# Deep learning-based infrared thermography reveals reproducible uniform and individual thermoregulatory responses during running

## TL;DR

This study shows that deep learning-assisted infrared thermography can reliably capture both shared and individual thermoregulatory responses during running.

## Contribution

The study introduces a reproducible deep learning method for analyzing thermoregulatory responses during exercise using IRT.

## Key findings

- TP metrics showed strong correlations with heart rate and oxygen consumption during running.
- TNV metrics varied based on individual anaerobic threshold and demonstrated high reproducibility.
- Deep learning-assisted IRT provides consistent metrics across repeated exercise sessions.

## Abstract

Infrared thermography (IRT) has recently gained attention in the field of exercise physiology, due to its ability to monitor thermoregulatory and cardiopulmonary responses non-invasively and in real time during physical exercise. However, the reproducibility of intra-individual measurement and standardization of region-of-interest selection in relation to the acute exercise response remain inconclusive. This study aimed to examine the reproducibility and physiological relevance of specific skin temperature (TSK) metrics processed automatically using deep learning-assisted IRT during running, and to synchronize these metrics with cardiopulmonary and thermoregulatory parameters. Eleven endurance-trained individuals performed three 46-min running sessions over 2 days, with the same average external load but different intensity distributions. Individual anaerobic threshold velocity (vIAT), previously determined by cardiopulmonary exercise testing, was used to prescribe running intensity. During exercise, oxygen consumption (VO2), core temperature (TCORE), heart rate (HR) and different TSK metrics, including non-vessel (TNV), cutaneous arterial perforator (TP), and superficial vein patterns, were continuously measured. All TSK metrics displayed consistent temporal dynamics aligned with external load, but their absolute temperature levels differed systematically. During intermittent running and recovery, TP exhibited robust correlations with HR and VO2 (r = − 0.63 to − 0.9, p < 0.001), and TP entropy showed consistent associations with TCORE during the warm-up (r = 0.59–0.83, p < 0.001). This indicates uniform response patterns across the cohort. In contrast, TNV demonstrated heterogeneous correlations with TCORE, depending on individual exercise capacity. A strong inverse correlation was identified between ∆TNV and vIAT (r = − 0.74 to − 0.88, p ≤ 0.009) and individuals with higher vIAT demonstrated greater TCORE-TNV gradients during running. Measurements of ∆TNV demonstrated high reproducibility, with intra-individual ICC(3,1) values of 0.89 for recovery and 0.76 for warm-up, and no statistically significant differences between the three sessions. Deep learning-assisted IRT provides reproducible, physiologically consistent metrics across repeated exercise sessions, regardless of the day or prior load. Distinct TSK metrics capture both uniform and individual-specific thermoregulatory responses. Variability in peripheral temperature regulation is more strongly associated with running velocity at the individual anaerobic threshold than with maximal cardiorespiratory fitness.

The online version contains supplementary material available at 10.1038/s41598-026-44102-6.