prepare_mds_ed_dataset.py

"""Prepare the UniPACT MDS-ED prompted dataset.

Reads the MDS-ED tabular splits and emits, for each (split, task family),
a JSON array of prompted Yes/No conversation samples paired with an ECG
reference. Optionally merges the per-task arrays of one split into a single
``<split>.json``.
"""

from __future__ import annotations

import argparse
import gc
import json
import os
from pathlib import Path
from typing import Iterable, Iterator

import pandas as pd


CONDITIONS_DETERIORATION = {
    "deterioration_severe_hypoxemia": "experience severe hypoxemia",
    "deterioration_ecmo": "require ECMO (extracorporeal membrane oxygenation)",
    "deterioration_vasopressors": "require vasopressors",
    "deterioration_inotropes": "require inotropes",
    "deterioration_mechanical_ventilation": "require mechanical ventilation",
    "deterioration_cardiac_arrest": "experience cardiac arrest",
}

CONDITIONS_ICU = {
    "deterioration_icu_24h": "require ICU admission within the next 24 hours",
    "deterioration_icu_stay": "require ICU admission during this hospital stay",
}

CONDITIONS_MORTALITY = {
    "deterioration_mortality_1d": "die within 24 hours",
    "deterioration_mortality_7d": "die within 7 days",
    "deterioration_mortality_28d": "die within 28 days",
    "deterioration_mortality_90d": "die within 90 days",
    "deterioration_mortality_180d": "die within 180 days",
    "deterioration_mortality_365d": "die within 365 days",
    "deterioration_mortality_stay": "die during the hospital stay",
}

TEMPLATES = {
    "deterioration": (
        'You are a cardiologist. Your task is "to predict whether a patient will experience clinical deterioration" '
        "based on the provided ECG and Electronic Health Record (EHR) data. "
        "{EHR} "
        'Will the patient "{condition}"? Answer strictly with Yes or No.'
    ),
    "icu": (
        'You are a cardiologist. Your task is "to predict whether a patient will require ICU admission" '
        "based on the provided ECG and Electronic Health Record (EHR) data. "
        "{EHR} "
        'Will the patient "{condition}"? Answer strictly with Yes or No.'
    ),
    "mortality": (
        'You are a cardiologist. Your task is "to predict whether a patient will experience mortality" '
        "based on the provided ECG and Electronic Health Record (EHR) data. "
        "{EHR} "
        'Will the patient "{condition}"? Answer strictly with Yes or No.'
    ),
    "diagnose": (
        'You are a cardiologist. Your task is "to predict the correct ICD-10 diagnosis code" '
        "based on the provided ECG and Electronic Health Record (EHR) data. "
        "{EHR} "
        'Will the patient be diagnosed with "{condition}"? Answer strictly with Yes or No.'
    ),
}

GENDER_MAP = {0: "female", 1: "male"}

DECIMAL_COLUMNS = [
    "biometrics_bmi", "biometrics_weight", "biometrics_height",
    "vitals_temperature_mean", "vitals_heartrate_mean", "vitals_resprate_mean",
    "vitals_o2sat_mean", "vitals_sbp_mean", "vitals_dbp_mean",
]

VITAL_FIELDS = (
    "vitals_temperature_mean", "vitals_heartrate_mean",
    "vitals_resprate_mean", "vitals_o2sat_mean",
    "vitals_sbp_mean", "vitals_dbp_mean", "vitals_acuity",
)

BIOMETRIC_FIELDS = ("biometrics_bmi", "biometrics_weight", "biometrics_height")

SPLITS = ("train", "val", "test")
TASKS = ("deterioration", "icu", "mortality", "diagnose")
RANDOM_STATE = 42


def load_split(csv_path: Path) -> pd.DataFrame:
    """Load a split CSV, decode gender, and round floating-point columns."""
    df = pd.read_csv(csv_path, low_memory=False)
    df["demographics_gender"] = df["demographics_gender"].map(GENDER_MAP)
    df[DECIMAL_COLUMNS] = df[DECIMAL_COLUMNS].round(1)
    return df


def render_ehr_text(row: pd.Series) -> str:
    """Compose a single-paragraph EHR description from one row."""
    demo = []
    if not pd.isna(row.get("demographics_age")):
        demo.append(f"{row['demographics_age']} year-old")
    if not pd.isna(row.get("general_race")):
        demo.append(row["general_race"].replace("/", " ").replace(" - ", " "))
    if not pd.isna(row.get("demographics_gender")):
        demo.append(row["demographics_gender"])
    bio = [
        f"{c.replace('biometrics_', '').replace('_', ' ')} {row[c]}"
        for c in BIOMETRIC_FIELDS if not pd.isna(row.get(c))
    ]
    vital = [
        f"{c.replace('vitals_', '').replace('_mean', '').replace('_', ' ')} {row[c]}"
        for c in VITAL_FIELDS if not pd.isna(row.get(c))
    ]
    sentences = [
        "The demographics information, " + ", ".join(demo) + "." if demo else "",
        "The biometrics information, " + ", ".join(bio) + "." if bio else "",
        "The vital parameters, " + ", ".join(vital) + "." if vital else "",
    ]
    return " ".join(s for s in sentences if s).strip()


def generate_task_json(
    df: pd.DataFrame,
    conditions: dict[str, str],
    template: str,
    output_dir: Path,
    suffix: str,
    balanced: bool,
) -> None:
    """Emit one task-family JSON for one split.

    If ``balanced`` is True, the majority class is downsampled per condition
    so that Yes and No counts match. Otherwise both classes are kept intact.
    """
    out_records: list[dict] = []
    for col, desc in conditions.items():
        yes_rows = df[df[col] == 1]
        no_rows = df[df[col] == 0]
        if balanced:
            n = min(len(yes_rows), len(no_rows))
            yes_sample = yes_rows.sample(n=n, random_state=RANDOM_STATE) if len(yes_rows) > n else yes_rows
            no_sample = no_rows.sample(n=n, random_state=RANDOM_STATE) if len(no_rows) > n else no_rows
        else:
            yes_sample, no_sample = yes_rows, no_rows
        chunk = (
            pd.concat([yes_sample, no_sample])
            .sample(frac=1, random_state=RANDOM_STATE)
            .reset_index(drop=True)
        )
        n_yes = n_no = 0
        for idx, row in chunk.iterrows():
            answer = "Yes" if row[col] == 1 else "No"
            n_yes += answer == "Yes"
            n_no += answer == "No"
            question = template.format(condition=desc, EHR=render_ehr_text(row))
            out_records.append({
                "id": f"{col}_{idx}",
                "ecg": row.get("general_file_name", ""),
                "conversations": [
                    {"from": "human", "value": f"<ecg> {question}"},
                    {"from": "gpt", "value": answer},
                ],
            })
        print(f"  {col}: Yes={n_yes}, No={n_no}", flush=True)

    out_path = output_dir / f"all_tasks_{suffix}.json"
    print(f"  -> {out_path}", flush=True)
    with out_path.open("w", encoding="utf-8") as f:
        json.dump(out_records, f, separators=(",", ":"))
    gc.collect()


def stream_records(path: Path) -> Iterator[dict]:
    """Yield items from a JSON-array file. Uses ijson if available."""
    try:
        import ijson  # type: ignore
    except ModuleNotFoundError:
        with path.open("r", encoding="utf-8") as f:
            yield from json.load(f)
        return
    with path.open("rb") as f:
        yield from ijson.items(f, "item")


def merge_split(output_dir: Path, split: str, tasks: Iterable[str]) -> None:
    """Concatenate per-task JSONs of one split into ``<split>.json`` (streamed)."""
    out_path = output_dir / f"{split}.json"
    print(f"merging {split} -> {out_path}", flush=True)
    n = 0
    with out_path.open("w", encoding="utf-8") as out:
        out.write("[")
        first = True
        for task in tasks:
            part = output_dir / f"all_tasks_{split}_predictions_{task}.json"
            if not part.exists():
                print(f"  skip missing {part}", flush=True)
                continue
            for record in stream_records(part):
                if not first:
                    out.write(",")
                out.write(json.dumps(record, separators=(",", ":")))
                first = False
                n += 1
        out.write("]")
    print(f"  merged {n} samples", flush=True)


def build_arg_parser() -> argparse.ArgumentParser:
    p = argparse.ArgumentParser(description=__doc__.splitlines()[0])
    p.add_argument("--data-dir", type=Path, required=True,
                   help="Directory containing mds_ed_{train,val,test}.csv.")
    p.add_argument("--output-dir", type=Path, required=True,
                   help="Output directory for per-task and merged JSONs.")
    p.add_argument("--icd-xlsx", type=Path, required=True,
                   help="Excel file with columns 'Original Column' and 'Description'.")
    p.add_argument("--splits", nargs="+", choices=SPLITS, default=list(SPLITS),
                   help="Which splits to process. Default: all.")
    p.add_argument("--tasks", nargs="+", choices=TASKS, default=list(TASKS),
                   help="Which task families to generate. Default: all.")
    p.add_argument("--balanced-splits", nargs="*", choices=SPLITS,
                   default=["val"],
                   help="Splits whose majority class should be downsampled to "
                        "match the minority class per condition. Splits not "
                        "listed here keep the original prevalence. "
                        "Default: val (train and test kept unbalanced).")
    p.add_argument("--merge", action="store_true",
                   help="After generation, merge per-task JSONs into <split>.json.")
    p.add_argument("--no-generate", action="store_true",
                   help="Skip generation; only run the merge step on existing parts.")
    return p


def main() -> None:
    args = build_arg_parser().parse_args()
    args.output_dir.mkdir(parents=True, exist_ok=True)

    icd_df = pd.read_excel(args.icd_xlsx)
    icd_mapping = dict(zip(icd_df["Original Column"], icd_df["Description"]))

    task_specs = {
        "deterioration": (CONDITIONS_DETERIORATION, TEMPLATES["deterioration"]),
        "icu": (CONDITIONS_ICU, TEMPLATES["icu"]),
        "mortality": (CONDITIONS_MORTALITY, TEMPLATES["mortality"]),
        "diagnose": (icd_mapping, TEMPLATES["diagnose"]),
    }
    balanced_set = set(args.balanced_splits)

    if not args.no_generate:
        for split in args.splits:
            csv_path = args.data_dir / f"mds_ed_{split}.csv"
            print(f"loading {csv_path}", flush=True)
            df = load_split(csv_path)
            balanced = split in balanced_set
            print(f"split={split} balanced={balanced}", flush=True)
            for task in args.tasks:
                conditions, template = task_specs[task]
                print(f"task={task}", flush=True)
                generate_task_json(
                    df, conditions, template, args.output_dir,
                    suffix=f"{split}_predictions_{task}", balanced=balanced,
                )
            del df
            gc.collect()

    if args.merge:
        for split in args.splits:
            merge_split(args.output_dir, split, args.tasks)


if __name__ == "__main__":
    main()