Artificial Intelligence Identifies Individuals with Prediabetes from Single-Lead Electrocardiograms

Contents codes.zip For ECG feature extraction (this will need original ECG signal data) ecg_feature_extraction.sh ecg_feature_extraction.py feature_extractor.py For training with hyperparameter optimization train.sh train.py For prediction of prediabetes/diabetes from ECG feature test.sh test.py raw_ecg_data.zip: 16,766 ECG records used in our analyses. Each record is a 5,000 x 12 matrix in a CSV file. (In this dataset, value 1 represents 4.88 µV.) external_ecg_data.zip: 2,456 ECG records used in our external validation. Each record is a 5,000 x 12 matrix in a CSV file. (In this dataset, value 1 represents 1 µV.) participant_characteristics.csv: Health check records of 16,766 participants where the information below are stored. participant_id: IDs for participant. Some IDs are duplicated because the dataset contains multiple records from some of the participants. ecg_id: IDs for ECG records, all of which are unique age: The age of each participant at the time of the health checkup male_sex: If the participant is male, "True" is recorded smoking: if the participant smokes, "True" is recorded drinking: 1 for "rarely", 2 for "occasionally" and 3 for "regularly" is recorded according to the frequency of drinking height: participant's height in centimeters (cm) weight: participant's body weight in kilograms (kg) BMI: body mass index, calculated using the formula: weight (kg) / [height (m)]^2 pulse_rate: pulse rate in pulse per minute (/min) sBP: systolic blood pressure in mmHg dBP: diastolic blood pressure in mmHg FPG: fasting plasma glucose levels measured in milligrams per deciliter (mg/dL) HbA1c: hemoglobin A1c levels in % dm_under_treatment: if the participant was undergoing treatment for known diabetes, "True" is recorded prediabetes_diabetes: classification label which is "True" if a participant meet either of the following criteria FPG ≥ 110 mg/dL HbA1c ≥ 6.0% Undergoing treatment for diabetes development_data: "True" in records used as development data in our study external_cohort_characteristics.csv: Health check records of 2,456 participants where the information below are stored. ecg_id: IDs for ECG records, all of which are unique prediabetes_diabetes: classification label which is "True" if a participant meet either of the following criteria FPG ≥ 110 mg/dL HbA1c ≥ 6.0% Undergoing treatment for diabetes FPG: fasting plasma glucose levels measured in milligrams per deciliter (mg/dL) HbA1c: hemoglobin A1c levels in % dm_under_treatment: if the participant was undergoing treatment for known diabetes, "True" is recorded ecg_feature_data.zip: extracted ECG features (unprocessed), for 12-lead and 1-lead ECG ecg_features_1-lead.csv [Single-lead (lead I) ECG] ecg_features_12-lead.csv [12-lead ECG] ecg_features_12-leads_external_cohort.csv [12-lead ECG of external cohort] feature_list.zip: List of ECG features used (to be used for ECG extraction for original data) feature_list_269_12-lead.csv [269 features for 12-lead ECG analysis] feature_list_28_1-lead.csv [28 features for single-lead (lead I) analysis] model_12-lead.zip, model_1-lead.zip: model trained with our 12-lead or single-lead (lead I) ECG data, and the classification thresholds, used for test model_fold_1.pkl - model_fold_10.pkl : model for each of 10-fold cross validation average_threshold.pkl : classification threshold, which is the average of 10-fold Codes and data for demo are also available in (https://github.com/dkoga4116/diabetes_detector)