Speaker
Description
Background and Aim: Thermal imaging is widely used to assess skin temperature. However, interpretation often relies on absolute values or qualitive patterns that provide limited insight into spatial thermoregulatory organization. Palm-to-finger thermal gradients (PFG) capture temperature distribution of the hand which reflects vascular tone and peripheral thermoregulation in response to a variety of physiological or pathological states. However, standardized, quantitative descriptors for palm-to-finger thermal gradient patterns and their population-level variability have not been established. The current study was designed using automated, deep-learning analysis to identify and characterize PFG phenotypes in an active population and determine their prognostic association with demographic and hand load and exposure variables.
Methods: Thermal and optical images of the hands were acquired using a FLIR C5 camera under controlled, indoor conditions. Participants were seated with hands exposed to ambient room temperature (22-25 °C) and allowed to equilibrate thermally for at least 5-minutes prior to imaging. During acquisition, participants positioned their hands approximately 1–2cm above a standardized hand template to ensure consistent posture and minimize conductive heat transfer. Joint localization and multimodal image alignment were performed using a customized, automated pipeline integrating deep-learning–based computer vision tools (Meta’s Segment Anything Model and Google’s MediaPipe HandLandmarker) with joint detection performed using a ResNet50-based convolutional neural network trained on annotated thermal images. Skin temperatures were extracted automatically at each hand joint, and PFGs were computed (ΔCenter–MCP, ΔCenter–PIP, ΔCenter–DIP). Unsupervised k-means clustering was applied to identify distinct hand thermal gradient phenotypes. The dataset included 139 participants (mean age 44.7±11.7 years; BMI 26.7±4.4 kg/m²; 109 men, 30 women). Demographic and hand load and exposure parameters were evaluated using the Jonckheere–Terpstra trend analysis and Somers’ D ordinal-association test.
Results: Four phenotypes were characterized according to absolute temperatures and PFG’s. The coldest cluster showed steep PFG’s (center 26.3 ± 2.1°C; ΔCenter–DIP 6.9 ± 1.4°C); whereas the warmest cluster exhibited nearly flat PFG profiles (center 33.6 ± 1.4°C; ΔCenter–DIP 1.4 ± 1.5°C). Increased body weight and BMI were associated with significant monotonic increases (from colder to warmer) across PFG clusters (Jonckheere–Terpstra test, both p < 0.001). Age and height did not show significant ordered correlations with these thermal phenotypes. Dominant hand laterality was associated with the warmer phenotypes (Somers’ D, p = 0.037). Biological sex was not significantly associated with phenotype rank. Hand load characteristics—including subjective load score, weekly procedural time, and daily computer use—also were not associated significantly with phenotype rank (all p > 0.18).
Conclusion: Using deep-learning automated analysis, we identified and characterized distinct palm-to finger thermal gradient phenotypes that establish reference patterns within the study population. This approach provides a reproduceable quantitative framework for objective assessment of peripheral thermoregulation that might be useful as an additional patient specific prognostic tool in the evaluation of progression and efficacy of treatment of a variety of physiological and pathological states.
