtype=pl.Float32)
... }).lazy()
>>> extract_seq_of_subject_events(df).collect()
shape: (2, 4)
┌────────────┬─────────────────┬─────────────────────┬─────────────────────┐
│ subject_id ┆ time_delta_days ┆ code                ┆ numeric_value       │
│ ---        ┆ ---             ┆ ---                 ┆ ---                 │
│ i64        ┆ list[f32]       ┆ list[list[i64]]     ┆ list[list[f32]]     │
╞════════════╪═════════════════╪═════════════════════╪═════════════════════╡
│ 1          ┆ [NaN, 12.0]     ┆ [[101, 102], [103]] ┆ [[2.0, 3.0], [4.0]] │
│ 2          ┆ [NaN]           ┆ [[201, 202]]        ┆ [[6.0, 7.0]]        │
└────────────┴─────────────────┴─────────────────────┴─────────────────────┘
Source code in meds_torch/utils/custom_tokenization.py 
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
def extract_seq_of_subject_events(df: pl.LazyFrame) -> pl.LazyFrame:
    """This function extracts sequences of subject events, which are sequences of measurements.

    The result of this can be naturally tensorized into a `JointNestedRaggedTensorDict` object.

    Args:
        df: The input data.

    Returns:
        A `pl.LazyFrame` object containing the sequences of subject events, with the following columns:
            - `subject_id`: The subject ID.
            - `time_delta_days`: The time delta in days, as a list of floats (ragged).
            - `code`: The code, as a list of lists of ints (ragged in both levels).
            - `numeric_value`: The numeric value as a list of lists of floats (ragged in both levels).

    Examples:
        >>> from datetime import datetime
        >>> df = pl.DataFrame({
        ...     "subject_id": [1, 1, 1, 1, 2, 2, 2],
        ...     "time": [
        ...         None, datetime(2021, 1, 1), datetime(2021, 1, 1), datetime(2021, 1, 13),
        ...         None, datetime(2021, 1, 2), datetime(2021, 1, 2)],
        ...     "code": [100, 101, 102, 103, 200, 201, 202],
        ...     "numeric_value": pl.Series([1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0], dtype=pl.Float32)
        ... }).lazy()
        >>> extract_seq_of_subject_events(df).collect()
        shape: (2, 4)
        ┌────────────┬─────────────────┬─────────────────────┬─────────────────────┐
        │ subject_id ┆ time_delta_days ┆ code                ┆ numeric_value       │
        │ ---        ┆ ---             ┆ ---                 ┆ ---                 │
        │ i64        ┆ list[f32]       ┆ list[list[i64]]     ┆ list[list[f32]]     │
        ╞════════════╪═════════════════╪═════════════════════╪═════════════════════╡
        │ 1          ┆ [NaN, 12.0]     ┆ [[101, 102], [103]] ┆ [[2.0, 3.0], [4.0]] │
        │ 2          ┆ [NaN]           ┆ [[201, 202]]        ┆ [[6.0, 7.0]]        │
        └────────────┴─────────────────┴─────────────────────┴─────────────────────┘
    """

    _, dynamic = split_static_and_dynamic(df)

    time_delta_days_expr = (pl.col("time").diff().dt.total_seconds() / SECONDS_PER_DAY).cast(pl.Float32)

    return (
        dynamic.group_by("subject_id", "time", maintain_order=True)
        .agg(pl.col("code").name.keep(), fill_to_nans("numeric_value").name.keep())
        .group_by("subject_id", maintain_order=True)
        .agg(
            fill_to_nans(time_delta_days_expr).alias("time_delta_days"),
            "code",
            "numeric_value",
        )
    )

extract_statics_and_schema(df)

This function extracts static data and schema information (sequence of subject unique times).

Parameters:

Name Type Description Default
df LazyFrame

The input data.

 required

Returns:

Type Description
LazyFrame

A pl.LazyFrame object containing the static data and the unique times of the subject, grouped
LazyFrame

by subject as lists, in the same order as the subject IDs occurred in the original file.

Examples:

>>> from datetime import datetime
>>> df = pl.DataFrame({
...     "subject_id": [1, 1, 1, 1, 2, 2, 2],
...     "time": [
...         None, datetime(2021, 1, 1), datetime(2021, 1, 1), datetime(2021, 1, 13),
...         None, datetime(2021, 1, 2), datetime(2021, 1, 2)],
...     "code": [100, 101, 102, 103, 200, 201, 202],
...     "numeric_value": [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0]
... }).lazy()
>>> df = extract_statics_and_schema(df).collect()
>>> df.drop("time")
shape: (2, 4)
┌────────────┬───────────┬───────────────┬─────────────────────┐
│ subject_id ┆ code      ┆ numeric_value ┆ start_time          │
│ ---        ┆ ---       ┆ ---           ┆ ---                 │
│ i64        ┆ list[i64] ┆ list[f64]     ┆ datetime[μs]        │
╞════════════╪═══════════╪═══════════════╪═════════════════════╡
│ 1          ┆ [100]     ┆ [1.0]         ┆ 2021-01-01 00:00:00 │
│ 2          ┆ [200]     ┆ [5.0]         ┆ 2021-01-02 00:00:00 │
└────────────┴───────────┴───────────────┴─────────────────────┘
>>> df.select("subject_id", "time").explode("time")
shape: (3, 2)
┌────────────┬─────────────────────┐
│ subject_id ┆ time                │
│ ---        ┆ ---                 │
│ i64        ┆ datetime[μs]        │
╞════════════╪═════════════════════╡
│ 1          ┆ 2021-01-01 00:00:00 │
│ 1          ┆ 2021-01-13 00:00:00 │
│ 2          ┆ 2021-01-02 00:00:00 │
└────────────┴─────────────────────┘
>>> df = pl.DataFrame({
...     "subject_id": [1, 1, 1, 1, 2, 2, 2],
...     "time": [
...         datetime(2020, 1, 1), datetime(2021, 1, 1), datetime(2021, 1, 1), datetime(2021, 1, 13),
...         datetime(2020, 1, 1), datetime(2021, 1, 2), datetime(2021, 1, 2)],
...     "code": [100, 101, 102, 103, 200, 201, 202],
...     "numeric_value": [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0]
... }).lazy()
>>> df = extract_statics_and_schema(df).collect()
>>> df.drop("time")
shape: (2, 4)
┌────────────┬───────────┬───────────────┬─────────────────────┐
│ subject_id ┆ code      ┆ numeric_value ┆ start_time          │
│ ---        ┆ ---       ┆ ---           ┆ ---                 │
│ i64        ┆ list[i64] ┆ list[f64]     ┆ datetime[μs]        │
╞════════════╪═══════════╪═══════════════╪═════════════════════╡
│ 1          ┆ null      ┆ null          ┆ 2020-01-01 00:00:00 │
│ 2          ┆ null      ┆ null          ┆ 2020-01-01 00:00:00 │
└────────────┴───────────┴───────────────┴─────────────────────┘
>>> df.select("subject_id", "time").explode("time")
shape: (5, 2)
┌────────────┬─────────────────────┐
│ subject_id ┆ time                │
│ ---        ┆ ---                 │
│ i64        ┆ datetime[μs]        │
╞════════════╪═════════════════════╡
│ 1          ┆ 2020-01-01 00:00:00 │
│ 1          ┆ 2021-01-01 00:00:00 │
│ 1          ┆ 2021-01-13 00:00:00 │
│ 2          ┆ 2020-01-01 00:00:00 │
│ 2          ┆ 2021-01-02 00:00:00 │
└────────────┴─────────────────────┘
Source code in meds_torch/utils/custom_tokenization.py 
106
107
108
109
110
111
112
113
114
115
116
117
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
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
def extract_statics_and_schema(df: pl.LazyFrame) -> pl.LazyFrame:
    """This function extracts static data and schema information (sequence of subject unique times).

    Args:
        df: The input data.

    Returns:
        A `pl.LazyFrame` object containing the static data and the unique times of the subject, grouped
        by subject as lists, in the same order as the subject IDs occurred in the original file.

    Examples:
        >>> from datetime import datetime
        >>> df = pl.DataFrame({
        ...     "subject_id": [1, 1, 1, 1, 2, 2, 2],
        ...     "time": [
        ...         None, datetime(2021, 1, 1), datetime(2021, 1, 1), datetime(2021, 1, 13),
        ...         None, datetime(2021, 1, 2), datetime(2021, 1, 2)],
        ...     "code": [100, 101, 102, 103, 200, 201, 202],
        ...     "numeric_value": [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0]
        ... }).lazy()
        >>> df = extract_statics_and_schema(df).collect()
        >>> df.drop("time")
        shape: (2, 4)
        ┌────────────┬───────────┬───────────────┬─────────────────────┐
        │ subject_id ┆ code      ┆ numeric_value ┆ start_time          │
        │ ---        ┆ ---       ┆ ---           ┆ ---                 │
        │ i64        ┆ list[i64] ┆ list[f64]     ┆ datetime[μs]        │
        ╞════════════╪═══════════╪═══════════════╪═════════════════════╡
        │ 1          ┆ [100]     ┆ [1.0]         ┆ 2021-01-01 00:00:00 │
        │ 2          ┆ [200]     ┆ [5.0]         ┆ 2021-01-02 00:00:00 │
        └────────────┴───────────┴───────────────┴─────────────────────┘
        >>> df.select("subject_id", "time").explode("time")
        shape: (3, 2)
        ┌────────────┬─────────────────────┐
        │ subject_id ┆ time                │
        │ ---        ┆ ---                 │
        │ i64        ┆ datetime[μs]        │
        ╞════════════╪═════════════════════╡
        │ 1          ┆ 2021-01-01 00:00:00 │
        │ 1          ┆ 2021-01-13 00:00:00 │
        │ 2          ┆ 2021-01-02 00:00:00 │
        └────────────┴─────────────────────┘
        >>> df = pl.DataFrame({
        ...     "subject_id": [1, 1, 1, 1, 2, 2, 2],
        ...     "time": [
        ...         datetime(2020, 1, 1), datetime(2021, 1, 1), datetime(2021, 1, 1), datetime(2021, 1, 13),
        ...         datetime(2020, 1, 1), datetime(2021, 1, 2), datetime(2021, 1, 2)],
        ...     "code": [100, 101, 102, 103, 200, 201, 202],
        ...     "numeric_value": [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0]
        ... }).lazy()
        >>> df = extract_statics_and_schema(df).collect()
        >>> df.drop("time")
        shape: (2, 4)
        ┌────────────┬───────────┬───────────────┬─────────────────────┐
        │ subject_id ┆ code      ┆ numeric_value ┆ start_time          │
        │ ---        ┆ ---       ┆ ---           ┆ ---                 │
        │ i64        ┆ list[i64] ┆ list[f64]     ┆ datetime[μs]        │
        ╞════════════╪═══════════╪═══════════════╪═════════════════════╡
        │ 1          ┆ null      ┆ null          ┆ 2020-01-01 00:00:00 │
        │ 2          ┆ null      ┆ null          ┆ 2020-01-01 00:00:00 │
        └────────────┴───────────┴───────────────┴─────────────────────┘
        >>> df.select("subject_id", "time").explode("time")
        shape: (5, 2)
        ┌────────────┬─────────────────────┐
        │ subject_id ┆ time                │
        │ ---        ┆ ---                 │
        │ i64        ┆ datetime[μs]        │
        ╞════════════╪═════════════════════╡
        │ 1          ┆ 2020-01-01 00:00:00 │
        │ 1          ┆ 2021-01-01 00:00:00 │
        │ 1          ┆ 2021-01-13 00:00:00 │
        │ 2          ┆ 2020-01-01 00:00:00 │
        │ 2          ┆ 2021-01-02 00:00:00 │
        └────────────┴─────────────────────┘
    """

    static, dynamic = split_static_and_dynamic(df)

    # This collects static data by subject ID and stores only (as a list) the codes and numeric values.
    static_by_subject = static.group_by("subject_id", maintain_order=True).agg("code", "numeric_value")

    # This collects the unique times for each subject.
    schema_by_subject = dynamic.group_by("subject_id", maintain_order=True).agg(
        pl.col("time").min().alias("start_time"), pl.col("time").unique(maintain_order=True)
    )

    # TODO(mmd): Consider tracking subject offset explicitly here.

    return static_by_subject.join(schema_by_subject, on="subject_id", how="full", coalesce=True)

fill_to_nans(col)

This function fills infinite and null values with NaN.

This enables the downstream functions to naturally tensorize data into numpy or Torch tensors.

Parameters:

Name Type Description Default
col str | Expr

The input column.

 required

Returns:

Type Description
Expr

A pl.Expr object that fills infinite and null values with NaN.

Examples:

>>> print(fill_to_nans("value"))
.when([(col("value").is_infinite()) |
       (col("value").is_null())]).then(dyn float: NaN).otherwise(col("value"))
>>> print(fill_to_nans(pl.col("time_delta")))
.when([(col("time_delta").is_infinite()) |
       (col("time_delta").is_null())]).then(dyn float: NaN).otherwise(col("time_delta"))
>>> df = pl.DataFrame({"value": [1.0, float("inf"), None, -float("inf"), 2.0]})
>>> df.select(fill_to_nans("value").alias("value"))["value"].to_list()
[1.0, nan, nan, nan, 2.0]
Source code in meds_torch/utils/custom_tokenization.py 
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
def fill_to_nans(col: str | pl.Expr) -> pl.Expr:
    """This function fills infinite and null values with NaN.

    This enables the downstream functions to naturally tensorize data into numpy or Torch tensors.

    Args:
        col: The input column.

    Returns:
        A `pl.Expr` object that fills infinite and null values with NaN.

    Examples:
        >>> print(fill_to_nans("value")) # doctest: +NORMALIZE_WHITESPACE
        .when([(col("value").is_infinite()) |
               (col("value").is_null())]).then(dyn float: NaN).otherwise(col("value"))
        >>> print(fill_to_nans(pl.col("time_delta"))) # doctest: +NORMALIZE_WHITESPACE
        .when([(col("time_delta").is_infinite()) |
               (col("time_delta").is_null())]).then(dyn float: NaN).otherwise(col("time_delta"))
        >>> df = pl.DataFrame({"value": [1.0, float("inf"), None, -float("inf"), 2.0]})
        >>> df.select(fill_to_nans("value").alias("value"))["value"].to_list()
        [1.0, nan, nan, nan, 2.0]
    """

    if isinstance(col, str):
        col = pl.col(col)

    return pl.when(col.is_infinite() | col.is_null()).then(float("nan")).otherwise(col)

main(cfg)

TODO.

Source code in meds_torch/utils/custom_tokenization.py 
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
@hydra.main(
    version_base=None, config_path=str(PREPROCESS_CONFIG_YAML.parent), config_name=PREPROCESS_CONFIG_YAML.stem
)
def main(cfg: DictConfig):
    """TODO."""

    logger.info(
        f"Running with config:\n{OmegaConf.to_yaml(cfg)}\n"
        f"Stage: {cfg.stage}\n\n"
        f"Stage config:\n{OmegaConf.to_yaml(cfg.stage_cfg)}"
    )

    output_dir = Path(cfg.stage_cfg.output_dir)
    if train_only := cfg.stage_cfg.get("train_only", False):
        raise ValueError(f"train_only={train_only} is not supported for this stage.")
    shards_single_output, include_only_train = shard_iterator(cfg)

    for in_fp, out_fp in shards_single_output:
        sharded_path = out_fp.relative_to(output_dir)

        schema_out_fp = output_dir / "schemas" / sharded_path
        event_seq_out_fp = output_dir / "event_seqs" / sharded_path

        logger.info(f"Tokenizing {str(in_fp.resolve())} into schemas at {str(schema_out_fp.resolve())}")

        rwlock_wrap(
            in_fp,
            schema_out_fp,
            pl.scan_parquet,
            write_lazyframe,
            extract_statics_and_schema,
            do_overwrite=cfg.do_overwrite,
        )

        logger.info(f"Tokenizing {str(in_fp.resolve())} into event_seqs at {str(event_seq_out_fp.resolve())}")

        rwlock_wrap(
            in_fp,
            event_seq_out_fp,
            pl.scan_parquet,
            write_lazyframe,
            extract_seq_of_subject_events,
            do_overwrite=cfg.do_overwrite,
        )

    logger.info(f"Done with {cfg.stage}")

split_static_and_dynamic(df)

This function splits the input data into static and dynamic data.

Static data is data that has a null time, and dynamic data is everything else.

Parameters:

Name Type Description Default
df LazyFrame

The input data.

 required

Returns:

Type Description
LazyFrame

A tuple of two pl.LazyFrame objects, the first being the static data and the second being the
LazyFrame

dynamic data.

Examples:

>>> from datetime import datetime
>>> df = pl.DataFrame({
...     "subject_id": [1, 1, 2, 2],
...     "time": [None, datetime(2021, 1, 1), None, datetime(2021, 1, 2)],
...     "code": [100, 101, 200, 201],
...     "numeric_value": [1.0, 2.0, 3.0, 4.0]
... }).lazy()
>>> static, dynamic = split_static_and_dynamic(df)
>>> static.collect()
shape: (2, 3)
┌────────────┬──────┬───────────────┐
│ subject_id ┆ code ┆ numeric_value │
│ ---        ┆ ---  ┆ ---           │
│ i64        ┆ i64  ┆ f64           │
╞════════════╪══════╪═══════════════╡
│ 1          ┆ 100  ┆ 1.0           │
│ 2          ┆ 200  ┆ 3.0           │
└────────────┴──────┴───────────────┘
>>> dynamic.collect()
shape: (2, 4)
┌────────────┬─────────────────────┬──────┬───────────────┐
│ subject_id ┆ time                ┆ code ┆ numeric_value │
│ ---        ┆ ---                 ┆ ---  ┆ ---           │
│ i64        ┆ datetime[μs]        ┆ i64  ┆ f64           │
╞════════════╪═════════════════════╪══════╪═══════════════╡
│ 1          ┆ 2021-01-01 00:00:00 ┆ 101  ┆ 2.0           │
│ 2          ┆ 2021-01-02 00:00:00 ┆ 201  ┆ 4.0           │
└────────────┴─────────────────────┴──────┴───────────────┘
Source code in meds_torch/utils/custom_tokenization.py 
 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
def split_static_and_dynamic(df: pl.LazyFrame) -> tuple[pl.LazyFrame, pl.LazyFrame]:
    """This function splits the input data into static and dynamic data.

    Static data is data that has a null time, and dynamic data is everything else.

    Args:
        df: The input data.

    Returns:
        A tuple of two `pl.LazyFrame` objects, the first being the static data and the second being the
        dynamic data.

    Examples:
        >>> from datetime import datetime
        >>> df = pl.DataFrame({
        ...     "subject_id": [1, 1, 2, 2],
        ...     "time": [None, datetime(2021, 1, 1), None, datetime(2021, 1, 2)],
        ...     "code": [100, 101, 200, 201],
        ...     "numeric_value": [1.0, 2.0, 3.0, 4.0]
        ... }).lazy()
        >>> static, dynamic = split_static_and_dynamic(df)
        >>> static.collect()
        shape: (2, 3)
        ┌────────────┬──────┬───────────────┐
        │ subject_id ┆ code ┆ numeric_value │
        │ ---        ┆ ---  ┆ ---           │
        │ i64        ┆ i64  ┆ f64           │
        ╞════════════╪══════╪═══════════════╡
        │ 1          ┆ 100  ┆ 1.0           │
        │ 2          ┆ 200  ┆ 3.0           │
        └────────────┴──────┴───────────────┘
        >>> dynamic.collect()
        shape: (2, 4)
        ┌────────────┬─────────────────────┬──────┬───────────────┐
        │ subject_id ┆ time                ┆ code ┆ numeric_value │
        │ ---        ┆ ---                 ┆ ---  ┆ ---           │
        │ i64        ┆ datetime[μs]        ┆ i64  ┆ f64           │
        ╞════════════╪═════════════════════╪══════╪═══════════════╡
        │ 1          ┆ 2021-01-01 00:00:00 ┆ 101  ┆ 2.0           │
        │ 2          ┆ 2021-01-02 00:00:00 ┆ 201  ┆ 4.0           │
        └────────────┴─────────────────────┴──────┴───────────────┘
    """

    static = df.filter(pl.col("time").is_null()).drop("time")
    dynamic = df.filter(pl.col("time").is_not_null())
    return static, dynamic