HandDx-200: Multimodal RGB and Thermal Hand Images Paired with Clinical Biomarkers

HandDx-200 is a multimodal dataset comprising RGB and thermal images of palmar and dorsal views of both hands, paired with clinical and laboratory biomarkers from 198 adult participants (73 male, 125 female; age 39.7 ± 11.7 years). Images were acquired under a standardized protocol using a Nikon D3200 RGB camera and a Mastfuyi FY12 infrared thermal imager, with controlled ambient temperature and humidity. Clinical data include age, sex, blood pressure, fasting capillary blood glucose (measured after a minimum 8-hour self-reported fast), height, weight, and BMI. Laboratory biomarkers include complete blood count (CBC), lipid profile (total cholesterol, triglycerides, HDL, LDL, VLDL), and HbA1c. LDL and VLDL are calculated by the Friedewald equation; no participant had triglycerides above 400 mg/dL (maximum 388 mg/dL), so the derived values are valid throughout. The CBC was obtained on a 3-part differential analyser, so eosinophil and basophil counts are not separately resolved. All RGB images underwent a fully deterministic preprocessing pipeline (white-balance correction, hand segmentation, orientation normalization, and cropping), with per-image JSON logs documenting every processing step for each of the 792 RGB images. Segmentation accuracy was quantitatively validated against 68 manually annotated images spanning the full cohort (P001–P198): an initial set of 48 images from participants P001–P050 plus 20 images sampled at random from P051–P198 (numpy seed 42, balanced across hand side and view). Mean Dice = 0.993 ± 0.004 (95% CI 0.992–0.995), mean IoU = 0.987 ± 0.009; all 68 images scored ≥ 0.90 (minimum 0.972), with no difference between the early and later cohort subsets, indicating no protocol drift. To quantify annotation uncertainty, a second independent annotator re-traced 15 of these images: inter-annotator agreement was Dice 0.993 ± 0.006, i.e. algorithm-versus-annotator agreement is indistinguishable from agreement between two human annotators. Both annotators' masks and the per-image results are included in quality_control. Ethics approval: Medical Research Ethics Committee, National Research Centre, Egypt (preliminary approval 25 June 2025, authorising recruitment and the full data-collection protocol; final approval 25 February 2026; No. 15120226). OSF preregistration: https://doi.org/10.17605/OSF.IO/4QUGB Important limitations for users: Thermal images are exported as 8-bit pseudocolor BMP files (Rain palette) representing relative spatial thermal patterns. They are not radiometrically calibrated and cannot be converted to absolute temperature. Each frame additionally carries device-generated centre/high/low temperature text, a palette legend and pointer markers baked into the image, including over the hand region; these must be cropped or masked before any texture or gradient analysis, and absolute pixel values must not be compared across participants. The RGB and thermal images are not spatially registered. 123 of the 792 RGB images (15.5%) received fallback whole-image white-balance correction instead of gray-card correction, because the gray reference patch was occluded. These are flagged in the WB_method column of handdx_image_metadata_qc.csv and should be stratified or excluded in colorimetric analyses. All participants fall within the Dark ITA category (ITA < −30 degrees; mean −52.1° ± 3.9°), consistent with the Egyptian recruitment context; independent validation is required before applying derived models to lighter skin tones. Each participant contributes multiple images, so machine-learning users must apply participant-level train/test splits to avoid data leakage.