genhpf 1.0.0__py3-none-any.whl

genhpf/scripts/preprocess/preprocess_meds.py

@@ -0,0 +1,584 @@
+import functools
+import glob
+import multiprocessing
+import os
+import re
+import shutil
+import warnings
+from argparse import ArgumentParser
+from bisect import bisect_left, bisect_right
+from datetime import datetime
+from pathlib import Path
+
+import h5py
+import numpy as np
+import pandas as pd
+import polars as pl
+from tqdm import tqdm
+from transformers import AutoTokenizer
+
+pool_manager = multiprocessing.Manager()
+warned_codes = pool_manager.list()
+
+
+def find_boundary_between(tuples_list, start, end):
+    starts = [s for s, e in tuples_list]
+    ends = [e for s, e in tuples_list]
+
+    start_index = bisect_left(starts, start)
+    end_index = bisect_right(ends, end)
+    assert start_index < end_index
+
+    return start_index, end_index
+
+
+def get_parser():
+    parser = ArgumentParser()
+    parser.add_argument(
+        "root",
+        help="path to MEDS dataset. it can be either of directory or the exact file path "
+        "with the file extension. if provided with directory, it tries to scan *.csv or "
+        "*.parquet files contained in the directory, including sub-directories, to process "
+        "all of them.",
+    )
+    parser.add_argument(
+        "--metadata_dir",
+        help="path to metadata directory for the input MEDS dataset, which contains codes.parquet",
+    )
+
+    parser.add_argument(
+        "--birth_code", type=str, default="MEDS_BIRTH", help="string code for the birth event in the dataset."
+    )
+
+    parser.add_argument(
+        "--cohort",
+        type=str,
+        help="path to the defined cohort, which must be a result of ACES. it can be either of "
+        "directory or the exact file path that has the same extension with the MEDS dataset "
+        "to be processed. the file structure of this cohort directory should be the same with "
+        "the provided MEDS dataset directory to match each cohort to its corresponding shard "
+        "data.",
+    )
+    parser.add_argument(
+        "--cohort_label_name",
+        type=str,
+        default="boolean_value",
+        help="column name in the cohort dataframe to be used for label",
+    )
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        default="outputs",
+        help="directory to save processed outputs.",
+    )
+    parser.add_argument(
+        "--rebase",
+        action="store_true",
+        help="whether or not to rebase the output directory if exists.",
+    )
+    parser.add_argument(
+        "--workers",
+        metavar="N",
+        default=1,
+        type=int,
+        help="number of parallel workers.",
+    )
+    parser.add_argument(
+        "--max_event_length",
+        metavar="N",
+        default=128,
+        type=int,
+        help="maximum number of tokens in an event.",
+    )
+
+    parser.add_argument(
+        "--mimic_dir",
+        default=None,
+        help="path to directory for MIMIC-IV database containing hosp/ and icu/ as a subdirectory. "
+        "this is used for addressing missing descriptions in the metadata for MIMIC-IV codes.",
+    )
+
+    return parser
+
+
+def main(args):
+    root_path = Path(args.root)
+    output_dir = Path(args.output_dir)
+    metadata_dir = Path(args.metadata_dir)
+    mimic_dir = Path(args.mimic_dir) if args.mimic_dir is not None else None
+
+    if not output_dir.exists():
+        output_dir.mkdir()
+    else:
+        if args.rebase:
+            shutil.rmtree(output_dir)
+        if output_dir.exists():
+            raise ValueError(
+                f"File exists: '{str(output_dir.resolve())}'. If you want to rebase the "
+                "directory, please run the script with --rebase."
+            )
+        output_dir.mkdir()
+
+    if root_path.is_dir():
+        data_paths = glob.glob(str(root_path / "**/*.csv"), recursive=True)
+        if len(data_paths) == 0:
+            data_paths = glob.glob(str(root_path / "**/*.parquet"), recursive=True)
+        if len(data_paths) == 0:
+            raise ValueError("Data directory does not contain any supported file formats: .csv or .parquet")
+    else:
+        data_paths = [root_path]
+
+    label_col_name = args.cohort_label_name
+
+    tokenizer = AutoTokenizer.from_pretrained("emilyalsentzer/Bio_ClinicalBERT")
+
+    codes_metadata = pl.read_parquet(metadata_dir / "codes.parquet").to_pandas()
+    codes_metadata = codes_metadata.set_index("code")["description"].to_dict()
+    # do not allow to use static events or birth event
+    birth_code = args.birth_code
+    # if birth_code not in codes_metadata:
+    #     print(
+    #         f'"{birth_code}" is not found in the codes metadata, which may lead to '
+    #         "unexpected results since we currently exclude this event from the input data. "
+    #     )
+
+    if mimic_dir is not None:
+        d_items = pd.read_csv(mimic_dir / "icu" / "d_items.csv.gz")
+        d_items["itemid"] = d_items["itemid"].astype("str")
+        d_items = d_items.set_index("itemid")["label"].to_dict()
+        d_labitems = pd.read_csv(mimic_dir / "hosp" / "d_labitems.csv.gz")
+        d_labitems["itemid"] = d_labitems["itemid"].astype("str")
+        d_labitems = d_labitems.set_index("itemid")["label"].to_dict()
+    else:
+        d_items = None
+        d_labitems = None
+
+    max_event_length = args.max_event_length
+
+    progress_bar = tqdm(data_paths, total=len(data_paths))
+    for data_path in progress_bar:
+        progress_bar.set_description(str(data_path))
+
+        data_path = Path(data_path)
+        subdir = data_path.relative_to(root_path).parent
+        if data_path.suffix == ".csv":
+            data = pl.scan_csv(data_path)
+        elif data_path.suffix == ".parquet":
+            data = pl.scan_parquet(data_path)
+        else:
+            raise ValueError(f"Unsupported file format: {data_path.suffix}")
+
+        data = data.with_columns(
+            pl.when(pl.col("code") == birth_code).then(None).otherwise(pl.col("time")).alias("time")
+        )
+        data = data.drop_nulls(subset=["subject_id", "time"])
+
+        cohort_path = Path(args.cohort) / subdir / data_path.name
+
+        if cohort_path.suffix == ".csv":
+            cohort = pl.scan_csv(cohort_path)
+        elif cohort_path.suffix == ".parquet":
+            cohort = pl.scan_parquet(cohort_path)
+        else:
+            raise ValueError(f"Unsupported file format: {cohort_path.suffix}")
+
+        cohort = cohort.drop_nulls(label_col_name)
+        cohort = cohort.unique()
+
+        cohort = cohort.select(
+            [
+                pl.col("subject_id"),
+                pl.col(label_col_name),
+                # pl.col("input.end_summary").struct.field("timestamp_at_start").alias("starttime"),
+                pl.col("prediction_time").alias("endtime"),
+            ]
+        )
+        cohort = (
+            cohort.group_by("subject_id", maintain_order=True)
+            .agg(pl.col(["endtime", label_col_name]))
+            .collect()
+        )  # omitted "starttime"
+        cohort_dict = {
+            x["subject_id"]: {
+                # "starttime": x["starttime"],
+                "endtime": x["endtime"],
+                "label": x[label_col_name],
+            }
+            for x in cohort.iter_rows(named=True)
+        }
+
+        def extract_cohort(row):
+            subject_id = row["subject_id"]
+            time = row["time"]
+            if subject_id not in cohort_dict:
+                # return {"cohort_start": None, "cohort_end": None, "cohort_label": None}
+                return {"cohort_end": None, "cohort_label": None}
+
+            cohort_criteria = cohort_dict[subject_id]
+            # starts = cohort_criteria["starttime"]
+            ends = cohort_criteria["endtime"]
+            labels = cohort_criteria["label"]
+
+            # for start, end, label in zip(starts, ends, labels):
+            #     if start <= time and time <= end:
+            #         return {"cohort_start": start, "cohort_end": end, "cohort_label": label}
+
+            # assume it is possible that each event goes into multiple different cohorts
+            cohort_ends = []
+            cohort_labels = []
+            for end, label in zip(ends, labels):
+                if time <= end:
+                    # return {"cohort_start": start, "cohort_end": end, "cohort_label": label}
+                    cohort_ends.append(end)
+                    cohort_labels.append(label)
+
+            if len(cohort_ends) > 0:
+                return {"cohort_end": cohort_ends, "cohort_label": cohort_labels}
+            else:
+                # return {"cohort_start": None, "cohort_end": None, "cohort_label": None}
+                return {"cohort_end": None, "cohort_label": None}
+
+        data = data.group_by(["subject_id", "time"], maintain_order=True).agg(pl.all())
+        data = (
+            data.with_columns(
+                pl.struct(["subject_id", "time"])
+                .map_elements(
+                    extract_cohort,
+                    return_dtype=pl.Struct(
+                        {
+                            "cohort_end": pl.List(pl.Datetime()),
+                            "cohort_label": pl.List(pl.Boolean),
+                        }
+                    ),
+                )
+                .alias("cohort_criteria")
+            )
+            .unnest("cohort_criteria")
+            .collect()
+        )
+
+        data = data.drop_nulls("cohort_label")
+
+        data = data.with_columns(pl.col("time").dt.strftime("%Y-%m-%d %H:%M:%S").cast(pl.List(str)))
+        data = data.with_columns(
+            pl.col("time").list.sample(n=pl.col("code").list.len(), with_replacement=True)
+        )
+
+        if str(subdir) != ".":
+            output_name = str(subdir)
+        else:
+            output_name = data_path.stem
+
+        if not os.path.exists(output_dir / output_name):
+            os.makedirs(output_dir / output_name)
+
+        with open(str(output_dir / (output_name + ".tsv")), "a") as manifest_f:
+            if os.path.getsize(output_dir / (output_name + ".tsv")) == 0:
+                manifest_f.write(f"{output_dir}/{output_name}\n")
+
+            must_have_columns = [
+                "subject_id",
+                "cohort_end",
+                "cohort_label",
+                "time",
+                "code",
+                "numeric_value",
+            ]
+            rest_of_columns = [x for x in data.columns if x not in must_have_columns]
+            column_name_idcs = {col: i for i, col in enumerate(data.columns)}
+
+            meds_to_remed_partial = functools.partial(
+                meds_to_remed,
+                tokenizer,
+                rest_of_columns,
+                column_name_idcs,
+                codes_metadata,
+                output_dir,
+                output_name,
+                args.workers,
+                d_items,
+                d_labitems,
+                warned_codes,
+                max_event_length,
+            )
+
+            # meds --> remed
+            print("Processing...")
+            if args.workers <= 1:
+                length_per_subject_gathered = [meds_to_remed_partial(data)]
+                del data
+            else:
+                subject_ids = data["subject_id"].unique().to_list()
+                n = args.workers
+                subject_id_chunks = [subject_ids[i::n] for i in range(n)]
+                data_chunks = []
+                for subject_id_chunk in subject_id_chunks:
+                    data_chunks.append(data.filter(pl.col("subject_id").is_in(subject_id_chunk)))
+                del data
+                pool = multiprocessing.get_context("spawn").Pool(processes=args.workers)
+                # the order is preserved
+                length_per_subject_gathered = pool.map(meds_to_remed_partial, data_chunks)
+                pool.close()
+                pool.join()
+                del data_chunks
+
+            if len(length_per_subject_gathered) != args.workers:
+                print(
+                    "Number of processed workers were smaller than the specified num workers "
+                    "(--workers) due to the small size of data. Consider reducing the number of "
+                    "workers."
+                )
+
+            for length_per_subject in length_per_subject_gathered:
+                for subject_id, (length, shard_id) in length_per_subject.items():
+                    manifest_f.write(f"{subject_id}\t{length}\t{shard_id}\n")
+
+
+def meds_to_remed(
+    tokenizer,
+    rest_of_columns,
+    column_name_idcs,
+    codes_metadata,
+    output_dir,
+    output_name,
+    num_shards,
+    d_items,
+    d_labitems,
+    warned_codes,
+    max_event_length,
+    df_chunk,
+):
+    code_matching_pattern = re.compile(r"\d+")
+
+    def meds_to_remed_unit(row):
+        events = []
+        digit_offsets = []
+        col_name_offsets = []
+        for event_index in range(len(row[column_name_idcs["code"]])):
+            event = ""
+            digit_offset = []
+            col_name_offset = []
+            for col_name in ["code", "numeric_value"] + rest_of_columns:
+                # do not process something like "icustay_id" or "hadm_id"
+                if "id" in col_name:
+                    continue
+
+                col_event = row[column_name_idcs[col_name]][event_index]
+                if col_event is not None:
+                    col_event = str(col_event)
+                    if col_name == "code":
+                        if col_event in codes_metadata and codes_metadata[col_event] != "":
+                            col_event = codes_metadata[col_event]
+                        else:
+                            do_break = False
+                            items = col_event.split("//")
+                            is_code = [bool(code_matching_pattern.fullmatch(item)) for item in items]
+                            if True in is_code:
+                                if d_items is not None and d_labitems is not None:
+                                    code_idx = is_code.index(True)
+                                    code = items[code_idx]
+
+                                    if code in d_items:
+                                        desc = d_items[code]
+                                    elif code in d_labitems:
+                                        desc = d_labitems[code]
+                                    else:
+                                        do_break = True
+
+                                    if not do_break:
+                                        items[code_idx] = desc
+                                        col_event = "//".join(items)
+                                else:
+                                    do_break = True
+
+                            if do_break and col_event not in warned_codes:
+                                warned_codes.append(col_event)
+                                warnings.warn(
+                                    "The dataset contains some codes that are not specified in "
+                                    "the codes metadata, which may not be intended. Note that we "
+                                    f"process this code as it is for now: {col_event}."
+                                )
+                    else:
+                        col_event = re.sub(
+                            r"\d*\.\d+",
+                            lambda x: str(round(float(x.group(0)), 4)),
+                            col_event,
+                        )
+                        event_offset = len(event) + len(col_name) + 1
+                        digit_offset_tmp = [
+                            g.span() for g in re.finditer(r"([0-9]+([.][0-9]*)?|[0-9]+|\.+)", col_event)
+                        ]
+
+                        internal_offset = 0
+                        for start, end in digit_offset_tmp:
+                            digit_offset.append(
+                                (
+                                    event_offset + start + internal_offset,
+                                    event_offset + end + (end - start) * 2 + internal_offset,
+                                )
+                            )
+                            internal_offset += (end - start) * 2
+
+                        col_event = re.sub(r"([0-9\.])", r" \1 ", col_event)
+
+                    col_name_offset.append((len(event), len(event) + len(col_name)))
+                    event += " " + col_name + " " + col_event
+            if len(event) > 0:
+                events.append(event[1:])
+                digit_offsets.append(digit_offset)
+                col_name_offsets.append(col_name_offset)
+
+        tokenized_events = tokenizer(
+            events,
+            add_special_tokens=True,
+            padding="max_length",
+            max_length=max_event_length,
+            truncation=True,
+            return_tensors="np",
+            return_token_type_ids=False,
+            return_attention_mask=True,
+            return_offsets_mapping=True,
+        )
+        lengths_before_padding = tokenized_events["attention_mask"].sum(axis=1)
+
+        input_ids = tokenized_events["input_ids"]
+        dpe_ids = np.zeros(input_ids.shape, dtype=int)
+        for i, digit_offset in enumerate(digit_offsets):
+            for start, end in digit_offset:
+                start_index, end_index = find_boundary_between(
+                    tokenized_events[i].offsets[: lengths_before_padding[i] - 1],
+                    start,
+                    end,
+                )
+
+                # define dpe ids for digits found
+                num_digits = end_index - start_index
+                # 119: token id for "."
+                num_decimal_points = (input_ids[i][start_index:end_index] == 119).sum()
+
+                # integer without decimal point
+                # e.g., for "1 2 3 4 5", assign [10, 9, 8, 7, 6]
+                if num_decimal_points == 0:
+                    dpe_ids[i][start_index:end_index] = list(range(num_digits + 5, 5, -1))
+                # floats
+                # e.g., for "1 2 3 4 5 . 6 7 8 9", assign [10, 9, 8, 7, 6, 5, 4, 3, 2, 1]
+                elif num_decimal_points == 1:
+                    num_decimals = (
+                        num_digits
+                        - np.where(input_ids[i][start_index:end_index] == 119)[0][0]  # 119: token id for "."
+                    )
+                    dpe_ids[i][start_index:end_index] = list(
+                        range(num_digits + 5 - num_decimals, 5 - num_decimals, -1)
+                    )
+                # 1 > decimal points where we cannot define dpe ids
+                else:
+                    continue
+        # define type ids
+        # for column names: 2
+        # for column values (contents): 3
+        # for CLS tokens: 5
+        # for SEP tokens: 6
+        type_ids = np.zeros(input_ids.shape, dtype=int)
+        type_ids[:, 0] = 5  # CLS tokens
+        for i, col_name_offset in enumerate(col_name_offsets):
+            type_ids[i][lengths_before_padding[i] - 1] = 6  # SEP tokens
+            # fill with type ids for column values
+            type_ids[i][1 : lengths_before_padding[i] - 1] = 3
+            for start, end in col_name_offset:
+                start_index, end_index = find_boundary_between(
+                    tokenized_events[i].offsets[1 : lengths_before_padding[i] - 1],
+                    start,
+                    end,
+                )
+                # the first offset is always (0, 0) for CLS token, so we adjust it
+                start_index += 1
+                end_index += 1
+                # finally replace with type ids for column names
+                type_ids[i][start_index:end_index] = 2
+
+        return np.stack([input_ids, type_ids, dpe_ids], axis=1).astype(np.uint16)
+
+    events_data = []
+    worker_id = multiprocessing.current_process().name.split("-")[-1]
+    if worker_id == "MainProcess":
+        worker_id = 0
+    else:
+        # worker_id is incremental for every generated pool, so divide with num_shards
+        worker_id = (int(worker_id) - 1) % num_shards  # 1-based -> 0-based indexing
+    if worker_id == 0:
+        progress_bar = tqdm(df_chunk.iter_rows(), total=len(df_chunk))
+        progress_bar.set_description(f"Processing from worker-{worker_id}:")
+    else:
+        progress_bar = df_chunk.iter_rows()
+
+    for row in progress_bar:
+        events_data.append(meds_to_remed_unit(row))
+    data_length = list(map(len, events_data))
+    data_index_offset = np.zeros(len(data_length), dtype=np.int64)
+    data_index_offset[1:] = np.cumsum(data_length[:-1])
+    data_index = pl.Series(
+        "data_index",
+        map(
+            lambda x: [data_index_offset[x] + y for y in range(data_length[x])],
+            range(len(data_length)),
+        ),
+    )
+    events_data = np.concatenate(events_data)
+
+    df_chunk = df_chunk.select(["subject_id", "cohort_end", "cohort_label", "time"])
+    df_chunk = df_chunk.insert_column(4, data_index)
+    df_chunk = df_chunk.explode(["cohort_end", "cohort_label"])
+    df_chunk = df_chunk.group_by(
+        # ["subject_id", "cohort_start", "cohort_end", "cohort_label"], maintain_order=True
+        ["subject_id", "cohort_end", "cohort_label"],
+        maintain_order=True,
+    ).agg(pl.all())
+
+    df_chunk = df_chunk.sort(by=["subject_id", "cohort_end"])
+    # regard {subject_id} as {cohort_id}: {subject_id}_{cohort_number}
+    df_chunk = df_chunk.with_columns(pl.col("subject_id").cum_count().over("subject_id").alias("suffix"))
+    df_chunk = df_chunk.with_columns(
+        (pl.col("subject_id").cast(str) + "_" + pl.col("suffix").cast(str)).alias("subject_id")
+    )
+    # data = data.drop("suffix", "cohort_start", "cohort_end")
+    df_chunk = df_chunk.drop("suffix", "cohort_end")
+
+    length_per_subject = {}
+    progress_bar = tqdm(
+        df_chunk.iter_rows(),
+        total=len(df_chunk),
+        desc=f"Writing data from worker-{worker_id}:",
+    )
+
+    for sample in progress_bar:
+        with h5py.File(str(output_dir / output_name / f"{worker_id}.h5"), "a") as f:
+            if "ehr" in f:
+                result = f["ehr"]
+            else:
+                result = f.create_group("ehr")
+
+            sample_result = result.create_group(sample[0])
+
+            data_indices = np.concatenate(sample[3])
+            data = events_data[data_indices]
+            sample_result.create_dataset("hi", data=data, dtype="i2", compression="lzf", shuffle=True)
+
+            times = np.concatenate(sample[2])
+            times = [datetime.strptime(x, "%Y-%m-%d %H:%M:%S") for x in times]
+            times = np.cumsum(np.diff(times))
+            times = list(map(lambda x: round(x.total_seconds() / 60), times))
+            times = np.array([0] + times)
+
+            sample_result.create_dataset("time", data=times, dtype="i")
+            sample_result.create_dataset("label", data=int(sample[1]))
+
+            length_per_subject[sample[0]] = (len(times), worker_id)
+    del df_chunk
+
+    return length_per_subject
+
+
+if __name__ == "__main__":
+    parser = get_parser()
+    args = parser.parse_args()
+    main(args)