Multimodal Diabetes Dataset: Near-Infrared Optical and Wearable Sensor Signals for Non-Invasive Glucose Screening

This dataset provides 15,720 measurement sessions for non-invasive diabetes screening research, combining three data modalities: dual-wavelength near-infrared (NIR) optical sensor readings (660nm and 940nm, plus a derived 660/940nm ratio), wearable physiological signals (heart rate, blood volume pulse, electrodermal/skin conductance, wrist temperature, and movement intensity), and demographic variables (age, BMI). Each session is labeled final_label__is_diabetic (1 = diabetic, 0 = non-diabetic), derived from a paired blood glucose measurement using the standard 126 mg/dL clinical threshold. The dataset aggregates 57 columns from five underlying sources — PhysioCGM, Nature Scientific Reports NIR Glucose, Kaggle Raman Diabetes, Raman Sugars, and NTNU NIR Glucose — combining 56.1% real-world sensor recordings with 43.9% statistically modeled synthetic data generated to be consistent with published correlation values. Class distribution is 66.7% diabetic and 33.3% non-diabetic. The data is pre-partitioned into a stratified 70/15/15 train/validation/test split (identified by a split column), preserving class ratio across partitions. The dataset is intended to support development of non-invasive diabetes screening models — i.e., predicting diabetes risk from wearable- and optical-sensor signals alone, without a blood draw. Exploratory analysis shows the NIR optical features carry the strongest signal (Pearson r up to −0.72 with the diabetic label), followed by heart-rate-derived wearable features (moderate correlation), while demographic features (age, BMI) show negligible correlation in this specific dataset. A full data dictionary, per-feature correlation/statistical significance tables, and the code used to generate all summary tables and figures are included.