This package contains a large, practice-ready patient table with 1,000,000 rows designed for education, demonstrations, and method development in machine learning and analytics.

Grain: one row = one patient.
Target column: insomnia_class (0/1) for training and evaluation in binary classification tasks.
Scope: demographics, utilization-style counts, comorbidity indicators, and outcome label.
Important: This dataset is not for clinical use.

This document focuses on how to use the dataset effectively. It does not describe its provenance or internal construction.

Files

patients.csv — the main table with 1,000,000 patients.

You may also produce your own analysis outputs (e.g., metrics CSVs, plots) alongside the data.

Table schema

Each row is one patient. Columns cover demographics, counts, comorbidities, and the label.

Column	Type	Description	Typical Values / Range	Missing?
`patient_id`	string	Unique ID	`P000001` …	No
`age`	float	Age in years	≥ 18	Rare
`sex`	category	Administrative sex	`Female`, `Male`	No
`bmi`	float	Body mass index	~10–70	Yes
`ethnicity`	category	Race/ethnicity bucket	`Caucasian`, `African American`, `Hispanic`, `Asian`, `Other`, `Unknown`	Yes
`smoking_status`	category	Tobacco status	`Current`, `Past`, `Never`, `Unknown`	Yes
`emr_fact_count`	int	Total EMR facts (activity proxy)	0–several thousand	Yes
`sleep_disorder_note_count`	int	Sleep-related note count	0+	No
`insomnia_billing_code_count`	int	Insomnia diagnosis code count	0+	No
`anx_depr_billing_code_count`	int	Anxiety/depression code count	0+	No
`psych_note_count`	int	Psychiatry-related note count	0+	No
`insomnia_rx_count`	int	Insomnia-related prescription count	0+	No
`joint_disorder_billing_code_count`	int	Musculoskeletal/joint code count	0+	No
Comorbidity flags	int (0/1)	Indicator variables (see list below)	{0,1}	No
`insomnia_probability`	float	Scoring column in (0,1) useful for ranking	[0,1]	No
`insomnia_class`	int (0/1)	Label for supervised learning	{0,1}	No

Comorbidity flag columns (0/1):
hypertension, lipid_metabolism_disorder, diabetes, gastrointestinal_disorder, anxiety_or_depression, psychiatric_disorder, pneumonia, obesity, congestive_heart_failure, coronary_artery_disease, asthma, copd, cerebrovascular_disease, afib_or_flutter, cancer, peripheral_vascular_disease, osteoporosis, ckd_or_esrd, renal_failure.

Some columns intentionally include missing values to mirror common data challenges.

Loading at scale (Python / pandas)

import pandas as pd

dtypes = {
    "patient_id": "string",
    "age": "float32",
    "bmi": "float32",
    "emr_fact_count": "int32",
    "sleep_disorder_note_count": "int16",
    "insomnia_billing_code_count": "int16",
    "anx_depr_billing_code_count": "int16",
    "psych_note_count": "int16",
    "insomnia_rx_count": "int16",
    "joint_disorder_billing_code_count": "int16",
    # binary flags as compact integers
    "hypertension": "uint8", "lipid_metabolism_disorder": "uint8", "diabetes": "uint8",
    "gastrointestinal_disorder": "uint8", "anxiety_or_depression": "uint8",
    "psychiatric_disorder": "uint8", "pneumonia": "uint8", "obesity": "uint8",
    "congestive_heart_failure": "uint8", "coronary_artery_disease": "uint8",
    "asthma": "uint8", "copd": "uint8", "cerebrovascular_disease": "uint8",
    "afib_or_flutter": "uint8", "cancer": "uint8", "peripheral_vascular_disease": "uint8",
    "osteoporosis": "uint8", "ckd_or_esrd": "uint8", "renal_failure": "uint8",
    "insomnia_probability": "float32",
    "insomnia_class": "uint8"
}

cat_cols = ["sex", "ethnicity", "smoking_status"]

# Option A: read all at once
df = pd.read_csv("patients.csv", dtype=dtypes)
for c in cat_cols:
    df[c] = df[c].astype("category")

# Option B: chunked reading to limit peak memory
chunks = pd.read_csv("patients.csv", dtype=dtypes, chunksize=100_000)
df = pd.concat((chunk.assign(**{c: chunk[c].astype("category") for c in cat_cols}) for chunk in chunks), ignore_index=True)

Quick EDA (Exploratory Data Analysis)

print(df.shape)
print("Label prevalence:", df["insomnia_class"].mean())

# Missingness overview
missing = df.isna().mean().sort_values(ascending=False).head(20)
print(missing)

# Example correlation matrix for numeric counts
key = ["sleep_disorder_note_count","insomnia_billing_code_count","anx_depr_billing_code_count",
       "psych_note_count","insomnia_rx_count","joint_disorder_billing_code_count","emr_fact_count"]
print(df[key].corr(method="spearman"))

Recommended preprocessing (for ML)

Imputation: median (numeric), most-frequent (categorical).
Encoding: one-hot for sex, ethnicity, smoking_status.
Scaling: standardize numeric features for linear models (trees are scale-agnostic).
Outliers: consider clipping or robust scalers for skewed counts.
Splitting: stratify by insomnia_class for train/test.

Example (scikit-learn pipeline):

from sklearn.model_selection import train_test_split
from sklearn.pipeline import Pipeline
from sklearn.compose import ColumnTransformer
from sklearn.preprocessing import OneHotEncoder, StandardScaler
from sklearn.impute import SimpleImputer
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import roc_auc_score

y = df["insomnia_class"].astype(int)
X = df.drop(columns=["insomnia_class"])

cat_cols = X.select_dtypes(include=["category","object"]).columns.tolist()
num_cols = X.select_dtypes(include=[np.number]).columns.tolist()

preprocess = ColumnTransformer([
    ("num", Pipeline([("imp", SimpleImputer(strategy="median")),
                      ("scaler", StandardScaler())]), num_cols),
    ("cat", Pipeline([("imp", SimpleImputer(strategy="most_frequent")),
                      ("ohe", OneHotEncoder(handle_unknown="ignore"))]), cat_cols),
])

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.25, stratify=y, random_state=7)

clf = Pipeline([("prep", preprocess), ("lr", LogisticRegression(max_iter=500))]).fit(X_train, y_train)
print("ROC-AUC:", roc_auc_score(y_test, clf.predict_proba(X_test)[:,1]))

Slices & reporting

When reporting results, consider per-group summaries for fairness diagnostics:

for col in ["sex","ethnicity"]:
    print("\nSlice:", col)
    print(df.groupby(col)["insomnia_class"].mean())

Caveats

This dataset is designed for education and reproducible demonstrations.
It must not be used for clinical decision-making or any activity involving real individuals.
Demographic proportions and feature distributions are provided as-is and should not be interpreted as representing any specific population or institution.

Citation

Kartoun, U. (2025). Patient Cohort — 1,000,000 Rows (Synthetic Dataset). DBbun LLC. Inspired by: Kartoun et al., Sci Rep (2018).

Downloads last month: 31